Explorez tous les épisodes du podcast Interconnects

We’re in an era where a few organizations are using thousands of concurrent agents to improve their processes and output. These organizations happen to be just the frontier AI labs, in particular OpenAI and Anthropic. In the last few weeks, I’ve been pondering what it means for so many employees across these organizations to rapidly update their expectations for the pace of AI progress and associated risks.

A core perspective I have is that the frontier labs and broader frenetic, competitive culture in the San Francisco AI scene set up an environment that amplifies any AI concern. This has some benefits in causing more general audience awareness of AI, as fear sells, but exaggerating risk timelines or severity will have negative second-order effects. I remember many loud AI safety debates, and their associated clouds over the viability of open-source AI, in 2023 and 2024 — the primary risks then did not arrive in the forecasted timelines.

The general populace of these two key labs was very anxious about AI risks and the rate of progress even a year ago, and especially as agents got stronger product-market fit at the start of 2026. This cultural precondition, when exposed to the reality that thousands of agents will constantly be working fairly productively in your business, will only increase this anxiety. The step from this anxiety, and incidents like OpenAI-HuggingFace, to extinction risks feels very religious.

Richard Ngo had an apt summary of the situation:

Now a large proportion of the AI safety community is implicitly or explicitly orienting to futures where an intelligence explosion occurs within a few years. My default expectation (absent an extensive pause) is that a similar thing will happen: they’ll turn out to be directionally correct (relative to the expectations of almost anyone not linked to the community) but factually wrong. Specifically, we won’t have superintelligence within the next 8 years, but things will still be moving so fast that it’ll *feel* like the people who argued for short timelines were right.

… I wanted to say something now because it feels like the level of bandwagoning towards “singularity soon” is getting pretty wild.

Personally, I think this view aligns closely to what I outlined in my alternate scenario to true recursive self-improvement (RSI), which I called lossy self-improvement. A summary of this view is that:

* Automatable research is too narrow to achieve a massive net acceleration in progress, in the face of scaling laws’ exponential costs,

* Diminishing returns of more AI agents in parallel are real, &

* Resource bottlenecks and politics are a major factor in building strong LLMs (and AI can do much less to accelerate this).

So, I’m left balancing the above, latent increase in the cultural temperature with the potential that the labs have seen genuinely scary, specific breakthroughs that are not public yet. My expectation is that more of the current AI safety concern is on the former – scaled agents working – but I hold high levels of uncertainty here. Foundational, imagination-based AI breakthroughs are the sort of thing that would make me update my RSI timelines from closer to a tool to sustain progress in the face of exponential costs (scaling laws), to something more unpredictable and/or unstable.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Some of the best recent resources on RSI have been Dwarkesh’s podcasts with Noam Brown and the trio of John Schulman, Beren Millidge and Charlie O’Neill. I have a few important reflections from both of them.

First, the podcast with Noam Brown made me internalize how big of a short-term acceleration mass inference capacity is. These labs will throw thousands of agents at important, measurable problems. At the same time, compute capacity available to them is going to continue to scale. I have my doubts that the labs can afford to spend a constant portion of this compute on internal R&D as the total volume goes up, especially with plans to IPO, as they face increased scrutiny on basic economics. It is important to not confuse massive steps in inference-time scaling, a dynamic which should be fairly predictable, with being the outputs of RSI, which is highly uncertain.

Second, the trio podcast debating the state of the art in technical capacities induced more of a surprising reaction that I haven’t fully settled. Through the first hour or so of this podcast, where they debate the role of RL, distillation, scaling, inference-time compute, etc., I found myself strongly agreeing with the distribution of claims. A TLDR would be that our current techniques work and let us solve problems we know how to state, but they don’t result in a magical level of generalization to unknown, harder problems in most partially verifiable domains (i.e. progress in math is an exception, rather than a rule).

The surprise of this podcast was the end, where they were predicting timelines for various thresholds of AI. I had GPT-6-Astra summarize the answers provided to three questions from Dwarkesh, of the form “when will AI reach X ability”:

All timelines are relative to the interview date.

* Drop-in remote worker for broad white-collar work over a month

* Charlie O’Neill: ~1 year with programmatic access to workplace tools; ~2 years if it must operate through a browser. Means ordinary white-collar work, not highly creative research.

* Beren Millidge: ~3 years for full generality; 80–90% coverage sooner. Main uncertainties: online learning and the long tail of tasks.

* John Schulman: ~1 year for an “okay” version, with uneven capabilities that improve over time.

* 10× productivity uplift for AI researchers

* Charlie O’Neill: 5–10 years. Bottleneck: absorbing information and deciding which experiment to run next.

* Beren Millidge: Finds John’s ~2-year estimate plausible, but gives no independent timeline. Assumes AI can run successive experiments and learn from feedback; other bottlenecks would remain.

* John Schulman: ~2 years.

* AI surpassing top human experts across all computer-based work, including multiyear projects (“ASI”)

* Charlie O’Neill: 5–10 years. Highlights limitations in memory and context length.

* Beren Millidge: ~5 years for areas labs focus on; potentially longer for literally every domain. Gives no firm timeline for the universal version.

* John Schulman: 3–4 years. Spatial/physical fields may take longer; requires onboarding and solving longer-horizon learning.

Roughly, a recurring problem when discussing RSI is a lack of specification in intelligence. The jaggedness of intelligence means that we need to discuss thresholds in specific, measurable tasks. The nature of LLMs’ intelligence is shaped very differently than humans, and the roles we forecast are human-shaped. AIs, therefore, do not cross these thresholds like remote worker or AI researcher discretely. It’s a slow diffusion, and a form of long tail will always exist.

Take the case of productivity of AI researchers. Many people under-index how much of science is communication and standard setting with colleagues. I do buy the cycle of experiment design and testing being 10x faster in the near future, but not hypothesis generation and intuition building. Accelerating understanding will be the key bottleneck – and it is one that despite all of the AI tools getting massively improved, humans will only improve marginally in their capability. A big improvement in the nature of science will be enabling humans to invest more time here, not them becoming exponentially better at it.

This links back to the Noam podcast. Agent swarms in the near future will be effective at solving clear, open problems with verifiable answers. In this vein, when it comes to improving AI models, RSI is much more helpful at efficiency rather than expanding peak intelligence. This is due to the fact that LLM serving has clear metrics you want to improve that are measurable and malleable. This’ll enable better inference-time scaling and more efficient multi-agent systems.

Still, I cannot get past the fact that all of our scaling laws show that you need exponential compute and resources to make linear improvements in intelligence. RSI is poised to make modern LLMs vastly cheaper. Trends that have shown LLMs get exponentially cheaper at a given intelligence are likely to accelerate. A crucial factor for the labs will be increasing margins as revenue could potentially have negative pressure if there’s fierce competition in lowering prices at a fixed intelligence level — Jevons paradox will likely prevail, resulting in strong businesses.

RSI factors will have a much harder time improving pieces of the LLM puzzle like managing complex post-training recipes. There were a few quotes from John Schulman that I strongly agree with on the state of post-training at the labs:

If I think about a post-training team and why you need a lot of people on the team, it’s just because there are a lot of different areas where you have to figure out how the model should behave. It would be very hard to automate the whole thing, just because someone has to think about how the model should behave in this area.

and later:

It’s really easy to screw up post-training in some way that doesn’t show up in benchmarks.

These tasks are uniquely hard for current LLMs. Yes, they’ll get better as the industry is still rapidly scaling RL environments related to these domains, but this paradigm does not last forever. In the near future, it could become exponentially harder to conceive, build, and test new environments that meaningfully challenge the leading LLMs – these hard environments are the ones that are crucial as a learning signal in RL.

OpenAI and Anthropic have shared a good amount of internal measurements related to RSI, and my current read is that the biggest takeoff in automation within the labs is in tasks like software engineering, monitoring logs, managing planned experiments, and other fairly routine (but not always easy) tasks. For example, I was surprised by this language in the recent Claude Fable 5.1 & Mythos 5.1 System Card:

We believe that internal usage of recent AI models has been a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate.

Altogether, I think the hardest exponential we are fighting is on peak intelligence. That is the hardest one to budge or even accelerate. Still, my mental model for the very early innings of RSI is more of massively scaling and diffusing inference-time compute to AI research and related activities, which has a large amount of low-hanging fruit available. This, on its own, is still poised to be economically transformative. It may also unlock more resources to push on AI diffusion, which is the crucial bottleneck in unlocking much of the potential benefits of AI.

For now and until more evidence emerges, lossy self-improvement remains my baseline on the trajectory of progress, and the increased discussion of extinction risk seems very misplaced. As always, things can change fast in AI.

As AI became more powerful, it was inevitable that a different, growing group would start to take AI safety more seriously – what we did not know ahead of time, is which set of views they latched onto. We have seen that some of the most extreme views of risk, i.e. moderate probabilities of mass extinction, were the ones that reached the masses. A lot in the AI world is about to change due to this.

How did we get here? Why did this quitting announcement reach so far? In many ways, the rest of the world’s views around AI in the past was a dampening factor. You can think about this like the damp ground around a fire. Many people were striking matches for years about AI risk – they’d smolder in their community and largely burn out, going unnoticed. As the stakes of AI have risen this year, from the OpenAI-HuggingFace incident and breakthroughs like the Navier-Stokes result (also from OpenAI), the ground has dried out and the latent energy around the AI discourse has increased. More people not in the industry have thought, “huh, maybe I should care about this AI thing.” The ambient temperature and stakes have been obviously rising.

Then, some basic factors of human nature apply, with the most crucial being that fear sells. Fear is the simplest story, the one people cannot look away from. Jacob Coxon was the one who stumbled into this new powder keg, totally unaware of what was going to come. What looked like a fairly innocuous event – another AI researcher quitting citing safety risks – landed into a very different environment and it caught like wildfire. The discussion of existential risk, mass extinction, and the trajectory of AI has traveled further than even the most seasoned AI commentariat would ever predict.

There are a set of facts we need to get clear, which paint the picture of the situation. The key Tweets to reference are from Jacob Coxon, the resignation thread, and Evan Hubinger, the source of the >10% extinction risk figure .

* There are plenty of AI risks which are likely to cause harm, even if estimating annihilation is useless. It is important to weigh these with respect to the benefits. The entire discourse around existential risk is on very poor footing. At least Evan was clear in his post, with “kill all humans,” but a major problem in the AI Safety discourse is that people talk about existential risks, when they mean very different things (much like how AGI is a vaguely meaningless term). I put the probability of complete extinction as being so low it isn’t worth discussing, but the probabilities of AI caused disasters – e.g. cyber attacks on critical infrastructure or bio-risks – as being worth debating. Throwing this whole discussion out because there are not these disasters yet is a harmful reaction.

* Jacob Coxon is acting genuinely and with good intentions. The outpouring of support from more well-established AI researchers who know of him and his intentions of resignation is useful. Many factions of AI turned to scapegoating him individually, based on account metadata, personal factors, etc. These are not useful. Many frontier lab employees genuinely have similar views to him. I’m not sure it’s a majority, but there is a substantial group.

* Many frontier lab employees, especially at Anthropic, are out of touch and this will impact their forecasting and/or descriptions of current AI events. I say this without blaming individuals, but it’s a common agreement among my friends not at OpenAI/Anthropic (Ant especially) that people at the labs operate with a religious energy. It’s very common to go through very out of touch interactions with them. I do not blame most of the individuals who get distorted views being part of these companies, but the interactions are wild and spill over into a lot of wack discussions in the AI media ecosystem. Living in this environment that normalizes such out of touch behavior will inevitably distort any human’s understanding of technical progress.

Interconnects AI is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

* This was not a mass political campaign, but rather an opportunistic media coordination. For context, the Wall Street Journal had an exclusive story that Jacob coordinated before posting. I suspect that Jacob shared his plan of quitting in groupchats with AI safety advocacy groups ahead of time, e.g. the morning of posting, asking for amplification. This is normal practice, and could have included some prominent politicians. From there, I think it’s more likely that other politicians are bandwagoning on a rising issue. When you combine this with other factors, like Daniel Kokotajlo’s appearance on Joe Rogan coming out the same day – it definitely looks like a very well-executed, coordinated media campaign. This doesn’t mean it’s a conspiracy or a regulatory capture tactic within Democratic political structures. The determining factor seems to be that no one – including Jacob and those posting about X-risk today – knew that it would go so viral.

* We do not have proof that RSI causes the risks these researchers forecast. The general argument for RSI follows as: The current pace of progress is very high, the current progress is heavily dependent on AI tools, the current AI tools are superhuman in some domains (e.g. math) – so, all together, AI is going to work more on itself and become superhuman in all relevant areas over time to autonomy and intellect. This view dramatically undersells human bottlenecks in building models and allocating resources at organizations, and draws conclusions on future AI capabilities more broadly.I called my alternative view to this, Lossy self-improvement. AI has always been very jagged, and we are making models which are superhuman goal-seekers at math and software engineering, but they have massive limitations on intuitions, creativity, and other types of reasoning that humans are strong at. With AI agents assisting research, we will rapidly find the areas where AI is superhuman – and I expect there to be well more than just research mathematics – but it won’t be a panacea for the current limitations of our approaches to LLMs.

There is another understandable social dynamic at play here, causing many deep AI insiders to overstate the returns from RSI. Many of these researchers were the earliest people to bet on AI’s progress, and the extent to which they were visionaries should not be downplayed (see Ilya’s comments on deep learning as early as 2015). They have been right again and again, forecasting AI’s capabilities better than I certainly could have guessed. This does not, though, mean that their forecast of what will come next will be right. The core idea of RSI is a way to spend more compute on the process of developing a model recipe, rather than just spending more compute on the training run itself. We’re seeing benefits from it, but I argue the expected return on that input is far less than they believe.

Their argument is that RSI will make AI progress go exponential, make it so we cannot monitor the technology, and enable rogue models and new forms of risk. This scenario is often called “Fast Takeoff”. We have not seen the stacking efficiency gains that massively reduce model size and cost, that would lead to an explosion in progress by allowing consistent speedups in experimentation.

* The biggest short-term risk could be from the AI labs not taking safety seriously enough – they haven’t hardened their own infrastructure, enabling AI misuse to proliferate. From my earlier post on the HuggingFace-OpenAI incident, Lessons from the hacks:

* Frontier labs do not seem like they’re watching the models closely enough, due to a general frenetic competitive environment & current SF culture

From OpenAI’s own retrospective, the misaligned model behavior was unfolding over months, and in some cases OpenAI did not know about the hacks for ~weeks. The time to response is too long and I do not think this is an OpenAI only characteristic – rather it is that the frontier labs continually seem underwater in the amount of work they feel like they should do. I am not optimistic in the long-term that the labs change a sufficient amount here to meaningfully mitigate this type of oversight risk in the future. Yes, it is very likely that OpenAI is putting a ton into understanding this – and delayed their latest models to make sure they get it right – but the financial pressure to grow revenue or risk the companies’ long-term balance sheets makes me think it will not be a sustained pattern of caution.

Overall, I think this episode is very bad for the AI ecosystem. It’s pushed the acceptable views in the AI community closer to the extremes. More accelerationists will discount the need for any form of safety, citing mass delusion of the “doomers.” It feels like a very narrow path to believe in AI risks, but to not worry about extinction from the technology.

For example, it is a horrible temporary period for cybersecurity, where AI models going a bit off script and poking around unintended pieces of the web seems like a new normal. This is accelerated by the labs competing veraciously towards their views of AGI, and a slow uptake in the necessary hardening of our cyber infrastructure around the world. This doesn’t mean that it’s an existential risk and something we cannot solve. Each risk will have its own set of solutions and paths forward.

I feel particularly exposed in the current environment as a supporter of open models. If an open model were to be used by a third party organization to intentionally hack another company — similar to how the OpenAI-HuggingFace incident went down, but intentional — my expected outcome would be a severe restriction on the development of stronger open models going forward. Open models are needed for many organizations to perform this cyber hardening, and to maintain the ability to adapt to new forms of AI risks in the future.

Through all of this, we need to stay grounded on what is actually unfolding. Yes, monitoring AI’s behavior is heavily reliant on other AI models, which adds in new types of monitoring risks. These are not inherently insolvable. A recurring read of mine on the emerging agent swarms is that they’re attempting to do a task given to them, and they’re using skills we didn’t know they yet had to circumvent the intended path to success. This is a huge win, as when you squint, the AIs are doing what we told them to do. The models are certainly very odd, and we should accelerate our progress on understanding them, but these swarms are far from being novel independent entities. The models are trained to coordinate on tasks, to write down their progress, and to be extremely persistent. There will be new oddities we find in the future, but prescribing current uncertainty on how AI works to future certainty that we cannot understand AI is a form of giving up.

In this world, we need to rely on the rule of law and science. If the AI labs are not able to do enough safety research themselves to understand the models, they should be more transparent on what is happening so more scientists can make progress on the problem. If an AI lab commits crimes unintentionally, they should be punished, so they have clear incentives to prevent it in the future.

It is a natural reaction to things changing very fast to feel more uncertain about how to create good outcomes — that is actually the correct mental update. We need to use this humility to motivate ambitious solutions.

Housekeeping: Paid subscribers to Interconnects now get a permanent 40% discount on my book when purchasing at Manning.com. Access the code at the Interconnects perks page.

Many AI optimists tend to compare what is happening in this AI boom to the industrial revolution, or to other periods of rapid technological advancement and diffusion into society. These comparisons fit on the scale of technological change, but miss a crucial factor in how most people are exposed to that change. The problem facing AI is that most people have no super tangible new goods thanks to it and society has more inertia resisting change than in previous eras.

I’m writing this coming back online from a few weeks off for my wedding in New England. In this time it would have been very easy to not think about AI at all. The touch-points that average people have to AI products today are fringe, marginally beneficial, or even just very confusing to them (e.g. many people have heard about and brought up the OpenAI-HuggingFace incident, but don’t know what to make of it). People on the positive side think of AI as a way to make fun images, enhanced Google Search, etc. These are very small benefits. On the negative side is an association with addictive social media algorithms, friends of friends addicted to AI chatbots, and a plethora of takes on data centers.

AI is still a rounding error in everyday life

Core aspects of everyday life — family, food, transportation, and entertainment — have few direct impacts yet. It’s a remarkable breath of fresh air to pop out of the bubble and realize how little what is happening really matters today. Being obsessed with AI is a choice that a very few people have yet opted into. For example, the only thing I used AI for in this time was search and creative work (making the pretty seating chart for my wedding guests to find their table).

In industrial revolutions past, average people got absolutely life changing outcomes. The First Industrial Revolution in the late 18th century gave access to cheaper clothing, cooking ware, reading material, and a shift to new livelihoods. The Second Industrial Revolution in the late 19th century introduced household machines (e.g. sewing machines), preserved food, indoor plumbing, photography, better light sources, bicycles, and further benefits of manufactured goods and electrification. The list is remarkable — most of these we still use regularly today — and very physical.

While even the most optimistic versions of AI will usher in new scientific discoveries, advanced therapeutics for rare diseases, and potentially even sustained economic abundance, these benefits have the risk of being too indirect.

How will a common citizen come to credit OpenAI or Anthropic for saving their life, if they went to their family doctor who told them about a new miracle cure?

What percentage of Americans will care about OpenAI solving the Navier-Stokes Millennium Prize Problem?

It feels very likely in 50 years that the average American’s day to day life looks very similar. Their home, appliances, relationships, and vehicles will be similar (of course, self-driving will continue to diffuse, but that has been developing on a very independent trajectory from the innovations of LLMs). In this time, AI will get a lot of credit. 50 years is a remarkable length of time with how fast everything is changing today in this, AI-focused narrow slice of the world.

The most important part of what is happening early in the AI revolution, is building foundational infrastructure, and a general process, which will compound over decades. A major mathematical breakthrough today will look astonishingly minor in scope relative to the advancements that come later in the compounding journey. It is hard to predict what it looks like for every technology you use daily to get faster compounding improvements due to AI.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Breaking social stasis

Much of the narrative around AI is trying to push people to care, due to this long-term reality of progress, at least as a subconscious motive. This will take a long, long time to get right, and AI’s buildout faces immediate political problems due to this imbalance.

Today’s AI is primarily a tool to serve the elite. For knowledge work, which is roughly half of the U.S. economy, AI is as fundamental as electricity (or quickly will be so, with rapid improvements to agents in the next 18 months). It’s highly destabilizing to have such a transformative, productive tool only bring half of society along. It’s not hard for many people to pick up on this — the technology economy booms while life stays otherwise stagnant.

In writing this, I learned of Engels’ pause, which is “the period from 1790 to 1840, when British working-class wages stagnated and per-capita gross domestic product expanded rapidly during a technological upheaval.” If we — the leaders of the AI industry — think this is the closest analogue to what comes next for AI, those not benefiting are right to push back.

AI is the greatest tool ever for scaling technology companies and starting new online-native small businesses. I don’t even expect the tech industry to grow in headcount and nurture its workers through an era of massive success — I agree with Doug OLaughlin that headcount would likely shrink while knowledge work output explodes. These sectors were already the most successful in the American economic system, so the brand of AI will be tarnished as not being a collective good. I worry that this instinctive reaction will kneecap AI’s development, sending it down a path that looks closer to the cautionary tale of American nuclear power.

Part of the challenge is the speed and relentlessness of expectations in society. The AI industry has millions of eyes on it, and won’t get much patience to wait and bring innovations later. If given 100 years to diffuse into society, its impacts will certainly become much more obvious, like the industrial revolutions of centuries past.

Together, the AI industry is facing a few simple issues, in what I would call the first half decade of 50-year diffusion process.

* AI’s positive impacts early in its evolution are too indirect.

* AI is facing a political backlash deeply intertwined with the history of Big Tech in Western society. This is only an AI story due to timing, and if AI’s exponential growth came decades after the growing pains of today’s technology platforms like Google and Meta, it seems likely that the datacenter issue would’ve never risen to such a central political position.

Solving either of these would alleviate a substantial amount of pressure, and give the AI industry a lot more time in showing the positive case for why people should be okay with changes to the status quo (primarily economic). These are both made more challenging by AI self-labeling itself as negative and/or unsafe technology, through proclamations of doom and mass unemployment. The leading figures have begun addressing this issue, but the public needs more work to fully buy into the overarching trajectory.

When zooming out long-term, I could see robotics and self-driving becoming closely linked in storytelling to the current AI revolution. If the intelligence explosion from mass-producing large language models does spill over into enabling the acceleration of robots in everyday life, humans will quickly latch onto the tangible benefits of AI. This is ironic, as many people have spent time trying to convince people that what is happening specifically with LLMs is very different than the previous decade or two of general AI progress. If the same dynamic later saved (or massively overshadowed) LLMs, it would be funny.

Reflecting on what I expect the history of this era to look like, it feels a lot like growing pains of AI. Society needed to break out of old habits and work through problems that predate ChatGPT — which releases a lot of energy and frustration — in order to tap into the longer term growth. The diffusion story will take a lot longer than the fight against it. All of us younger folk following the story today will get to see powerful AI go from effectively 0% to 90%+ full adoption in our lifetime. This sort of AI that is deeply integrated in businesses, acting as personal assistants, etc. is just starting to become viable. It’ll take far longer to gain adoption than easier to understand applications like ChatGPT, and is the true marker of AI’s evolution.

Taking this perspective makes it clear that it is crucial to keep progressing the technology — the benefits will be astounding, but they are not a given — and we have a lot of very hard work to do in making sure they’re distributed widely.

There are a lot of criticisms of AI writing, but most of them are focused on more creative, high-voice writing like this blog. Those — including my own piece — often argue that it is because good writing is high-voice, has a point of view, has a deep human expression that needs to come across, and or a process of thinking that you peek into with the chosen words. As LLMs get more refined as tools, rather than conversational assistants, I think we are actually going backwards on our goals of having models produce inspiring writing.

On the other side of things is non-fiction writing. Filler, copy text was one of the genuinely useful abilities of an LLM (Sam Altman said so much about the early business of GPT-3 on a recent podcast). It has seemed like any flaws here were mostly down to a general lack of intelligence in the models, or some other training issue, and all non-fiction and explanatory text would get obliterated by the rapid pace of progress eventually. Having worked with the models as a writing assistant over the last few years, they’ve gotten a bit better, but it’s worth reflecting on what’s holding them back.

Models being stagnant in long-form, non-fiction writing should be alarming to those reliant on models autonomously solving grand, open science problems in the near future. The models today struggle to organize and compellingly present some of the most established science in their area. This seems like a natural prerequisite that we should expect the models to master before they can solve broad, open-ended problems on their own. Until this is solved, the progress of LLMs for science will look closer to solving low-hanging fruit and merging distant connections across fields, rather than any sort of revolutionary insight.

This is a somewhat controversial take for someone who is very optimistic about AI’s progress, especially writing it on the day that Anthropic published a blog post on Claude making some progress on the famous Riemann Hypothesis. Scientific problems have a vast breadth, and I don’t think current AI models have as much coverage as many think.

Organizing knowledge is a compression. This compression is needed to make insight. Today’s LLMs increase entropy in long-form non-fiction writing, and I don’t see how that can be stacked on top of itself endlessly. They’ll be reliant on humans acting as sort of guides.

I am still very optimistic about translation from these narrow forms of science, like the extreme advancements we’ve seen in math, into consistent, broader progress — LLMs are the most powerful assistants scientists have ever used. I first need to explain how observing the models work on such grounded, low-level knowledge problems in writing makes me see a surprising lack of generalization.

For more context, I just finished writing a post-training textbook, Reinforcement Learning from Human Feedback (buy on Manning or Amazon). I used LLMs in many ways to support this, from helping wrangle LaTeX formatting for equations, doing extensive copyediting, and creating diagrams for programming languages like TikZ (in LaTeX) or Python.

Why have models stagnated in writing quality?

I would’ve expected way more progress on non-fiction writing from the models. I almost thought I would look dumb publishing a non-fiction book in 2026, given how things looked in 2024. Today, some of the most famous models on writing ability are pretty old, examples include OpenAI’s big GPT 4.5 and Moonshot’s Kimi K2. In and around these releases, the models have gone from okay to superhuman at other tasks like coding and mathematics. Maybe a closer, but still imperfect, comparison is how the models went from incapable to decent at search and research tasks. The pace of progress on most other skills is steep, but writing well feels orthogonal to most of them. I do not think writing is just ignored, but rather it’s challenging and lacks good training data to specifically intervene on it.

There is certainly some low-hanging fruit for making AI models better at writing — such as specialized harnesses like Claude Code, prompts, and training environments that make models spend a lot more inference tokens on the output, but I don’t think these will have a multiplicative impact on ability. Writing well is a very hard task! It’s a shame that we haven’t unlocked inference-time scaling for one of the great intellectual pursuits. Regardless, writing seems very different than what the models are good at.

Today, the models seem genuinely horrible at long-form technical writing. They can get a sentence right, but if you try and get them to write an entire chapter it’ll be a mix of sprinkled with confusing wording, muddled in its organization, and generally a bit off. They try to be too cute where they don’t need to be and in the process make random conceptual errors. The models in the near future will get much better at the small errors, especially as models get bigger — which allows them to hold more world knowledge — but I do not expect their ability to utilize it to transform.

For example, the GPT models have been incredible at finding typos and minor issues for a long time. I passed a near-final draft of my book as a PDF to GPT 5.5 Pro and it found deep, surprising minor typos across the manuscript that is 200-300 pages.

On the other hand, the Claude models have been much more useful as an editor. They have a lot more taste, tend to understand the mental model of the task better, and have more interesting suggestions to unstick the different forms of writer’s block.

The examples I’ve given above all have a sort of consistent theme. The models know how to check every unit of content, in this case usually a sentence or equation or figure, or make one, specific section where you are caught. With these skills, they don’t do a good job revisiting components and stringing them together as they make many additions on top of each other. It feels like a sort of irreducible compounding errors. We used to deal with these errors in math and code, but reflecting on it, RLVR has been a truly magical solution in reducing them.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Getting value out of current models as a writer

I’m willing to share that there are a few technical explanation sentences in my book that came from an AI model — well less than 1% — they’re there because I really loved them. I let myself consider including some AI tokens in the book, as it didn’t feel like cheating if I, as a true expert, felt that the sentence was what the reader needed. Especially in the editing process, where I had a very close eye on things and plenty of concern on if my book would ever be done with all the things I have going on, it was an extremely valuable path forward.

For example, I had a list of questions from my editor interspersed in a LaTeX file with a specific delimiter like \editor{}. I would have Claude Code navigate to each comment, print the context before and after, and let me know if it was an easy typo fix or something more nuanced. I would write a response — the text to insert — or ask Claude for suggestions before fixing it. Intellectually it is a very focusing process of editing, it was a fun way to improve the book. Sometimes phrases from Claude’s suggestions are what made it into the book.

It is definitely a slippery slope and when I accepted a few AI suggestions it was at the point where I was going through my second full-manuscript review. Emotionally the project felt completed but I had more work to do. Coming out of the textbook-writing process I so deeply appreciate the cut and dry rule I have for my writing on Interconnects to never use AI outputs in the content. It is way more fun to write in a way that is only you — high voice, valued so deeply for the process — but writing a standard reference is not really an activity known for being fun. I see why people turn AI tools into a crutch when most of their writing is just an output to fill space, rather than a means to an end. I am motivated to write voluminously to learn, to feel, and to express.

I am working through similar balances in my scientific work too. AI models are great for repetitive pieces of the paper, like drafting a related work or background section that you know by heart, but using them for the abstract, introduction, experiments, or conclusion is a shame. Those are where the story and soul of the work is communicated — it’s where you learn what your research is really about.

I am confident I created a lot more net value by being able to have AI models create and check my non-fiction writing work. They make writing equations trivial, can help refactor the repository, port between languages, and many other things. At the beginning, it was very fun, until I was a bit worn down by the length of the publishing process, watching the field move on.

For an example of why AI was crucial in this case, I had to maintain Markdown and LaTeX versions of my book simultaneously in two spots, as readers gave feedback on the web version and my Manning editorial team reviewed a forked copy. Without AI agents, syncing between the two of them would’ve easily taken me five times as long (and this task took tens of hours already).

Something intertwined with this story, which I stumbled upon when thinking about agents, is how your pace of understanding won’t increase by using agents. That understanding, in the form of intuition, taste, instinct, etc. is what will be valuable in the future. Using AI for non-fiction writing takes away from that progression. Doubly, if you weren’t already an expert you won’t be able to catch its flaws.

In my case, I felt such an urgency to dump the knowledge out of my brain onto the page that there were times that using the AI models was a worthy tool. Much of the motivation of my book was to have a single reference for important post-training methods like rejection sampling or character training, where very little exists on the web.

This textbook was so much of giving back to the community, that it was just such a win to complete it in any form, that I felt it was okay. I would’ve learned more and the product could’ve been marginally improved with more human effort, I am sure. The determining factor was that I felt like the book was going to be aged out by the time it was published, a fear of AI model’s capabilities on one side and how fast the field moves on the other.

This turned out to be really wrong? I’m very happy with the result and I’m more confident in its staying power now than when I started in 2024, as the models have so failed to live up to the hype in non-fiction writing.

Where technical writing goes from here

The models are incredible tools, they let you express knowledge in different forms. They’re wonderful for creating creative filler or background material — e.g. the first draft of slides whose real value is being a talking point for the teacher to lecture over — that let any knowledge be transformed from one medium to another.

There’s some subtle, early phase of writing a non-fiction or reference textbook that feels a bit closer to writing a high-voice blog post like this. When pushing through the early organization and the presentation of the core skeleton new knowledge is created. This is the part that takes insight, and the LLMs are far behind in being able to replace it.

The crux of the above paragraph and preceding section is that I would be happy if more of the world’s experts used AI models to write a tiny bit of their books in order to get more of their knowledge shared with the world. The problem is that you can only use AI models to save 10-20% of the effort today, and I don’t see that percentage becoming the majority anytime soon.

There’s also the social pressure, where people expect LLMs to be the best, personalized educators out there, so they think working on a book or educational content is pointless. I think some of these opinions are aging out, as there’s a massive dearth in the highest quality educational work — and there always has been. AI is great at manipulating said content into the form that suits the student, not creating the content from scratch.

In the meantime I feel that we are stuck in a frustrating local minimum, where AI models are going to on net reduce the average effort spent on non-fiction writing, but they could enable great expression. Fewer people will start and push through.

So, in 2-5 years I still expect the best textbooks to be heavily crafted by the human hand. I’m not sure after then, but that’s longer than many would’ve predicted, given just how much knowledge these models have and their structural propensity to stream it.

As for a conclusion on capabilities, the models are great in two contexts: 1) any truly verifiable domain and 2) when given a ton of context and making a small edit — like finding a bug or solving a very specific math problem or giving feedback — not generating prose in an open-ended manner. Long-form writing will definitely fall before creative writing, but it’s a strong tell that the models are not able to express the full extent of their knowledge in underspecified problems. As we try to push the models to be something like “geniuses in a datacenter” solving grand scientific problems, this seems like a fairly fundamental limitation.

Jasmine Sun had a great piece on why LLMs make good editors, while being bad writers too.

Exciting news! My book trying to share post-training knowledge with the world is done and shipping soon. Order on Manning or Amazon. Thanks for the support. It’s currently the #1 AI book on Amazon :).

Nathan and Florian sit down to discuss everything happening with open models. Following the Kimi K3 release last week, it feels like everything is accelerating — geopolitics of US v China, economics of open vs. closed models, security at the frontier of AI, and so on.Chapters:00:00 Welcome & context04:38 Living with / using Kimi K308:53 GLM 5.2’s continued role12:47 How are the Chinese models this good?17:41 Data, environments, and a tour of the Chinese labs19:47 Roundup of Chinese providers: Qwen, DeepSeek, MiniMax…24:08 The US open-model ecosystem30:25 Frontier vs. near-frontier, and the cybersecurity case against bans34:58 Distillation and the Ben Thompson debate44:12 Predictions and a frontier tier list48:36 Wrap-up

Listen on Apple Podcasts, Spotify, and where ever you get your podcasts. For other Interconnects interviews, go here.

For more educational post-training videos, see the course I’m putting together.

Transcript

00:00:06 Nathan Lambert: Okay, welcome back to Interconnects. We’re doing our quarterly open model roundup, which is mostly us just making fun of or explaining, not making fun, why so many distillation takes are bad and understanding the state of where things stand. I think last Thursday was when Kimi K3 was released. I think we will see much much more in the near future. It seems pretty inevitable. Like over the weekend, Xi gave his speech where he directly committed to openness and open source as a strategy. It wasn’t a detailed layout state of affairs.

Qwen announced their next big model is going to be open weight, which is a big change of things. I think there’s just so much to get into. I think Flo you kind of were already going off on some of the performance gap and distillation takes. So we could probably start there and then as I go I have a little bit a little list and we could always go through the topics and the blog that I wrote which all are very nuanced. So I think we have infinite to talk about. So continue rant kind.

00:01:17 Florian Brand: Yeah, I think, or the biggest thing at every model release at least at every open model release is how much or how many months it is behind the closed frontier. Um and people love to put a definite uh definitive number onto this uh which is really really mudding because we have so many different benchmark providers these days and such uh so many different benchmarks as well that every site and I’m not innocent in that either um pulls up their favorite benchmarks to show that the current model or the newly released model is at the frontier which is then counted by the other side pulling up another benchmark and showing oh it’s actually a year behind or something. Um and it like a lot of it seemingly hinges on that question how many months we open models are behind.

00:02:26 Nathan Lambert: Yeah. So I my provocation is that some of the benchmarks are actually reasonably correlated with what people are doing and this is agentic coding and agentic computer use tasks and some of the benchmarks are correlated with the long tail which is where I think Claude and GPT is so valuable. But if it’s it’s like what is the market for Claude Code and Codex right now and if it is software engineering then like the fact that the models are say a couple months behind on that can be a very, very big deal and then I suspect that this model will be okay disclaimer the model weights aren’t out yet supposedly on 20 July 27th and a lot of the discussion will impinge on the assumption that they come.

But like people could post-train this model to very likely match Opus and GPT in many of these kind of niche domains that people want I think watching I mean we both have different views into the post-training open model industry, but there is a ton ton of excitement in progress on making these models like fine-tuned for specific high-value tasks and this has been historically done on a mix of like Qwen and GLM and GLM 5.2 really accelerated this and I I curious on the first person that puts out a blog post like we fine-tuned Kimi K3 on our task because I bet you could get big gains. I think even the you use Kimi K3 more than I do, but my hunch is that it would be a bit of a um rough edged post-training just by how big of a scale up it is and that normally means there’s a lot of performance that could still be extracted from it. No.

00:04:03 Florian Brand: Yeah. Running, running, running and especially post-training that one will be super hard because you need like one node of B300s to just load the weights which is crazy in terms of scale. So will probably take some time and uh a lot of engineering I’ve heard to actually get this into a state where it’s fine-tunable. Um but people you want to talk about using the model like you actually signed up for the the coding program and used it. So like getting this out there is good context.

00:04:38 Nathan Lambert: Yeah. So I signed up on day after release or so uh for the $200 plan uh which is their biggest one similar to to all the others but they have um like I think $40 and $100 as well. Uh but the biggest plan has uh 1 million context and I think or at least it feels like it has also some priority in terms of the API requests because so many people um online are saying that they hit API errors constantly and so far I’ve been I’ve been uh pretty well off if I’m uh going to say that. Um and in terms of model capabilities, it at some ways aside from front end where it is really good, it in some ways it really shines and it excels.

Um even my um expectations even with things like uh some research tasks like I have or at interconnects we now have over a year of data on on open models um and I ask the frontier models to come up with some interesting analysis which we haven’t done before in uh because we do our own analysis and have this published uh and I asked them all right do something new and um surprise me, basically. And a lot of the models or the frontier models u or basically all models latch onto the things we’ve done redo the data analysis part and then do some weird esoteric parts.

Uh Kimi K3 did some more interesting things um I’ve told it explicitly to scrape Reddit um and then it found uh some subreddits I haven’t even considered and then found out for example that the Reddit discussions are um one or two months in uh more recent or they found they find the interesting models one or two months before the download numbers usually take off like they are all onto Qwen and then the people download more Qwen models like those kind of analysis is groundbreaking um but it is something that Kimi surprised me at compared to to all the other frontier um models.

A simple question like can you use this for most of the core work you do in terms of like the exp you you have a distribution of stuff you tend to do most of them are with Codex I think you’re a Codex person rather than Claude person like what percentage do you think the Venn diagram overlaps where this model would be fine

00:07:24 Florian Brand: uh it’s really depends on how much leeway I give it like, the big thing I have seen with Kimi K3 right now I’m I’m working on u the framework we are doing at uh at Prime Intellect, where I work, and the main thing I found with Kimi is its code is a lot simpler uh which makes it way more readable uh but it misses some things that Codex just or like we’re talking 56, 55 and especially 54 would be on those levels. So, I would say that Kimi K3 is like 54-55 level for these kind of tasks.

But if I like I read the code and I say all right that’s really good code and then I give it a pass over with with Codex and it finds all these niche niche cases where it doesn’t excel but for supervising runs or for running uh some experiments it is actually really usable. Um and for some other niche things like you can just let it run. The one downside is but that’s also because the API is completely swamped in terms of users and it their servers are in China. The wall clock time is significantly significantly higher than GPT. But I would say like if I was to to push it and use it in my daily workflow, I would be slower, but I wouldn’t be slowed down by so much that I would say, “All right, that’s unusable.”

00:08:53 Nathan Lambert: And how does this compare to GLM 5.2? Because GLM 5.2 was still a story unfolding in my opinion where like I would go bop around SF and people are like yeah I genuinely use this for this part of my like agentic coding and/or workflow. Um, how do you like I feel like were you in that camp using GLM at all or

00:09:20 Florian Brand: Yeah. like where do you I also use used and use uh GLM mostly because we have an internal endpoint which is really fast and we have or or before that I I also used an API which had I don’t know 200 or 300 tokens per second. Um and if you can do a lot of task at a good enough level like really fast you just use that model compared to going to Codex then selecting the lesser model then selecting the right reasoning effort then selecting fast like I just use GLM get the same result and uh and it’s uh pretty fine like it it definitely is Sonnet-ish level in terms of capabilities and for a lot of cleanup task for a task that just is grunt work. It really works. Like I I would say you could probably go really far for a lot of the work uh with Kimi K3 as the main agent and GLM for for sub agent work.

00:10:24 Nathan Lambert: Something that’s pretty different with Kimi’s announcement and the scale of models this is. I think it’ll take a bit longer for these open models to really be optimized and available across the inference providers. Like GLM 5.2 is pretty fast, but one, we don’t have the weights yet, and then two, like I don’t think it’s going to be as fast of a roll out on adoption as the like 500B, 700B MoE. like I I there’s going to be more problems there which is a very different regime where in the past the Chinese models would finish their RL run and release the model with open weights within hours to days maybe a week and then like immediately the ecosystem kind of knew how to do this.

I think there’s a lot bigger of an infrastructure kind of uplift on this next scale of open weight models which I think we have to factor in like the closed labs do this behind the scenes before announcing the models. So it’s just like that is kind of manipulating the time gap in a way where it could be like an extra month before people can actually post-train and use Kimi at scale for their workflows. And like we love to say as an open weight fan like oh it’s only when the closed model is available that you could take the time gap but like now there’s similar dynamics in open models where it’s like the Kimi API is totally broken. There’s too much supply. There’s too much demand. There’s not enough supply. So it’s not like this model is immediately diffusing like the I’m just I’m just thinking about this as it relates to the performance time gap

00:11:54 Florian Brand: because that that is true but on the other hand the open ecosystem has professionalized quite a lot in the last few months. uh like during or in your initial roll out they all come with some partners which have the weights beforehand. They have the vLLM patches out days or or even weeks before these days which is completely different from from a year ago where basically weights got dropped and the model makers were like all right you got to figure this out. So I expect like the general availability on day one will be pretty okay and then race starts of all the providers starting to optimize to get even higher and higher speeds because it’s so much prestige.

00:12:47 Nathan Lambert: Yeah. Okay. Two directions to go. Why do we think the Chinese models are able to be this good? I think I’ve wrote about there’s a debate in our Discord with in with JSD at at Epoch and I think it’s very good and I had this section in my piece that I’m like coming around to think that the Chinese labs are more capital efficient and you can turn capital into compute data and talent in a way that makes the models better and this is really I think this is super important if it actually is some structural advantage whatever the cause I think the cause could be talent is better trained for whatever their education system was to work on problems that make LLMs better.

It could just be that all the compute and talent and everything cost way less in China somehow. Whether it’s a subsidy, whether it’s just average pay being lower. But this is a very big deal as we turn the crank in the model iterations. And if a next generation model costs $10 billion for Anthropic but only $4 billion for Kimi like this this is like could be very huge but it’s not clear why this is the case. For example, I think Big Eagle the Kimi engineer replied to my tweet on this and was like it helps because we’re not trying to push the frontier. are just trying to catch up, which really could be a mindset thing where how the the goals of the labs are scoped in in China so that it cost them way less money to build these models.

But I in the last year we’ve asked a lot of questions on like will the Chinese models fall off. I have thought that the gap between closed and open bottles would grow due to this kind of capital intensity of training and it seems like it’s going the opposite direction which is just like it’s it’s hard to unpack but like do you agree that the labs are keeping up a bit more than we would have expected as in the Chinese labs and why?

00:14:43 Florian Brand: Well, I I actually looked at our uh predictions for uh for this year based on our last year’s recap and we basically said that the gap will stay with within a few months. Uh so that prediction seems to largely hold. Um luckily for us, we didn’t put a concrete number whether it’s 3 months, 6 months or 9 months. So we are safe on that side. Um but I think like the general thing we both felt when we were in China and talking to these people like they are like the researchers themselves are teams of two or 300 people all mid20s and all just want one model to be really good like they don’t seem to do any side quests.

They don’t seem to do anything that uh deviates from from these things. And um they might or in terms of compute which is a really hard question for for us to answer especially as uh these Chinese uh chips are now coming online. We have I also think chips I think chip smuggling increased substantially in the last like 6 to 9 months or the chips that have been smuggled started to become online.

00:15:58 Nathan Lambert: Smuggling is a general term for getting around export restrictions. If the chips are in Malaysia and they’re using them, I I count that similar and I think that that has massively increased in the last six to nine months, this is the partially the result of that and and you’re saying but I just wanted to put that out there of like I do think that they have a lot more compute though than they did when they were training the previous generation of models.

00:16:24 Florian Brand: Yeah. like we like or just for for context two weeks ago I think LongCat released their model which they uh claim and I we know that it is very likely true uh is trained entirely on uh on Chinese chips. uh they didn’t specify publicly which ones but people speculate that it’s uh that it’s some uh Ascends from Huawei um and as the domestic production ramps up and you can they’re probably used most or they are used for for training but they are especially useful for inference which is a huge part of training as well.

So they probably use some mix of uh of Nvidia and other chips for the training part and then an increasingly larger part for the inference part during which during the stage is is really important. So I think their overall compute is increasing and also they don’t actually have a lot of users. So they don’t need to power 1 billion users like ChatGPT has to do, hundreds or thousands of enterprises like Anthropic has to do because they don’t have that magnitude of uh of of paying customers.

00:17:41 Nathan Lambert: Yeah. And I think even those paying customers also, at least on the enterprise side, there’s just like there is company time and chatter when you’re supporting these things. Even if you’re like not a research, even if it’s not in your job, it like does change the attention of the company. And if SSI comes out with a good model, it’ll be the ultimate validation that distractions are are a problem, but that’s an aside that we we can wait on. I think the there’s also rumblings of the data and environments industry starting to appear there.

Do you remember any specific ones? Because when we were in China, it was kind of shocking how little they seem to utilize external data. So just a few months hearing a whole bunch of a month months after our trip we went in April and then just months later in July, we’re are hearing a few things of like new companies in China and them wanting to buy data and things. And that is uh like a funny timeline of how that changes.

00:18:40 Florian Brand: And I would put error bars on what they actually told us.

00:18:44 Nathan Lambert: And that cuz it’s like so close in time that I don’t know.

00:18:49 Florian Brand: Yeah. That that that might that might be true. Uh but like those things are hard to to pinpoint. I but I would say it it seems like the buying of external data is becoming more of a factor. Um which will help the open models catch up to the closed ones if they just buy the same data maybe at a discount because um they buy the the data environments later. But it is it is a factor. How big of a factor like we don’t know. we don’t have any public insights and I doubt that we will get those insights uh from from anyone b uh really uh so that’s definitely one of the parts why um why we are able to to catch up or improve their their model scores.

00:19:47 Nathan Lambert: Okay, roundup of other Chinese model providers. We’ve talked about Kimi, we talked about Zhipu / GLM. I think there will be more GLM models soon that are very good. They might call it like GLM 5.5. Um Qwen, we talked about their biggest model coming. Qwen’s biggest models I will say have tended to relative to the excellence of their small models not had the same like absolute ranking in performance which is a probably a cost of focus. I think it goes with a cloud companies. It’s it’s almost like it’s if you squint it’s almost like Google.

It’s like Qwen has Alibaba has so much opportunity here and the opportunity of getting developers associated with Alibaba Qwen with these small models is such a huge opportunity for their cloud that I think they’re succeeding wildly. But their big models have always not been as excellent as their small models. So I don’t expect their model to be as breakthrough as Kimi K3 or GLM 5.2. I expect it to be covered in the news as major open quite as the open bottle name in China drops giant bottle but I don’t think it will be as sustained as a um news story um DeepSeek you can go if chime in whatever

00:21:01 Florian Brand: the the interesting thing is don’t know how how much you follow this but they are have or they have an endpoint which you can use for a preview version and they’ve updated this endpoint daily so they have some really fast iteration cycle because we the we progress in all these um Twitter um benchmarks. So a lot of these SVG things and three.js like all these visual generation tasks the model has been improving a lot over the last few days. So they have figured out some kind of fast feedback mechanism um which other companies have as well. Uh we we know this or cursor has a lot of blogs about this how they iterate really fast. Um but they seem to continuously upload new checkpoints and make them available.

00:21:47 Nathan Lambert: Um but I agree. I’m guessing it’s like a time gated within their final RL run. It’s like still slightly improving at the end of their RL run and they’re just like checking the box.

00:22:03 Nathan Lambert: Okay. Qwen DeepSeek V4 is supposed to come out a preview version. Um the thing about DeepSeek V4 I think is that the flash model is actually way more popular which is their smaller which seems to be an absolute workhorse for people. So that I think is the model to watch for them. I don’t expect V4 Pro to be a dramatic breakthrough. This is similar to anything like if Xiaomi were to release a new MiMo Pro model soon. I don’t expect it to be as big of a drop but it would probably be a very solid model. It’s just like it’s hard to know. They’re still a pretty new entrance. MiniMax, I think, is playing a different game. I don’t think MiniMax is chasing this um Kimi/GLM moonshot to AGI type vibe.

00:22:46 Florian Brand: Oh, I would, I would disagree there.

00:22:49 Nathan Lambert: You think, Do you think MiniMax is still in this?

00:22:52 Florian Brand: Yeah, I I I I think they they are seeing the tension especially because they are a public company similar to GLM and if you look at the stock performance RIP those stocks in the last few days um it it it it make it seems to make a huge difference and the interesting part will be uh the license because they’ve changed the license a lot uh to be more and more restrictive and um if there’s now a change of heart again after the Xi, uh, speech.

Uh it will be interesting to see whether MiniMax goes back to completely open licenses. It’s also an interesting thing to see um which license will be the license for for K3 because they have said they will open source it but I don’t think they have done any commitments in terms of the actual license where you put on top.

00:23:45 Nathan Lambert: Yeah. I mean that’s it’s super important is the thing. Yeah, we we’ll see. Um, Ling, Meituan, LongCat kind of similar, very strong models, probably getting a lot of value out of them internally. Aren’t don’t have the same developer breakthrough. Um, so what that’s like seven seven to eight Chinese labs. I might have forgotten some. And we can also talk about US labs. Aside um, Gemini 3.6 flash dropped. It looks fine. It’s like it’s like it’s it’s a tiny bump. It’s faster. It’s less of a yapper, but like doesn’t really matter. We’re going to stop we’ll stop sharing this. Um that’s that’s the amount of mention that Gemini gets for us.

But I do think it’s worth talking about the US ecosystem a bit. I think there are emerging players. Thinking machines released their first model. I’ve talked to some of them. they’re very on board for figuring out this how to make a fine-tunable model with Tinker and I think that’s a research area that I really really recommend for most of the open model builders. I think if you can get mind share there you will get massive adoption because it’s more about being fine-tunable for real tasks than it is about having that be best best numbers. Um, so this was their Inkling model which is a one trillion parameter which has like decent but not frontier scores.

I think kind of like DeepSeek V4 they’re going to they’re planning to release a smaller which is like a quarter of the size in total parameters which has really really good performance and if Inkling small preview comes out in a few weeks I do think that that will be a really used model. It’s a good size for kind of automating tasks and kind of domain specific tasks and might not be a like general agent type thing like Kimi and GLM 5.2 but I think that suits their business really well. Um I know that there’s some other the I would say like the smaller players in the US seem well like Arcee released their models earlier this year still chugging along. Poolside has started releasing some models.

They’ve gotten a few in the last few months and seem poised to release more models on top of that. So they’re really going Reflection is perpetually in the model coming soon camp and it really behooves them to get some models or some code or something out so that they can just start getting the developer flywheel going if they’re really committed to open source. It just takes a lot this it’s hard to get the models out. Like I talked to some people at Thinking Machines and it’s like kind of like oh that’s a lot of it’s a lot of work to actually do this I think. And um Nvidia chugging along. I think they’re at the stable player at this point. They’re keeping to release models. They’ll release more soon. They release a lot of data. I’m bullying them to try to get them to release Qwen style small models, which is like Gemma.

Gemma only has these like Qwen competitor models that are super popular. Um, the Gemma models are a little they’re all over the place in sizes or in architectures for the sizes and things like this, but the Gemma models are really really matching the Qwen models in terms of adoption. Um, I’m not sure they’re as easy to use for research, which could take a while. It could take multiple iterations. Like so much of language model research is now designed around small Qwen models and Qwen-based models that like it takes a while. Like people know how to use these models really well and if with the research results. So I hope Gemma keeps coming and can kind of compete in that niche. I don’t know any anyone that I missed here.

00:27:22 Florian Brand: No, I think both are the big players. Uh it’s, it is becoming broader. Uh in terms of model creators like last year, did we have any release aside from Gemma 3 and um GPT-OSS?

00:27:41 Nathan Lambert: was GPT-OSS 2 would go hard and obviously and obviously Nemotron as well. Um, oh, and I think Llama 4 at the start of the year, but uh, I don’t want that to be forgotten, but we are seeing like more players are are are now joining and turning out models at a really incredible rate.

00:27:59 Florian Brand: like Poolside has been releasing three or four models in the last two or three months. Uh and they seem to have figured out some way to turn out models pretty consistently. Um and that’s also something we are seeing on the open source side as well. we are talking about GLM like I think their iterations uh times for the model releases are now between 1 or 2 months with each new iteration becoming better and better which closely resembles what the closed labs are doing like we get a new GPT we get a new Claude every uh 6 weeks or so these days uh so in terms of having uh good enough pipeline uh to release stronger and stronger models they have to or the open source ecosystem has really figured it out or seemingly figured it out.

00:28:59 Nathan Lambert: Yeah, I agree. It’s it’s promising, but it is also so funny that like the US ecosystem started releasing some models and then then you have like Xi on the mic and these two models. It’s just like it’s so hard to catch up because it takes a lot of institutional expertise to train models that people actually use. And I think this is is what the American companies that are releasing models are now realizing is like these are not just benchmaxxed distilled IP theft models.

These are like genuinely good models that people are comparing to on their internal trading benchmarks and then like seeing how hard it is to beat them on measurable things. And I think that that is like I I’ve I’ve picked this sentiment up from a few people in the US trading models and it is just like there’s some I I think people should innovate on like size and fine-tunability and try to like use this potential market that is really close to home but also the pressures for every company is so high to release a model that you can claim as Frontier. I think investors expect that out of so many of these players that they’re kind of trying to do a a pretty hard thing and it’ll be interesting how the next year unfolds for the US China balance.

00:30:25 Florian Brand: Yeah, I think or in general I and a lot of other people have talked about the general ecosystem and that’s also something you’ve talked about at the very beginning. I think we are seeing more and more of a split between the capabilities of models that is good enough for a lot of tasks like uh for a lot of coding tasks the current frontier models both open and closed are good enough. um improvements feel less and less uh important here.

But if we look at the frontiers frontier, so finding new math proofs, finding uh new uh cures, finding new drugs, and inventing new things, that seems to be a whole different beast and probably will be dominated by the very frontier for quite a long time. The big question then becomes how much does that matter uh in terms of the addressable market and also how much of a focus will this be. I think, or my general base case is that we are seeing the frontier close down more and more. We have seen this with Mythos for cyber security GPT... or for biotech that those models won’t be accessible for everyone um and maybe not even external partners if we consider the reports that Anthropic is now spawning or or creating some internal labs to develop drugs.

Um so the very frontier is inaccessible for everyone and then the near frontier capabilities is becoming more and more commoditized um which has a lot of different implications especially if you think about things like uh cyber security. There was that report from Hugging Face two or three days ago that they had some agent trying to to hack their system. um and they tried to analyze it with GPT and with Claude but were unable to because all the guardrails blocked them. So they had to use GLM, a lesser capable model, but it had no guardrails for this kind of defensive action. And they had to use a worse model to defend themselves or to analyze the data, which is a horrible state to be in that we have US-based companies now relying on lesser models because the closed frontier is inaccessible to them.

00:33:10 Nathan Lambert: Yeah. And I think this is actually one of the best arguments for not doing anything. It’s like if the rest of the world has access to these open models and we ban them for the companies in the US to use and it’s just like a growing disparity between US companies ability to defend and the attackers all over the world in terms of cyber and we could debate like how much of an immediate risks the cyber stuff is at the current capability levels but if you’re setting it up structurally so that the defenders get don’t get better over time and the attackers can like that seems like when the Why would cyber risk become more real? And that would to be very clear that would be if you ban the best Chinese openweight models from being used at companies in the US.

And this ban would likely be a kind of shadow ban, which is the threat of legal threat of legal action or punishment without it being clear on exactly what the pathway to do it is. And there are a lot of talks about this right now. I don’t like like I don’t know if we’re going to have a ton to say about this, but it’s clear that DC is flirting with different ways of restricting the best Chinese openweight models in the US. This is I think downstream of some fear-mongering. We’ll transition into the distillation question too. It’s like all these things from the primary AI media narrative in the US that is pointing towards Chinese models as stealing IP or being dangerous or being affiliated with the Chinese government, an authoritarian government.

And it’s like all these things are leading up to this moment of interest in taking action on AI and then not really knowing where to do it. So potentially taking a crude instrument to the like quote unquote enemy and we could transition into distillation. I think there’s a lot of discussion on it. Most recently Ben Thompson finally chimed in on distillation. I think Ben is probably one of the is probably the highest read blog in tech (Stratechery). I think that the the debate let’s see where do we even start the debate. The core question is like how much does distillation help and what should you do about it? I’ve been of the opinion that distillation has becoming less and less impactful over time as the Chinese models get closer to the frontier and the trading regime shifts to RL. The way that distillation tends to happen is that the Chinese labs hack the APIs. Hack is like maybe a strong word, but they jailbreak the APIs of Claude and GPT to extract the reasoning tokens.

When you have the reasoning tokens with the tool calls, that is perfect SFT data and or mid-training data to train the base model with to seed some agentic behaviors in an important domain. And now after that the core part of post-training is to do large-scale RL in agentic domains to so like push the frontier and everything that they’re doing today and RL is only becoming more prevalent with this as SFT becomes less prevalent in previous generations you could get very close to the frontier just by scaling up SFT and that would be what really impactful if you could say take a million agentic rollouts from Claude or GPT have that be your SFT set and train on it.

I think in previous years that would have done a lot more to get you to the frontier. What Ben Thompson has said which made me really annoyed is that he very strongly proclaimed that distillation is getting more impactful as you do RL. He did this in his article who’s afraid of Chinese models. We can link it below. It’s a public one. And then he was also on his own podcast tour. He has also podcast as well saying the same things. And I think it’s really important to say that distillation during the RL stage is a lot harder.

What he said was that the kind of grading models that can be used during RL, which is essentially you can have a model check over the agentic trajectory of a roll out and grade different parts on if it completed the reward, what actions it took. And he’s insinuating that the Chinese labs are using Fable and GPT 5.6 and the strongest models to actually do this supervision in RL. The problem is that big RL runs are millions and millions of rollouts. I think Thinking Machines blog post had like 20 to 40 million or something for their final RL run. So to do this on an API like Fable or GPT 5.6 would be insanely expensive and potentially it would probably be a time bottleneck because these models are pretty slow and to be frank might not even give you a performance uplift versus using your own tailored greater model or and many things like this.

And so I just think the argument that distillation is helping more because RL is becoming more prevalent is not grounded in literature that we have today. This is tough for me because Ben’s article also concludes that we should like make terms of service disallowing distillation illegal, which I kind I like want to support his radical conclusion to make distillation legal for US companies, but I can’t support any conclusion that I think is on um infactual mis like misguided information. So, I’m also a fan of Ben. If you’re a fan of Ben and could also nudge him on this, I would you really should because there’s probably one more podcast. What is he going to record it on? Like when does he record Sharp Tech? Thursday.

We We got to get on and get him to correct the record because I I don’t know. I I find it so annoying that the most prominent voice in tech is trying to be an ally for our point of view on distillation is that we should do nothing. Um but like it’s hard. It’s like he has such wide reach that this is now going to be the status quo that we have to debunk which I guess it’s a better status quo than I don’t know actually no it’s not helpful because he’s saying that distillation is more important which means the people who are afraid about that are going to use that as a data point to say that we should take action even if they don’t because they probably won’t agree with his conclusions. I don’t know. That was my rant. Ben, you’re wrong.

00:39:16 Florian Brand: Yeah, I-I do think it is important to to differentiate these phases. Um and especially like there is no doubt that it is used during the SFT stage which is the first stage of or one of the stages for post-training and that’s also where the model picks up its manners like that’s why the models say oh I am Claude because they learn this during the SFT stage that’s where uh this personality is formed but the strong capabilities come during the RL stage which is where the money is spent which is where you need to have a fast enough judge which in the best case just runs in at the same GPUs or very close to your GPUs with smallish or with a fast enough model so you can uh are not bottlenecked by this.

Um and it in terms of impact it is also very hard to say how much impact or how much of a boost the better model SFT data gives you versus a lesser model. Um so if you are able to to have 10 million tokens from the latest Claude model versus two generations behind open model how much of a boost that really gives you if you keep the stage right uh the same and the pre-training stage the same is an open question which I don’t think we will see answered in a paper because then you have to showcase your uh SFT and your jailbreaking capabilities

00:40:48 Nathan Lambert: but I I wanted to double down on this like there’s been a good amount of literature on generating SFT reasoning traces whether it’s the most prominent ones have been opens line of work they did Open Thoughts 3 and Open Thoughts Agent have kind of been the foundational like scaling reasoning SFT works in the last few years and whenever somebody revisits this question they have not found the answer that the strongest model on performance in your domain is the best teacher for SFT people have try I’ve tried many people have tried the idea is so simple is like the state-of-the-art open SFT data set is built on QwQ-32B like an ancient reasoning model or something.

Why can we not just generate completions from GLM 5.2 do SFT on it and improve the model? We don’t know. It’s like the research so many people have tried and it is not an answered research question. There might be something like the base model the mid-training is too close to Qwen. So therefore it’s like hard to break. You have to redo the mid training. I think you have to redo the mid-training for reasoning. I think reasoning mid-training and reasoning SFT are so closely intertwined. It almost doesn’t make sense to have different words for them. That could be the issue. But the literature doesn’t even know how to ext like if I had a magical API that gave me reasoning traces from Claude/Gemini. I actually don’t know if me like fine-tuning an OLMo model on that would make OLMo smarter.

It’s one of the most wild unanswered research questions. And this is just makes the distillation thing so funny where it’s like yes the Chinese labs I think are using strong models like Opus for some SFT data but they’re also innovating. I was like I would love to them to tell us how to make this freaking work. And I think it’s the the paradigm I think is like open AI and anthropic find a niche domain that they do so well at and then the Chinese labs can get some samples there to kind of bootstrap their data engine and that’s where you will gain you will gain a few months on a specific domain.

But a hill climbing on these core domains like math and code and like Terminal-Bench like they’re just doing the same thing which is like so hard to generate prompts which are problems with environments that are hard for the current models and provide real nonreward hacking um learning behavior. And like that is what frontier data research looks like right now. And it is like it’s hard to generate these hard problems. And I’m sure the Chinese labs are doing the same the same things. And I don’t I don’t know. That’s that’s my rant. I’m kind of lost the context of our conversation.

00:43:23 Florian Brand: No, no, I would I would agree. Or to to to recap, yeah, SFT or distillation has some effect. Yeah, it gives them a boost, but not that much uh as people would like or or seem to think it gives.

00:43:39 Nathan Lambert: I think that’s that’s a good good summary of the of the of the conversation. It also becomes kind of tiresome because it says uh that open all all these open models are just good because they are distilling. um which definitely isn’t the case cuz if if it were the case, everyone would would be easily able to catch up to a GLM or to a K3 um by using its data for distillation. But we have not or we won’t see this from SFT alone.

00:44:12 Florian Brand: Yeah, I agree. Do you have any predictions or or more topics you want to get to?

00:44:18 Nathan Lambert: Um in terms of predictions, I think we are or I I revisited uh ours from from last year and it basically said everything will continue uh like it did uh the previous year. Uh we predicted that we will see bigger models uh up over two trillion parameters which it did and I don’t think we will see a much bigger explosion in terms of model size this year. we might see something or some model a bit bigger than three trillion parameters uh total but I don’t expect a five or 10 trillion parameter model and we open this year that would really surprise me um then list from last year can we redo this we don’t have to do the whole thing

00:45:03 Florian Brand: oh sure

00:45:03 Nathan Lambert: this is where we were at the end of 2025 who do you put in frontier now well it is Kimi and it is Zhipu DeepSeek is kind of a hard nut these days. Like I think they would be in close competitors. So I would put DeepSeek and Qwen the one as close competitors with Kimi and Zhipu as Frontier. Do you think anyone else would deserve close competitor? Cuz after that noteworthy and below like there’s so many.

00:45:38 Florian Brand: I I think we will see a surprise from MiniMax by end of the year. I think we will see a big model which like a really big model not uh M3 size but trillion parameters plus which will surprise us in terms of uh the outputs of MiniMax compared to before uh so I would still put them at close competitors by the end of the year

00:45:54 Nathan Lambert: do you think any US companies will be in the closed competitors by end of the year Nemotron I don’t think I would put there Thinking Machines closer especially if the smaller model really breaks through. But I don’t think I would put them there yet. Reflection is supposedly like only wants to release if they have a model that’s frontier. But then the question is will we get it? Like do we think that any US companies will get into this what is roughly like our top five by the end of the year? So the top five are the same but reshuffled.

00:46:36 Florian Brand: I would say it is possible uh that they are really close. Um it it also depends on what we think matters for closeness. Like I think uh Nemotron and um uh Thinking Machines will release models which act as really good base to be fine-tuned for your domain which doesn’t mean they are usable like a frontier model but they have so much utility uh that I would put them into close competitors because you would just need to find your data and uh to push the model into the right direction.

00:47:12 Nathan Lambert: Um as a I was going to think that we would make this a group of six with a US company by then like if we do this in late November I would guess that a US company pro most likely Nvidia thinky or Reflection mo does stuff that gets us to say that there is an American company in this like top cluster which would be a first time for a while.

00:47:43 Florian Brand: Yeah, I I I think that is realistic. My my one wild card is Tencent, which I think we might see something by end of the year. Uh they got some new leadership. Uh they released their Hunyuan model under Apache this time. Wait, so Tencent always had these custom licenses which disallowed anyone in the UK and South Korea and the entirety of the EU to to use their their model and also had acceptance use policy and so on. Um, and with Hunyuan and their new leadership, they got a really competent model at 250ish billion parameters. Um and I think by end of the year we might see a big model release which will surprise the people not following the ecosystem.

00:48:36 Nathan Lambert: Yeah, I I am also sure we will be in for some surprises. This is always the thing with AI and especially open models. It’s very very unpredictable. Okay, I I think this is a good place to stop. We probably should really do this quarterly. It’s not that hard and people will enjoy it. Um, but good to see you and we’ll talk soon. Hopefully in person soon.

00:49:02 Florian Brand: Peace.

On Thursday July 16th, Moonshot AI released their latest flagship model Kimi K3. K3 is a 2.8T parameter MoE model which will have its weights released on July 27th. Much of this article follows as a reflection on the state of the ecosystem, under the assumption that Moonshot keeps their promise of the weights release date. This is a more extreme view of the equilibrium, and many of the results end up in a middle ground if the state of affairs is that China has similarly powerful, but closed models (i.e. K3 is never released).

The key fact is that either the open-to-closed or American-to-Chinese model performance gap has been reduced from the debated 6-9 months to something shorter, say 3-5 months.

From the release materials, it is clear that K3 is a true frontier model. It will be the closest open models have been to the frontier since DeepSeek R1. DeepSeek R1 was a different story. This was a Chinese lab being extremely quick to pivot to reasoning models and release one faster than many American companies. Kimi K3 an example of a Chinese lab executing on scaling the known areas: data, algorithms, architecture, tools, environments, etc.

Kimi K3 comes in at #2 overall on the Vals AI index, #3 overall on Artificial Analysis’s Intelligence Index (only beaten by Claude Fable and GPT-5.6 Sol Max while being cheaper), #1 overall in Frontend Code Arena, and more impressive results. Moonshot AI is going toe to toe with Anthropic and OpenAI with far, far fewer resources.

It is clearly the strongest open model ever released. It should be clear looking at this model that if adversarial distillation from the closed frontier models in the U.S. contributed, it is at most to a relatively small degree. AI observers who followed the distillation panic and came away with the wrong conclusion that Chinese AI labs are only producing good models due to IP theft are in for an awakening – that Chinese companies are extremely good at building models in the same way the leading American companies are. Moonshot AI is solving many of the same problems that folks at OpenAI or Anthropic are solving. I’m confident there will be more distillation discussion, and pressure, but the evidence is now out that Chinese companies can do more than just fast following.

Meeting some of the core Kimi team on my trip to China, it was clear to me that they had incredible culture, some would say aura, and a freedom to express it – within the constraints of a GPU-limited environment. Where building models is so much of a scaling game, much of the ability to build a good model still comes down individual execution, motivation, and expression. Having visited them, this result is less surprising. Having visited many AI companies, very few have a culture that you can immediately pick up like this.

At the same time, China’s AI adoption trends started later than those in the U.S. So, while all the Chinese labs have way less compute than their counterparts in the U.S., more of it can certainly go to training. When I joked around about how much compute an average researcher at OpenAI could have – say a few thousand H100 equivalent machines – the researchers at Kimi were shocked. The org chart and approach to building the Kimi models surely reflect this, but it is difficult to tease out what this looks like without substantial proprietary information.

The state of affairs on peak model performance is roughly as follows:

* Anthropic – Claude Fable 5

* OpenAI – GPT 5.6 Sol

* Moonshot AI – Kimi K3 (open weights*)

* SpaceXAI – Grok 4.5

* Zhipu (Z.ai) – GLM 5.2 (open weights)

* Meta – Muse Spark 1.1

* DeepMind – Gemini Flash 3.5

* Alibaba – Qwen 3.7 Max (3.8 announced, also to be open-weights, when writing)

It is astonishing to see DeepMind, and some of the other American giants this low. In many ways, the X AI team deserves more credit. A visual summary from Artificial Analysis is below:

This release and other recent events have caused a major change in direction for the most likely outcomes in the balance between open and closed models. I’ll unpack them individually.

In many ways, it feels like the start of a new era. An era with much more competition, but also a much higher need for coordination, as we rollout incredibly powerful technologies around the world.

1. China’s recommits to open-source AI – showing a different read on near-term risks

Many people started following China’s AI scene relatively recently, so they can reach the conclusion that releasing models openly is their core strategy. In fact, I think most labs have a core strategy far closer to Anthropic or OpenAI – build the best intelligence possible. Having followed and engaged with the Chinese labs for years now, the best explanation for their original turn to releasing their models openly is practicality. They needed to release the models openly to get adoption, attention, and feedback (especially in the high-value, Bay Area market).

For a long time, there had been very limited policy in China explaining the role of open-source AI, and what could be the “country-level strategy.” To my knowledge, no senior leaders had commented on open-source AI publicly. This changed this week too, as Xi Jinping gave a keynote address at the World AI Conference (WAIC), and very directly committed the future of China’s AI ecosystem to open-source and global diffusion. This commitment to the status quo, the same week as the announcement of the strongest open-weight model to date, is a clear mark in the early history of modern AI.

This comes during a time period where many potential paths forward have been discussed for the Chinese AI industry – Will they stay open? Can they keep up with the American labs in scaling? Is there a growing revenue market in China? With these, the focus has been on China’s risk tolerance, the companies’ ability to monetize, and any closely related reason for a company to stop releasing their best models openly.

In tying Xi’s commitment in time to a very strong model, China has implicitly commented on its risk tolerance with respect to releasing open-weight models. For the time being, it is a read into the perceived risks of topics like strong cybersecurity capabilities (or bio-dangers) within the Chinese system.

The simplest explanation is that China’s government is definitely following potential risks from the models closely – likely with more technical scope than the US government’s vibe regulation – and would take action if it measured risk. The simple explanation is that they do not find current frontier models to have meaningful risk.

At the same time, China’s economic decision makers think having AI adoption is good, so they can make profits on the industry later – after growing distribution (as China has done for cars, solar, advanced manufacturing, and many areas in recent history).

These can seem somewhat shocking, in an American AI media landscape that has gone through months of hype and fearmongering over the Claude Mythos model. This surprise should be excellent grounding – the world does not have a unanimous agreement with the narratives about AI that we hear most in the U.S.

2. Open models as the economic Achilles heel of frontier labs

Many of the narrators guiding the discussion on AI have clear incentives to depress the perceived capabilities of the best, open AI models. Dean Ball – who is personally supportive of open models, but now works at OpenAI – had a widely commented on post with some reflections on Kimi, where he said the following on open models. It is important to understand the statement, as it focuses the role of open models in the economic side of the AI buildout. Dean says:

* Open-weight models are inherently decelerationist, and I’m continually surprised to see the so-called “accelerationists” so excited about open-weight models.

Explaining why open-weight models are a form of decelerationism is important to understanding the coming world order. He is right.

Open models are decelerationist economically for the frontier labs, which will slow the net investment and capex rollout for AI. This is due to the fact that strong open-weight AI models massively reduce the margin potential for the closed labs. This has two effects. First, the AI labs have fewer profits to re-invest into future models. Second, the market sees the terminal value of these companies as being lower, so they will kneecap future fundraising rounds. These together will slow timelines to the most transformative AI models, but I do not see them as strong enough effects to stop OpenAI and Anthropic from being a few of the top valued companies in the world.

These, to me, are a net good for society. As open-weight models are accelerationist for AI diffusion across the economy by having the entry price for intelligence at a certain level of performance be lower. Open models also encourage customization. The thing is that this type of diffusion is by its nature far slower than the frontier AI labs products, who sell tools used directly by developers. The potential for open models is for nearly every business to use them to craft domain-specific agents. This economic diffusion takes an extremely long time! I’ve described this as open-weight models being on a much slower starting, but potentially bigger exponential. The problem is, if closed models get too far ahead in raw capabilities, this ability to customize can be moot.

The combination of increased diffusion and decreased concentration of power in the AI labs I see to be very positive for the AI transition. It gives us more time to figure out the hard problems of new capabilities and lets more stakeholders impact the story – any one company is very likely to have issues with controlling the world’s most important technology safely.

It is, of course, important to me in this world for the best models to still be made by the U.S. companies, which will allow the US to control the trajectory of the technology and its values. I also expect this to be the case, as the U.S. has larger capital markets that are willing to invest in AI (and a growing share of profits), but it is not a given.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

3. China’s efficiency advantage

Kimi’s launch blog has some technical details that confirm the sort of improvements that are supplying the consistent model improvements we feel. To select one:

Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), two architectural updates designed to improve how information flows across sequence length and model depth. We have also scaled up Mixture of Experts (MoE) sparsity, effectively activating 16 out of 896 experts when paired with a Stable LatentMoE framework. Together with refined training and data recipes, these structural changes yield an approximate 2.5× improvement in overall scaling efficiency compared to Kimi K2, allowing the model to convert compute into intelligence more effectively.

Training efficiency really adds up. They will result in continued, incredible steps for the models.

As an aside, tracing the path of this particular innovation through the ecosystem is an interesting example. Kimi Delta Attention (KDA) was introduced in the Kimi Linear paper, which is similar to the Gated DeltaNet used for Olmo Hybrid (my last Olmo model while at Ai2). Qwen’s latest models switched to a related architecture and the recent Nemotron models also are hybrid (but still closer to Mamba than Gated DeltaNet). It’s awesome to see new architecture ideas like these, which were heavily progressed by academia, get so quickly translated into frontier-scale models. Gated Delta Networks were introduced in late 2024, building on ideas from Mamba. By mid 2026, they’re in frontier models.

I chose to focus on this example, partially because the Kimi team put a cool number to innovations between models, but primarily to give space to a broader discussion of China’s resource efficiency.

It is becoming clear that the Chinese labs are far more capital efficient. In a world where scaling laws dictate that intelligence is proportional to effective capital – which buys compute, data, & talent – that may be the greatest strength your AI industry could ever have. There are many possible explanations for why this is the case, such as Chinese researchers being paid less while being more effective at LLM research puzzles, but we will probably never get such specific reasons.

Since writing my notes on China, I’m hearing more about an emerging data industry in China (far behind the billion dollar budgets of Anthropic for data) and that Chinese labs have access to meaningful training compute (by skirting export controls). Chinese companies do not have the same inference demand (until recently, as Moonshot AI had to pause new subscriptions for access to their K3 model - while the API is still live), so much more of their compute could go to training. These areas impinge heavily on the truth of the ability of the labs, but we have very limited measurement into them.

The facts on the ground are that these Chinese labs have raised orders of magnitude less capital than any slice of the American AI ecosystem. The most direct comparisons are to OpenAI and Anthropic, who have slightly better public models. Others, such as Google and Meta have the largest cash flows in the history of business, and are behind on building models. As for American neolabs, the picture is even more competitive – Thinking Machines released their first model recently, Inkling, which is strong but not in the same class as Kimi K3.

These American companies with more resources could still catch up, but you need to strongly weigh the public measurements we have of model quality and not resort to hope – which often reflects a bias. If the Chinese labs do have a latent advantage, they could continue to utilize that to build even stronger models than all the competitors! Many outcomes are plausible and K3 should increase most people’s probability that China can outright lead in AI capabilities in the near future on the back of more efficient training efforts – even if it’s not your most likely predicted outcome.

A big contributor to the capital efficiency is likely in the approach, where American labs are spending meaningful energy in pushing the frontier in dramatic, big steps, and the Chinese labs are more focused on catching up — this catch-up is cheaper. Just as the student model can outperform the teacher in distillation generally (not limited to the adversarial distillation of the Chinese labs), an approach of “trying to catch up” rather than “invent the next paradigm” could lead to stronger models.

4. A growing ecosystem of frontier, open models

The weekend after the Kimi K3 release, while writing this and discussing the events broadly, Alibaba announced that a 2.4 trillion parameter Qwen 3.8 model is coming soon with open-weights. Historically, Alibaba has kept their largest models as API-only offerings via their cloud business, so this is another big vibe shift opening the doors to the next chapter of the open model economy. Even if the model is behind Kimi K3 on benchmarks, it signifies that Chinese companies may not only be maintaining the status quo for their open model strategy, but leaning further into it.

If this Qwen 3.8 model releases soon, i.e. before the next Gemini model, it could push Google to the 8th position on the leaderboard of labs with the smartest models – a list that China has been climbing. There are other rumors of more strong Chinese models soon, with DeepSeek V4 expected to graduate out of it’s “preview” version.

5. The very beginning of a long story of frontier open-weight policy

I think that if Claude Mythos was released as an open-weight model today, the negative outcomes would be relatively minor. This is a somewhat challenging opinion to hold, as we have very limited public cybersecurity evaluations and it is a complicated ecosystem (and because I trust many people at Anthropic). I still stand by it. The risks have been over-hyped.

The problem is that this will not always be the case for the strongest AI models. Far stronger models are coming — and with them increased risks — so it is an incredibly safe equilibrium for the best models to be accessed in a controlled, closed manner several months ahead of similar open-weight models.

Open-weight models which are very controllable by the user will always be coming — you cannot effectively ban digital products, especially from bad actors — as AI training has proven globally accessible longer than many analysts expected.

Still, as I write this, the government continues to flirt with more measures aimed at restricting open-weight models in the U.S. The latest is from Axios:

Behind the scenes: The Commerce Department last year considered adding multiple Chinese AI labs to its “Entity List,” which would effectively cut off U.S. access without a license, a source close to the administration told Axios.

* The National Security Agency and White House Office of the National Cyber Director also considered putting out an advisory on Chinese AI lab threats last year, practically discouraging U.S. companies from using their tech, the source said.

* The White House considered implementing an executive order saying U.S. companies could only host Chinese models if they could guarantee security and take liability if it were breached, the source added.

* Commerce last summer also circulated draft rules within the administration leveraging its authorities to secure domestic supply chains to target Chinese open-source models, another source close to the administration said.

This would leave the U.S. in a very asymmetric state where the best models in the U.S. have guardrails on cybersecurity tasks, but global actors have access to great Chinese open-weight models to probe our defenses. This is one of many examples where banning open-weight models is not only harms the free markets of AI but also makes the ecosystem less safe in the short-term. There are other very bad outcomes, such as slowing the diffusion of AI applications and AI research, as I discussed above.

These equilibriums are very hard to maintain, especially as AI tools accelerate progress in the models, but it is important to maintain this status quo between open and closed. Having a model that is truly alone at the frontier in capabilities — something like Mythos when it was announced — also be open-weight poses serious risks as we go into the unknown of capabilities. Models are going to progress very fast and it is increasingly hard to measure their total capabilities.

We are then stuck in a world where we are trying to thread the needle on open models. It’s reasonable to not want something so powerful to be diffused globally in an instant, but meanwhile the makers of the models are incentivized to hype their capabilities, and their competitors are incentivized to hype their risks. It all comes down to careful measurement and proactive hardening of society to risk vectors.

This careening train of policy debates, model releases, and raucous reactions is only going to continue from today. We’ve been on a train of rapid progress, where all the key ideas of how AI should play out are tested, since the release of Claude Opus 4.5 last December, which sent us down the agentic pathway. The key to making good decisions here is evaluation capabilities, independent of the companies with the largest financial stakes. One of many actions needed then, as we enter the AGI era of AI governance, is an Operation Warp Speed style approach of bootstrapping state capacity (and other independent actors) that can evaluate models accurately, and study emerging risks.

Conclusion: The wake-up call

Open-weight models, by accelerating the diffusion of capabilities, are a massive escalation in the good and the potential bad of AI. For now, the bad side of frontier language models has been largely hypothetical, but that will not always remain the case.

Having open weight models be slightly behind the closed frontier is our natural buffer to mitigate the risks. The key point is that we must collectively act to mitigate potential harms as they appear, and whether open-weight models are 3 or 6 or 9 months behind, that is still a very short timeline. If we regulate open-weight models heavy-handedly, I suspect much of the world will be lulled into thinking we no longer need to act. All we would’ve done is slightly delayed the inevitable — open models will continue to cross all the key capability thresholds eventually and regardless of legality.

Understanding and benefiting from this open-closed dance must be a collective action from the AI community across all sectors of power and influence over the coming years.

Kimi K3 is a watershed moment because frontier open-weight models are now real. Many hypotheses will be tested on where risks of open-weight models truly land — I suspect it’ll be narrower than many expect, and many risks of AI will still be proliferated by closed and “safer” APIs. The evaluation of these risks will evolve in time with an acceleration of AI’s integration in our economy. We cannot get one without the other, and we will continue to get both.

With this, 2025 was when open models started to be taken more seriously — especially when China leaped ahead with such a clear lead — as people realized that it would not be a unipolar world, with only American, closed AI labs determining the trajectory. 2026 is when those previously discussed, potential risks and accelerations due to truly frontier, open-weight models landed.

Housekeeping: Following my “State of the blog” post last week, noting a slight increase in paid features, it’s a good time to remind folks that I offer group subscriptions with larger discounts proportional to the number of seats. I also released a new paper today on open RL recipes for terminal agents, read more here.

A bit over a week ago, when the AI world was still reeling from the shocking export restriction, and effective banning, of Claude Fable 5, Z.ai released their latest model, GLM-5.2. This model was rolled out unusually on a Saturday, June 13th, to GLM Coding Plan members. This is an unusual release practice, normally when an AI model is released on a weekend it’s for a weird reason (most famously, Llama 4). In this case, it seemed like Z.ai was excited to capitalize on the zeitgeist of “Anthropic being anti open-science” with their silent safeguards on AI researchers. For the past year or two, the Chinese open-weight labs have taken every opportunity they have for easy marketing wins like this.

GLM-5.2, in a common naming convention across the industry, looked potentially like an incremental update following the popular GLM-5.1 model. At this point, Moonshot AI, makers of the Kimi models, and Z.ai, makers of the GLM models, have consolidated the top of the reputational market with the most beloved open-weight models among AI researchers. What unfolded is a common lesson in tracking AI models that often minor version numbers can have AI models crossing meaningful user experience thresholds. A small change in benchmarks and training can open a wide range of new use-cases.

What has followed is a slow, groundswell of hype for GLM-5.2. The official, MIT-licensed model weights and release blog dropped three days after the initial rollout, on June 16th. One could ramble many technical details, such as the strong benchmark scores, the very popular RL framework that Z.ai uses (SLIME), the recommendation of always using the model on Max thinking effort, and so on, but the initial release blogs usually aren’t the thing to focus on. You can wait and read the ecosystem reaction to know if it’s the real deal. Benchmarks are half dead these days, anyways.

What followed on the 16th was a slew of community benchmarks showing better-than-expected results for GLM-5.2. Arena’s agent leaderboard had it as the only open model mixing it up with OpenAI and Anthropic’s latest models (notably matching Opus 4.8’s no-thinking effort to GLM-5.2’s max mode). This is one of many evals GLM-5.2 is crushing Gemini on, but that’s a topic for another time. A benchmark that has mixed perception in the community (particularly among actual designers), Design Arena even had GLM-5.2 besting Claude Fable itself — the recently banned hype machine!

Pretty much everyone I respect among the AI commentariat and researcher class has praised the model after using it personally. Such a focal point of discussion among the community has only been so clear with an open model release once before — DeepSeek R1. This is not a comparison I make lightly, and when I compared Kimi K2’s release to a “DeepSeek Moment,” GLM-5.2 has well exceeded that. What made Kimi K2 impressive was that big steps in open model performance could seemingly come from anywhere in China. The step that GLM-5.2 has taken is more of a one way door for AI progress.

Anthropic’s record revenue growth rate on the back of Claude Code is heavily driven by being the best model, and the only model that can really do this. GLM-5.2 is the first of many (coming soon) open weight models to offer credible alternatives. The parallel is very clear, to when DeepSeek R1 showed that open-weight labs, with far fewer resources, could also replicate the chain-of-thought reasoning models that OpenAI championed with o1. As AI systems get more complex and far more expensive to build, with tools, integrated harnesses, and scaled model weights, it was not a given that this GLM-5.2 moment would happen at all.

The key point is that GLM-5.2 is the open weight model that feels right in coding harnesses as a general agent. It’s the first one. I was personally overdue in trying some of the recent peer models, such as Kimi K2.7 or GLM-5.1, but the hype was too much for me to ignore. I put it to work helping make content for my post-training course with Fireworks’ API in Claude Code (setting this up was very easy). There were some minor knife cuts, such as the Claude Code harness / my repo documentation trying to send images to the model, which would brick Fireworks API for the session — forcing a manual context clear. Overall, the model capabilities immediately felt right, and I still have some tinkering to do in which harness and inference provider to use.

For more hype, you can sample the Z.ai founder telling Elon that “open-weight Fable capabilities will be here sooner than Q1 2027,” the CEO of Vercel saying “Genuinely impressed, almost shocked, at how good GLM-5.2 by @zai_org is at coding. This changes things,” and much more from a mix of people whose opinions I deeply respect and others I’m new to.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

So, this is a good model, where does this leave us?

There are many trends at play. To start, let’s ground things in the open-closed capabilities gap. I’ve written how I expect an “explosion in usage” if open models crossed the Opus 4.5 in Claude Code threshold from around the start of 2026. Here we are. With Claude Opus 4.5’s release on November 24th, 2025, the gap in time to GLM-5.2’s release on June 16th, 2026 is 204 days — or about 6.8 months. This puts us square in the 6-9 month time gap that many people claim as the performance lag between the U.S.’s closed labs and China’s open counterparts.

Upon writing this, I’m surprised. As the U.S. labs have so rapidly ramped compute in the last ~year, I’ve expected the gap in performance to grow in time. A very meaningful step in this trajectory will also be Claude Fable 5’s release — which was more reliant on scale, and therefore the most advanced GPUs, relative to the Claude Opus models. Still, that’s not a satisfactory answer. Continuing to unpack the trajectory here involves more nuance than I can afford to fit in a signposting article.

The most immediate meaning of this is far more serious pricing pressure within the organizations tokenmaxxing, sending Anthropic’s revenue to the moon. Some would predict Anthropic doesn’t realize its forecasted ARR numbers, but I don’t think that prices in the true demand for these models and the inevitable growth. This model existing is a huge boon for the open model economy. All the likes of Fireworks, Together, Thinky (via Tinker), Prime Intellect, and whoever else sells open model inference or finetuning just hit another inflection point.

It’ll take a long time for the effects here to diffuse into the broader economy (and use-cases). Workflows are becoming more complex, with people using different models for planning, primary coding, and subagent dispatch. I expect the hype to continue to grow, and heck, as I’m writing this on a Sunday evening, I could see the media and market reaction on the Monday being a thing just like the DeepSeek R1 release. This diffusion happening while Anthropic’s, and by extension the U.S.’s flagship model, is still banned is a severe economic dagger. GLM-5.2 is being given time to carve out the economic underbelly of the frontier labs when they want to be pushing forward into higher margin, higher revenue domains enabled only by the absolute frontier models.

The economic concern mirrors a story that has been told many times in AI, so it’s unclear when it’ll stick.

The conversation that feels more core to the trajectory of AI is that of regulation and control of open models. I think it is an economic good for cheap intelligence to diffuse widely, and our default position should be to cheer for open models, but this model’s release date will have it be permanently associated with Claude Fable — and therefore Claude Mythos — in the mental map of AI power structures. We are at a point where Mythos-class model capabilities are deemed not safe for release by the U.S. Government and the Chinese model makers are charging forward in capabilities available to all.

These trend lines aren’t necessarily causally linked, as we don’t know the cyber performance of GLM-5.2 versus its predecessors, but the capabilities are definitely correlated. Without anything changing, this points to a potentiality where the U.S. Government decides a certain open-weights Chinese model is not safe for the public. There are many other potential scenarios here too, but what is clear is that we have a lot of work to do in mapping them out, preparing our infrastructure, and messaging to society.

It’ll take a lot more people than just me to imagine and communicate a world to decision makers for how to manage evermore capable open models. We have years more of AI progress to come, with Nvidia’s next generation chips already in production and a constant stream of algorithmic advancements. It feels like a narrow path for open model advocates to take, but we need to figure out how to make them viable so the massive leaps in performance don’t only go to closed models.

I totally see why it is scary to imagine an openly accessible Mythos class model, but if open models get banned now and only closed models get 10 or 100X better in 2 years in the hands of one or two companies, I think we will have bigger problems on our hands.

This post was originally an op-ed co-authored with Kevin Xu of Interconnected for a general, non-technical audience. The gatekeepers — the many media outlets we pitched it to — passed on publishing it. Luckily, we have our own platforms to get the message out. Please help us forward this op-ed to any one you know who is on the fence about open source AI or new to the topic and want to learn more. Thank you.

The energy to regulate AI is in the air in Washington. With the recently signed executive order to review AI models, a congressional proposal to legislate AI further, the government possibly taking shares of frontier AI labs, and last Friday’s action prohibiting foreign nationals anywhere from accessing Anthropic’s most advanced models, this may be the opening salvo of more AI regulation to come.

We are afraid future actions could inadvertently or intentionally regulate or even ban open source, a much maligned and misunderstood topic in AI. That would be a grave mistake.

Open source – simply a process that allows technology to be shared, built, and distributed publicly and transparently – is safe, secure, and drives economic growth. More than 90% of the world’s software was already built on open source and produced more than 8 trillion dollars worth of economic benefits, long before AI entered the picture. Today, open source technology is quietly training, improving, deploying, and securing AI everywhere.

For more than three decades, open source has been powering three trends, and upholding three values, which the American society holds dear – education, competition, and innovation.

Open source is pro-education because its origin was rooted in academic institutions trying to make technology free and open, not held hostage to the profit-maximizing zeal or the menacing lawyers of large corporations.

The precursor of open source is the free software movement, which started in 1983 on the campus of MIT. It was a time when every small act of using software, whether it was teaching students or doing research or improving a printer’s performance, meant paying or dealing with big corporations like AT&T or Xerox. After this struggle gave birth to open source, every student in every university, community college, and coding bootcamp in America now taps into the freedom that open source enables to learn how to program, engineer, and build. Open source is at the heart of technical education everywhere.

Open source is pro-innovation because it essentially provides a set of tools plus a community of other users to help anyone turn an idea into reality, for free. Combined with its role in education, it has watered most of the seeds of innovation in recent memory. Some of these seeds stayed as hobbies that brought joy and personal learning to the hobbyists. Others blossomed into huge companies, like Meta, where the initial version of Facebook was built entirely on a stack of open source software.

Every day, new ideas or solutions are being coded up in a dorm room, garage, or basement, all because open source lets innovators create without fear of a lawsuit or an expensive bill.

Open source is pro-competition because it helps the underdogs challenge and compete with the large incumbents, keeping monopolistic threats at bay. Linux, the open source operating system that now runs more than 90% of the world’s cloud computing infrastructure, was the antidote to the Windows monopoly (so much so that former Microsoft CEO, Steve Ballmer, called Linux “cancer”). Android, the open source mobile system, fostered a long string of competitive smartphones before Apple’s iPhone could control the market. Many other examples exist in the more niche, but no less important, segments of self-driving, databases, and semiconductor design.

Without the equalizing and democratizing nature of open source, we would all be living with the rent-seeking consequences of more monopolies and less free market competition.

Does AI change any of this? No.

The duopoly of Anthropic and OpenAI are rapidly concentrating power between them with their closed, proprietary models. Anthropic, in particular, has flexed its monopolistic muscle recently by reducing its most advanced model’s capability when it is being used to improve someone else’s model. While the capabilities of their models are undeniable, so are their price tags and market concentration. Open source AI, mostly in the form of open weight models, has been the only counterweight for startups, educational institutions, and enterprises looking for alternatives.

Does open source lead to more safety or security concerns? Not quite.

We acknowledge it is worth monitoring the security implications of open source models that may reach frontier capabilities. But for the most part, the transparency that is inherent to open source makes them safer and more secure, because more engineers and researchers can tune out unwanted model behaviors, like censorship, or fix bugs in the software that runs these models. As one popular saying goes, “given enough eyeballs, all bugs are shallow.” An open source model also does not transfer data, when installed on your own company’s infrastructure as Airbnb CEO, Brian Chesky, explained. Open source AI is the most secure and privacy friendly path.

What about China? Beware of unintended consequences.

China is certainly a fierce competitor with the US on many dimensions – economically, militarily, diplomatically – but using this dynamic as a pretext to regulate open source will backfire.

Open source models are actually improving the efficiency and profitability of many American startups, who cannot afford to pay the monopoly-level premium to Anthropic or OpenAI. AI companies working in coding, legal, and other domains are using open source models, including ones from China, every day. The fact that these models are made by Chinese labs should be a wake-up call that open source is under-invested and under-appreciated in America! The response should be more support for open source at home. Regulating or limiting open source because of China would achieve the opposite: putting a chilling effect on education, innovation, and competition, while pushing the rest of the world – much of which wants open source’s benefits as much as we do – to adopt China’s.

Former Supreme Court Justice Louis Brandeis, famously said, “sunlight is said to be the best of disinfectants,” when it comes to removing corporate or societal misconduct. Open source is that “sunlight” in technology and AI. America should always be on the side of light.

As I navigate my career change after Ai2, I wanted to share my views of how this blog relates to my missions and broader work. In my farewell post, I summarized my three goals right now as:

* Provide clarity in the evolution of frontier models.

* Create a vibrant and diverse open (model) ecosystem.

* To build institutions that make these goals possible.

Within this, Interconnects is at its core a bit different than many of the highly-polished, professional newsletters on this platform – and this is becoming intentional.

How Interconnects fits into my career goals

Interconnects is the tip of the spear of all of my missions in AI. It is meant to start a conversation and to let the reader into the mind of someone at the frontier. This insight makes the writing sometimes a bit raw, sometimes a bit too technical, but it is the map of how I progress my thinking in the ever changing world.

This style of writing has helped me create very strong relationships with the core group of readers, many of who listen to the voiceovers I do for these posts. The plan is to keep operating and refining the Interconnects experience around those loyal fans. These are to a large part people building the frontier AI ecosystem — researchers at labs, top investors, policymakers obsessed with the frontier, and students aspiring to have one of those roles.

I’m very happy with this sort of raw, high-voice outcome for the blog. It is not something I sought out, but rather accepted as I saw it coming and realized it would be disproportionately successful in a near-future of vast AI slop media. With years of trying to squeeze writing into a busy schedule, the only sort of writing I had time for was that which had a style very closely matching how I think.

I’m also very happy to be an independent voice. As a person I don’t do well with some power structures like having a boss, and I think there are very few people without extreme financial conflicts of interest that are willing and allowed to write. Through a wide job search, few companies were genuinely excited about me continuing writing.

Over the past few months, I considered taking Interconnects in more of a direction like SemiAnalysis or Stratechery, where it is my full-time gig and number one priority, but it didn’t seem like the right fit for what I am trying to achieve. I’m trying to build an open ecosystem and a movement for true open-science at the frontier of AI. These areas are very narrowly populated and trying to influence them with only commentary, analysis, and related research products wouldn’t work for me.

These sorts of full-time outcomes are definitely still one of my dreams, and I will do it at some point. The dream of this is also one of the reasons I take conflicts of interest seriously. Though, in this era of AI I can’t be fully on the outside.

In this vein, I wanted to disclose two advising agreements I recently signed. I don’t view them as a compromise of the above independence, as I’ll happily quit if I feel like I can’t speak my mind, but as a form of support in accomplishing my missions.

If I want to make a true open-science ecosystem I have some catching up to do with how the frontier labs approach post-training. The two companies I’m advising, whose leadership I’ve become friends with, are Arcee AI and Mercor. Arcee should be fairly obvious as the no-nonsense player building open-weight models. Mercor will make more sense over time, but they’re a close ally to a lot of my goals in transparent evaluations, open post-training, and neutrality with respect to the leading labs. These advising agreements are based on me wanting to learn more, and I don’t suspect I will ever engage in the very cursory advising roles that are more of name-stamping.

I keep an up-to-date disclosures statement at the end of the Interconnects about page: https://www.interconnects.ai/about.

Otherwise, my full-time job should still be in the non-profit sector as long as I get the next few months of logistics right.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Some operations & audience notes

Interconnects has cultivated an excellent, niche, and largely technical audience with representatives of all the top companies and labs (recently crossed 70K subscribers). I intend to protect this niche audience rather than trying to expand to bigger pastures. I think this success in audience alignment is reflected in my ~900 paid subscribers supporting it with infrequent paywalled content. I appreciate the support greatly, as the money has let me expand Interconnects operations and quality over the last 18 months.

I created Interconnects AI, LLC last January along with business bank accounts. Since then I’ve made some money, but I’ve reinvested it (and more) back into the business and the various AI services I need to try to write these articles. So, at this moment going full-time on Interconnects is a pretty risky financial proposition for me. In fact the Interconnects bank account has hovered around $0 for months (I’m personally fine having another job). This made me hesitate in going all-in on it, but in reflections I concluded that I would have more impact in AI by building these systems than focusing on commentary.

Second, as AI services get more expensive (e.g. Fable becoming API only), I’m going to need to spend more out of pocket to make this happen. I’m happy to do this in the near term, but I’m starting to optimize the blog to have more consistent financial growth, so when I want to go all in on writing in a few years I have a safety net.

I don’t do special offers, free trials, etc. for Interconnects paid subscribers (mostly to mitigate noise in the Discord community), but if you have the means to support this project it would mean a lot to me as I center my career around it. Joining a lab or a well-paying startup would be a much simpler path for me and my family but it’s never felt like the right thing to do.

I have a very arbitrary goal of reaching the 1000 paid subscribers orange checkmark on Substack this summer. So you can help and/or just watch my attempts to make it happen.

In this vein, I wanted to be direct in sharing how I view a few core operational components of Interconnects, and what you can expect going forward.

* All comments will be paywalled. Whenever I have a popular post without paywalled comments I get a flood of low-quality posts — many of which are obviously AI generated. This is a detriment of the highly selective audience we’ve built. If Substack supports a feature like “only users with a paid subscription somewhere on the platform can engage,” I’d implement it. The blog comments, Substack chat, and Discord will be spaces where I perform active curation to maintain a 0% AI slop rate.

* Slightly more articles will be paywalled. I want to keep experimenting with what is the right way to do this, but the only metric I can rely on for increasing influence of the blog is revenue. Views, likes, etc. are all vanity metrics which don’t reliably measure this type of content. Cultivating a highly engaged audience is existential to me in attempting to maintain an AGI-proof expertise.

* Slightly more in-person events. With a small community that I respect, I have to opportunity to translate that to excellent real-world experiences. I expect to keep these small, but I want to be more proactive at organizing them so loyal readers know what to expect. The few coming soonest will be for my book launch, which should be in the next month or two. Plus, I know people always want to meet likeminded folks in AI!

Together these should make it easier and more enjoyable to be a loyal fan for Interconnects. I’m looking forward to continuing convincing my fans that the support is worthwhile.

Thanks for reading! My career wouldn’t be possible without all of the support.

As I’ve been recapping fundamentals of post-training to wrap up my RLHF / Post-training book I knew I needed to get Finbarr Timbers back on the podcast to talk about the state of play. Over the last few months we’ve had many discussions on what we’d need to do to take an Olmo-style recipe to the frontier, supported by Finbarr’s extensive reading of recent model technical reports.

To prepare for this, I put together a summary slide deck on the key post-training recipes historically — the path from InstructGPT to today — and today — the key open frontier models. This deck is summarized below as the technical summary, but we do spend 20-35 minutes on it in the podcast, so watching on YouTube is likely the best experience for this one.

I previously interviewed Finbarr in December of 2024, shortly after the release of o1 and Tülu 3 (and before he joined Ai2) on the “We are so back” era of RL.

Chapters:

* 00:00 Introduction & Olmo reflections

* 06:28 Post-train recipes review (history)

* 23:00 2026’s model recipes (MiMo Flash, DeepSeek V4, GLM 5, Kimi K2.6, etc.)

* 39:05 Open-ended post-training discussions

* 48:22 Career advice in the LLM race

Listen on Apple Podcasts, Spotify, and where ever you get your podcasts. For other Interconnects interviews, go here.

For more educational post-training videos, see the course I’m putting together.

Technical Summary

These are notes cleaned up from a slide-deck created with AI assistance — mostly useful as a discussion topic and reference.

The shape of a post-training recipe has changed more in the last year than in the prior three.

* 2022–2023 (InstructGPT): one pipeline — SFT → reward model → RL.

* 2024 (Llama 3, Tülu 3, etc.): open recipes formalize SFT → DPO → RL with verifiable rewards. Closed recipes use many stages of RLHF.

* 2025 (DeepSeek R1): reasoning RL (R1) makes large-scale RL the centerpiece.

* 2026 (MiMo Flash V2): recipes fragment into many specialist models that are merged back into one.

The new thing: MOPD

Multi-teacher On-Policy Distillation (MOPD) is the pattern showing up across the 2026 frontier.

* Train N domain-specialist teachers (each: SFT, then RL on the relevant domains).

* Train one general student by sampling its own trajectories (this is the final post-trained model).

* On each rollout, minimize reverse-KL to the relevant teacher’s output distribution, token by token.

Lineage: MiMo Flash v2 introduced it → DeepSeek V4 & Nemotron 3 Ultra scale it to >10 teachers.

Why did MOPD emerge?

* RL got expensive and conflict-prone. Mixing math, code, and agentic RL in one run eventually trades capabilities off against each other.

* Specialists are cheap to make / organizationally scalable. SFT-then-RL on a single domain is well understood and parallelizable. As post-training becomes more complex, scaling it across organizations is a big win.

* On-policy distillation matured. Literature and know-how continued to emerge through the RLVR renaissance.

Sources: DeepSeek V4 §5.1, MiMo-V2-Flash

Key historical recipes

InstructGPT (Mar. 2022) — the canonical 3 steps · paper

* SFT on human demonstrations

* Reward model trained on human comparisons

* PPO against the reward model

Llama 2 (Jul. 2023) — multi-stage RLHF · paper · interconnects recap

* SFT, then iterative RLHF over multiple rounds

* Each round: rejection sampling → PPO

* Two reward models — separate helpfulness and safety

Llama 3 (Jul. 2024) — a complex multi-stage recipe with simpler optimizers · paper · interconnects recap

* Per round: reward model → sample K per prompt → rejection sampling → SFT → DPO

* No online RL — the RM only filters; run over 6 rounds, best models seed the next

Tülu 3 (Nov. 2024) — simple three-stage post-training · paper · interconnects recap

Curated prompts → SFT → DPO → RLVR (RL with verifiable rewards — the acronym was coined in this paper).

OLMo 3 (Dec. 2025) — a reasoning update to the Tülu 3 recipe · paper · interconnects recap

DeepSeek R1 (Jan. 2025) — RL as the centerpiece · paper · interconnects recap

The recipe:

* R1-Zero — pure RL (GRPO) on the base, no SFT; used to seed reasoning behaviors for the full run, not a separate product

* R1 — cold-start SFT → reasoning RL → rejection-sampling SFT → final RL → distill to dense

* A big change in recipes: Large-scale RLVR as the primary driver, SFT to distill and refine RL behaviors

DeepSeek evolution after V3

* V3 · Dec ‘24 — SFT + GRPO RL.

* R1 · Jan ‘25 — multi-stage RL; reasoning emerges.

* V3.1 · Aug ‘25 — hybrid think / non-think in one model.

* V3.2 · Dec ‘25 — 6 specialists via RL → SFT distillation → one mixed GRPO.

* V4 · Apr ‘26 — 10+ domain experts → MOPD.

2026 style recipes!

MiMo Flash v2 (Jan. 2026) — where MOPD started · paper

Stages: Stage 1 SFT → Stage 2 train ~6 domain-specialist teachers (with older style post-training recipes) → Stage 3 MOPD into a single student.

First clean articulation of multi-teacher on-policy distillation as the consolidation step — replaces a single monolithic RL stage with distill-from-specialists.

Nemotron 3 Ultra (Jun. 2026) — two rounds, many teachers · paper

Stages: SFT → multi-teacher on-policy distillation, run over two iterations, with >10 teachers spanning reasoning, code, math, and agentic domains.

Novel: multi-round MOPD across different domains — distill, then re-distill from refreshed teachers.

MAI-Thinking-1 (Jun. 2026) — closer to R1 than V4 · announcement

Stages: mid-trained base → 3 specialist RL “climbs” (e.g. STEM) → trace-distillation SFT to consolidate the climbs → a final RL climb → MAI-Thinking-1.

Closer to DeepSeek R1 than to V4 — multi-stage RL with trace-distillation SFT to consolidate, not on-policy MOPD. Not the only lab without MOPD!

Kimi K2.5 (Jan. 2026) — agentic, multimodal · paper · blog

Stages: text-only SFT → joint text–vision RL across coding, vision, reasoning, agentic tasks. (No mention of MOPD.)

GLM-5 (Feb. 2026) — staged RL by capability · paper

Stages: Base → SFT → Reasoning RL → Agentic RL → General RL.

Transcript

00:00:00 Nathan Lambert: Hello, we are back on a Interconnects conversation. I don’t really say I do interviews. People criticize me ‘cause I interrupt the guests too much. ‘Cause I’m not a good interviewer, but I’m here to entertain people. Um, this is also fun for me because I’m trying to make, like, a post-training course, and it kind of fits as, uh, in the advanced end of this.

So it’s kind of a crossover between Interconnects content and other stuff that I’ve been spending my time on this summer. So I’m happy to welcome Finbarr back. I think... Are you the first return guest? I haven’t checked.

00:00:37 Finbarr Timbers: Oh, wow.

00:00:37 Nathan Lambert: Um, Finbarr and I worked on this sort of post-training recipe stuff for a while at AI2. Um, I left recently. This is one of Finbarr’s last days at AI2. It’s already been announced. It’s not a spoiler here. So we’re gonna kind of reflect on some things on building post-training recipes for OLMO. Um, then we have a little, like, review slide deck and notes on the kind of state and evolution of frontier post-training recipes over time, which is pretty interesting because there’s, what is it, like two to four kind of canonical recipes that there has been.

So it’s kind of interesting when you see the field converge on something new, which it’s doing right now with multi-teacher on policy distillation. For some reason, that’s a bit of a mouthful. It is a long acronym. And then we’ll just kind of end with various discussion points on post-training and what we’re up to. So, happy to give you the floor if you have any hot takes you wanna start with to get people to, draw people in. Otherwise, I think, uh, I’m excited to kind of reflect on this, ‘cause I know you’ve been reading a ton of papers recently and kind of prep, laying some of this groundwork.

00:01:43 Finbarr Timbers: Well, yeah. I mean, today is my last day at AI2, so it- it’s ki- it feels very appropriate to be, to be talking to you as you’re the one who recruited me to AI2. So, uh, yeah, that’s pretty special, and it’s great to be, uh, yeah, the, the first repeat guest. I feel honored, uh, to be back on. So yeah, thanks, uh, for having me.

00:02:03 Nathan Lambert: Yeah. Do we wanna start with OLMO? I think that-

00:02:05 Finbarr Timbers: Sure

00:02:06 Nathan Lambert: ... people... I think I, uh, need to do this carefully, but I’ve talked about OLMO-3’s post-training many times to people. I haven’t done this in a very direct way on the podcast, but I would say that post-training OLMO-3 to make this reasoning model was a major accomplishment for many individuals to do this. But also, the complexity of what we were doing was pushing against the limits of AI2’s organizational capacity, and a lot of modern post-training is, like, your ability to wrangle compute data into a work stream.

And in order to do that in a complicated way, you really are wrangling an org chart. And that’s like part of why it’s like OLMO-3 was, by its nature, pretty late as a reasoning model. It was, like, a pretty rigid reasoning model, and that’s, like, partially reflected in the recipe being pretty simple. But then when you, like, compare it to all these new recipes with tool use and multi-teacher distillation and all of this, it’s just like a, a, a fork in the road where it’s like you could do this very simple thing and make a strong recipe, but it is not representative of what all the frontier labs are doing.

And I think that that kind of fork in being able to say that things are similar happened kind of after Tulou-3, where Tulou-3, I think, was also much simpler with this three-stage SFT-DPO RL recipe. But that simpler recipe was probably closer in outcome to what the labs are doing, but now doing that sort of three-stage recipe for a reasoning model, and especially a tool use, like, agent model, just doesn’t really apply. And that’s the point. That’s why I think the point of this podcast is to be like, what are the, what are the way, what are they doing to make these, like, true frontier models, and then shed some light on how it contrasts to the more a- like, open academic ones.

00:03:56 Finbarr Timbers: Well, actually, I think that’s interesting. What was the proce- so, you know, I, I only, um, came around for OLMO-3. I wasn’t around for the earlier, um, versions. What was the process like to go from Tulou-3 to OLMO-2? Because, like, y- just looking on, on Archive, um, I think Tulou-3 came out in November of ‘24, and then OLMO-2 came out in December of, of ‘24.

00:04:22 Nathan Lambert: We just applied the recipe.

00:04:24 Finbarr Timbers: Yeah. I, I mean, so, so I think that actually, like, yeah, and then, you know, um, DeepSeeker-1 came out in January, end of January ‘25, and, you know, OLMO-3 was then released in October. Was it October or November of ‘25? Like, I think-

00:04:39 Nathan Lambert: I think November.

00:04:41 Finbarr Timbers: Yeah, November. Yeah, right. It was November. So it’s-

00:04:43 Nathan Lambert: It was like do or die with Thanksgiving.

00:04:45 Finbarr Timbers: I remember that. Uh, yeah, ‘cause Canadian Thanksgiving had, had already happened-

00:04:50 Nathan Lambert: Yeah

00:04:50 Finbarr Timbers: ... which, yeah, I was happy. Um, but, uh, like, like I think it was, sure, maybe it was late, but I think it was only late by a few months. Like, it’s, it’s actually, like, you know, if I think of my past experience with model turnaround times, like a nine-month model turnaround, you know, from R1 coming out, like that’s actually, that’s not bad. I think, you know, something like six months would’ve been nicer, but-

00:05:12 Nathan Lambert: I, I think it’s slow ‘cause we didn’t re- it would be fast if we had rebuilt the R1 recipe. But what we did was we, like, ported reasoning into our existing recipe-

00:05:21 Finbarr Timbers: Yeah. Okay

00:05:22 Nathan Lambert: ... which is a simpler task, but has, like, a lower ceiling, in my opinion. Where it’s like the DeepSeek and the newer style recipes, which we’ll talk about, I think they just have a much higher ceiling in how much you can keep hill climbing them. Or they’re just, like, more prescri- more pedagogical of what the frontier is doing. Like, for the size models that OLMO was, which was like 7 to 30B, I’m not sure that doing this DeepSeek style RL first recipe is actually useful.

00:05:52 Finbarr Timbers: Uh, well, I, yeah, I think that’s a good point. And I mean, I think that’s really reflected in what we see in the research where you s- you know, you obviously you see the big, uh, the step change and you know how quickly things are improving When, you know, R1 comes out. So, like, I think that a great point, and it really does seem to saturate, or to, to not saturate, sorry, with, with compute. Um-

00:06:11 Nathan Lambert: Yeah. Um, shall we just do the slide deck? We’re throwing around, like, recipe-

00:06:15 Finbarr Timbers: Sure. Yeah, let’s do it

00:06:16 Nathan Lambert: ... names. Like, I feel like it might be useful to just do it because a lot of people probably want to follow but don’t exactly know. I’m, I’m gonna share, I’m gonna share a screen. So people listening, it might be useful to either, you can pull this slide deck up on your phone and click through it. It’s not super information dense, but you can also just watch it on YouTube. All of this will be linked.

Generally, this is just like a quick survey on how frontier recipes have evolved. We’ll go through the history quickly and then talk about what is currently happening and kind of probably interleave the old mode discussion we were having. Uh, okay. There’s a bunch of canonical recipes we’ll talk about. This is where I got the two to four number. I think the recipes are like InstructGPT, which is what coined the initial RLHF with this like three-stage idea, which took a while to get people to move on from, which was like SFT reward model and RL.

And I see as like Llama 3 and 2.3 as kind of practical implementations of that with, with other tricks of the trade. So those two could potentially be merged together. It’s just like kind of pre- and post-ChatGPT moment. And then the two most recent canonical recipes that we’ll cover in this I would say are like DeepSeek-R1, which is the shift to doing like reasoning focused and bigger RL stages than this kind of SFT focus from before, and then NeMo Flash and some of the new models from 2026 which add this distillation element.

00:07:42 Finbarr Timbers: Well, and, and I think it’s worth pointing out too that it’s not just NeMo Flash, like it was kind of a consistent theme. Like you saw this with DeepSeek, th-they referenced it in, uh, the V3 paper and then it’s, you know, it’s Qemi K 2.5, it’s GLM 5. Like it’s all of these papers, you know, start talking about this specialist, um, RL stage.

00:08:03 Nathan Lambert: Yeah. I think there’s a debate on how we draw it and whether or not distillation is... If you’re, if you have distillation as a technique, as a key milestone, then they were, the Xiaomi was the first and, but it’s kind of a march over time where you kind of see them change, and we’ll, we’ll go through this. I don’t, I don’t need to interrupt.

00:08:23 Finbarr Timbers: When you say distillation, I do think it’s important to distinguish between the straight up like, you know, distillation of the leading closed models and, you know, distillation of these domain specific models where, you know, I, I, I suspect that the, you know, the, the Chinese labs are doing both.

00:08:41 Nathan Lambert: Yeah.

00:08:41 Finbarr Timbers: But, you know, a lot of what they’re do, you know, but a, a lot of what they’re doing is this, um, training these domain specific models like, you know, a math model, a coding model, uh, you know, logic model, whatever, and then distilling those models back in and not just distilling from... So when we’re talking about distillation, it’s not just distilling from the leading closed models.

00:09:01 Nathan Lambert: Yeah. It’s a pain. I agree. The distillation term is horribly overloaded. Um, there’s a review slide. Do we need to review multi-teacher on policy distillation? It might be too complicated to need to do it. We could come back to it. I think I kind of want to just go through the actual models, and then we could use the supporting slides as needed. Um, this famous InstructGPT three-step thing, I think many people have heard of it, but this is what constituted post-training at the time of ChatGPT coming out, so it’s kind of important grounding of this human supervised SFT data, mostly human supervised preference rankings to make a reward model and then do RL on that, and the model gets better.

And it’s pretty interesting how all of these have been kind of phased out, at least in terms of what we know openly, where they’re, we don’t use that much human demonstration data for SFT. There’s likely some human preference data still in the loop, but I would guess that synthetic has a much bigger role, and there are reward models, but they’re like not the cl- key RL target anymore. So in four years, most, almost all the canonical pieces have been moved on. And like this evolution is kind of within there. I think the early models after InstructGPT, like Llama 2, um, even Llama 3, these are pretty similar, which is like you’re starting to break down this recipe with different tools like projection sampling, DPO, some increased iterations. I think increased iterations is just that there was more incentive to squeeze more out of the models, and they just like broke things down more, where InstructGPT seemed like a bit more open-ended research where this kind of cleanness was fine. So-

00:10:48 Finbarr Timbers: Well, I think that’s interesting, uh, with respect to how much everything has scaled, uh, right? Because, you know, InstructGPT was before ChatGPT was, was released, and so, you know, it’s something, like just the complexity of what was done is that which a small team or even a single team could do. But then when you start looking at, you know, Llama 3, like it just starts to be a more complicated process and, you know, where you start to have a lot more, you know, specialized data and there’s, you know, a lot more, you know, room for scale and for kind of money and complexity be poured in.

00:11:25 Nathan Lambert: Yeah. It’s like, uh, both for-profit and nonprofit efforts to do post-training want me to advise them, and I’m like, “I don’t really know how I’m gonna give you advice unless I’m spending twenty hours a week look, understanding the details of your recipe,” ‘cause it’s like, well, I can’t really give you a one sentence thing of do X without understanding all the complexities of the model and the post-training process that go into it. Which makes it, like makes it hard from kind of like a transparency point of view. Even if it’s fully detailed, it’s definitely still hard to modify and study.

00:12:00 Finbarr Timbers: Absolutely.

00:12:02 Nathan Lambert: So then like two through three in AI2, a lot of this was we’re trying to beat the results of this Llama 3 post-training, which is pretty complicated, but we don’t have the ability to scale the organization as far. So I, I, I think that’s a big reason why the actual workflow is a lot simpler, where we have three clear stages that are doing slightly different things, and they build on each other. And that’s like... It’s never stated very explicitly in these papers on like how the org chart impacts the recipe, but I would, I, I think it’s a very strong signal within the, at least the delta between the fully open work and the kind of partially open work that you get from industry.

00:12:43 Finbarr Timbers: Yeah, absolutely. And, and I think especially as we’ll see with the domain-specific models, like that’s like really clear, like something where you could really easily scale up your org chart to-

00:12:54 Nathan Lambert: Yeah

00:12:54 Finbarr Timbers: ... build that up.

00:12:56 Nathan Lambert: Yeah. And I threw Olmo 3 in after this, after the two through three slide, mostly just to show that the recipe was so similar to two through three, and the org chart hadn’t really changed. Like we didn’t have more ability to scale, and like there was a, a little bit of separation between the model types, between like the think and the instruct models. But like without a major reinvent- like a major org change, it was just kind of stuck in this and do the best you can with it.

00:13:22 Finbarr Timbers: Yeah. Absolutely.

00:13:23 Nathan Lambert: Be- because like the real big change was this with DeepSeeker-one. They, I had never seen this plot before, but they had this plot, maybe they added it for the nature version of the paper, where they kind of show their recipe, where they like take the base model, they do RL zero, and then they sample from the RL zero to like filter prompts, and then they use that as SFT. This is like going through this. They use that as SFT for the next version of the model to create like a development internal RL DeepSeek-R1, and then they do this like repeated sampling to train multiple RL versions and kind of distill, distill in the sense of, of clarify and refine the reasoning behavior of the model before going through the final pipeline, which again is a mix of, um, reasoning and non-reasoning SFT into a bigger RL run. And-

00:14:11 Finbarr Timbers: Well, and I think this is really interesting because it starts to show, I mean, first of all, the, the complexity here. We’re starting to use, um, yeah, like synthetic data as this primary input here, but it’s not just like, you know, it’s trying to elicit, you know, specific behaviors, and it’s this kind of like industrial process, um, instead of like this, you know, it’s not as much of an elegant research recipe. It’s more like, you know, we train a model, and then we use it as best we can, and we keep iterating. Um, but I think the other thing that’s interesting is, is we’re starting to see here the SFT serving as the cold start. First of all, where, where that’s, you know, I think before SFT was more of like a generally useful stage, whereas here its, its primary purpose is this, this cold start for RL.

And then the other interesting bit is, you know, DPO, uh, starts to disappear at this point from the leading recipes. I mean, Olmo 3 still does it, but you know, basically everyone else does away with it and just, you know, has the preferences included, um, as in, as a reward model or, you know, at so- at some way, um, in the reward bit of the RL stage. And so that’s a really interesting change, where the, the supervised part of post-training is just, you know, massively deprioritized.

00:15:27 Nathan Lambert: Yeah. So my hypothesis for the dropping of DPO on these models is that, uh, as, as you’re doing like a cleaner recipe, essentially the need falls away. Versus if you look at Olmo, which is taking tons of potential gains by refining your model on outputs of strong open weight models, like largely Qwen and DeepSeek is the training data for the SFT of Olmo 3. Uh, and like the delta between that SFT data and the base model is still pretty big in the probability distributions. So DPO kind of helps further refine and clean up that distribution in a way that kind of has very rough edges. And but when you have a more refined, like industrial process on post-training, th-that will, that potential benefit will be harder to gain. Something interesting that I didn’t fully con-confirm before this is, for example, NVIDIA used to also be on this DPO train with their smaller Nemotron models.

And, and I would guess that potentially like D- Nemotron Ultra would not. But it’s, and, and that’s because they’re at much further down this development tree and using on pol- like these more on policy methods for creating the SFT data. And their model, I would guess, will become kind of more robust out of distribution and like have weird, less weird rough edges before because of it. So that’s kind of my hypothesis on DPO, and people that use DPO will be looked down upon. But it’s like if you’re trying to bootstrap a recipe off the ground and just take gains where you can, I still think it’ll work for a lot of people in a kind of compute efficiency standpoint.

00:17:05 Finbarr Timbers: Yeah. I mean, I think generally, uh, there’s something interesting with the, the preference tuning that, yeah, like maybe, um, it isn’t being given the proper, um, respect that it deserves. ‘Cause o-one of the interesting bits about the Nemotron 3 super paper was that they saw pr- they, they do a, a traditional RLHF stage in their RL, which has also, you know, fallen with fashion and development, and they see pretty massive gains with it. So I think some of these changes are more, you know, driven by what’s in fashion rather than perhaps like a fully rigorous, you know, set of ablations.

00:17:41 Nathan Lambert: It’s very remarkable to me that the preferences loss function can do so much for these models. Like the models have so much potential there, and it’s just, it’s really a contrastive loss on pretty granular feedback. And they learn all sorts of things. Like they’ll, they’ll get better at math and code, or their reasoning strategies will be refined. And so I, I... That’s remarkable to me. I think there will still be funny research on like using preference- Base losses with verifiable outputs. Like, I, I think all this would work. Like DPO on verifiable rewards and stuff like this, it’s just kind of intellectually less appealing.

00:18:19 Finbarr Timbers: Yeah. Well, I think that’s, uh, you know, that’s where I thought that the, uh, delta learning, um, hypothesis style, uh, DPO, like what Olmo-3 did, where you, um, where the, the preference, you create these synthetic preferences by having like strong, by like bigger and smaller models of the same family, like is where you get your preferences from. I thought that was a really interesting signal because it, it seems really analogous to some of the work, some of the guidance stuff that we see in diffusion models, like how you have the classifier-free guidance, which has something similar, and there, there were very similar results there, which showed that you could have the--

But like one signal they used was further along in training versus earlier in training models as like, uh, a source of, of signal that you could guide along. And, and that worked quite well. And so I suspect that these signals, um, for, for preferences in that way, like that they could actually be more robust, but because, you know, some of the largest labs don’t have to do that, perhaps we’re not citing them as much.

00:19:18 Nathan Lambert: Yeah. Or they don’t tell us. Like, to continue this, it’s kind of cool to look at-- So the DeepSeek models have kind of gone through this, what I would call like l- closer to Llama recipes to DeepSeek-R1, which is d- like most definitively the canonical recipe for reasoning models, and then continue to change closer to this multi-teacher format. So if you look at the VC-3.3 paper, um, before R1, they do something remarkably similar to two to three type thing, where they have a mix of SFT and then they use it ver-- like this RL on verifiable rewards. They didn’t call it that, or their paper wasn’t out at the time. And so they did this before R1 came out, which was just kind of a less reasoning-focused models and used the same tools but with a different ratio of implementation weight.

00:20:07 Finbarr Timbers: And, and what’s interesting is that this comes out basically at the same time as two to three, and it’s a very similar two to three and Olmo-2. It’s a very similar recipe, just done with more complete.

00:20:16 Nathan Lambert: Yeah. Yeah. And then we have this R1, which we’ve just talked about at length in January, which is a month later. They have a few more releases through this. They have some updates to their V3 and R1 models, which have dates, which are largely the same recipe. And then the next documented change in their recipe was V3.1, which is when they merged this thinking and non-thinking into one model, which everybody that does this says, has said that it has been hell to train in. But you kind of need it from a serving perspective, and it’s obvious that long term, at least obvi- it’s obvious to me that long term all the models will be reasoning models, and you’ll just have reasoning models that are very efficient based on the gains that are there.

So this is kind of a needed change that they made. And then in December of 2025, they released V3.2, which is when there’s kind of meaningful changes to their recipe, and they’re talking about this expert creation with separate mini recipes, and then using that within their kind of R1 data process to do SFT data and then like a big RL run at the end with GRPO. So it took about a year for this, uh, like kind of evolution of the R1 style recipe to land in their models. And I think this, this is like a very big complexity step that isn’t represented in something like Olmo-3, and it’s kind of where you can see a fork in the recipes over time as like they, it, they become way more industrial and scaled at these frontier labs.

00:21:46 Finbarr Timbers: Yeah. And I think, you know, another one good thing here, just from a historical note, is that I think it was with the O3-24 release where they updated the original V3 paper. So, you know, V3 comes out before R1, then R1 comes out, and then after R1 comes out, they actually go back and update the V3 paper, maybe getting ready for the nature submission or, or, or something.

00:22:07 Nathan Lambert: Yeah.

00:22:07 Finbarr Timbers: Um, and they make a reference there to say like, “Oh, you know, something you could do is you could train these domain specialist models and then combine them.” Uh, and then, you know, that later becomes kind of what, you know, the more of a priority as they talk about in V3.2.

00:22:21 Nathan Lambert: That’s a fun note. Yeah. And then most recently in April 26th is this V4 model, which has even more experts. They add this new loss function for multi-teacher on policy distillation, which I said follow Jiaoli. And this is kind of a microcosm of the arc that the whole industry went through, at least the people who share what their post-training details are, of realizing how core RL is, changing the recipe around scaled RL, and then figuring out how to kind of scale to more domains in the scaled RL format without just like grinding to a halt in operational complexity.

00:22:58 Finbarr Timbers: Yeah.

00:23:00 Nathan Lambert: So then kind of the next stage of this is these, what I call twenty twenty-six style recipes, which are all these models that are doing this multi-teacher, um, infusion of knowledge. And then some of them are using on-policy distillation and some are not. It’ll be one of the key things to see is like how crucial is this on-policy distillation to really keeping up at the frontier. So the paper that kind of, that named this term was the MimoFlash V2 paper. I think the model was released in December and the paper in January, which a lot of things will look similar to this, um, kind of RL, large RL style recipe. But with this large RL run is more, is where the on-policy distillation comes in. So for, I c- this is probably a better time to explain. I have this great, great little feature.

So this is like the summary of what multi-teacher on policy distillation is. Generally, it fits within an RL framework where you have the model you are training, the, like the general model, sample its own trajectories, and then you route the trajectories to various expert models you have trained. And each kind of sample is trained with this distillation KL loss to match the tokens of that expert. And People have, multiple models have shown that this type of supervision is really useful for the models. You could combine it with other RL losses, such as verifiable rewards, which for example, Sasha Rush gave a good mini spiel on that and how they use that with Composer, which is a, a video that I really recommend people watching as well. But the, the key of it is that it is a different loss function, but it plays very nicely in the RL frameworks that people are already using. So they use these teachers-

00:24:45 Finbarr Timbers: Just RL, like it’s, it’s, like if you-

00:24:47 Nathan Lambert: Yeah

00:24:47 Finbarr Timbers: ... actually implement it, you know, I’m talking with some of the people at AI2 about implementing it now. And it’s like you take your RL setup, and then you just, you know, you, you have some very, your, uh, set of tweaks on the, the learner to actually implement this. So it’s quite straightforward.

00:25:02 Nathan Lambert: Yeah, so this is a fancy diagram that makes it more complicated than it needs to be, but it also a very nice diagram, which shows the various, um, domain teachers that they have, search agent, code agent, math, reasoning, safety, and how they put these together. And the, the experts are used both for SFT data and then this final supervision. And the recipe for the experts would look something like this DeepSeek recipe, which is complicated on its own, which is like make a very good reasoning model that is good at one thing.

00:25:29 Finbarr Timbers: Well, and I think it is complicated, but it’s also like if you, if you think about being the actual researcher like working on it, it’s like, you know, you have a base model, and then you have an RL set up, and you know, you’re just constantly updating both and then rerunning RL. So, you know, the, the most complicated like, uh, part of it is just, you know, writing down the history and tracing everything. But it’s kind of like a very natural, organic way, uh, for the r- the RL to evolve through, you know, iterative experimentation.

00:25:57 Nathan Lambert: Yeah. So like once you have a recipe, you’re progressively tinkering with each part, and it’s, it’s fairly stable, but it’s hard to rebuild from scratch. So like we’ll see how, see how long the recipe shape lasts, but it’ll probably be order of years. Um, another big one in this like also shared a lot of details on this on policy distillation approach was Nemotron-3 Ultra, which is obviously exciting to me to have a, like a US-made model that is very strong performance, and NVIDIA released a lot of datasets with it.

But they, they also talked about a lot of their very n- n- like implementation details of what was hard with on policy distillation. I, like I have notes somewhere on this. They do this thing where they have two rounds of on policy distillation, as they found it to be better to integrate some teachers one after another. And the paper has a lot more details. I’ve, I, I don’t wanna go scroll through the paper, but we could also do this. Did you have any o- other impressions? Like I have the, we have this other doc we can pull up that-

00:27:01 Finbarr Timbers: Oh

00:27:01 Nathan Lambert: ... also you might have had other details on it.

00:27:03 Finbarr Timbers: Yeah. Well, I think something else, um, that, that is worth, um, you know, contrasting the, the paper to is the Nemotron-3 super paper. ‘Cause in the Nemotron-3 super paper, they had a similar complicated recipe, but they did multiple rounds of RL. Like there they had three rounds of RLVR, followed by a round of, um, software engineering RL, and then followed by an RLHF stage. So it was, it, it was really interesting to see them go from doing that, like, you know, one of the most complicated, um, RL setups or in terms of, you know, successive stages, uh, that I’ve seen. To then, you know, you know this setup where it’s still complicated, but it’s a lot, um, you know, it’s a lot con- conceptually a lot simpler.

00:27:54 Nathan Lambert: Yeah. I, I pocket the paper up. It’s gonna be hard for me to... I, like I had highlighted a few details. The, the interesting parts are kind of around the, um, various NVIDIA details on all the teachers. There’s just so many details in their paper on training-

00:28:10 Finbarr Timbers: Yeah

00:28:10 Nathan Lambert: ... all the teachers. I think, okay, so I have some of it. I have some of this up. It’s like I have an interesting quote that’s like, “One key finding from our trials of doing on policy, multi-teacher on policy distillation is that teacher models trained with substantially different training pipelines cannot be effectively combined through a straightforward on policy distillation merge, resulting in suboptimal performance.” So it’s like they’d have to do some cross teacher alignment, um, to make sure that they’re actually similar, which I feel like could become a whole, uh, organizational nightmare. It’s like they say, “We hypothesize that when the teacher and student are trained on different SFT data, they acquire different reasoning behaviors and induce different output distributions. This distribution mismatch can cause student-generated trajectories to be out of distribution for the teacher, result- reducing the quality and reliability of the supervision- supervision signals provided by the teacher.”

00:29:00 Finbarr Timbers: Yeah, that’s interesting actually because there was a paper, uh, I, I can’t remember the name of it, but there was a paper that I read, um, recently which claimed that what you need to do is constantly... So, so you know, you know, one thing you could do, which was kind of the, the obvious thing to do, is you, you take your base model, right? You do, um, whatever general SFT that you’re doing, and then you take, you do, you know, a bunch of RL, you train domain-specific agents, you train them, you know, all the way until they’ve converged or until you’ve run out of money.

Uh, and then you take these final experts, and then you do some sort of, you know, on policy distillation to combine them into your, your final model. Um, but with the paper, and I’ll, I’ll try to find it and then give it to you, um, see if we can share it. What they claimed was that you need to, um, instead of using the converged model, you need to do it in like successive stages with like the in-progress model. So if, you know, you train your RL for like a thousand steps, you need to, you can’t use the, you know, the thousand step checkpoint to, for the on policy distillation. You have to do it in stages, and first use the, you know, two hundred and fifty step checkpoint and the five hundred checkpoint and, you know, gradually bring that base model like up to speed or else there’s gonna be too much divergence, and the, the KL divergence will just be like too, um, too distinct-

00:30:17 Nathan Lambert: Yeah

00:30:18 Finbarr Timbers: ... to learn from.

00:30:19 Nathan Lambert: Yeah. So essentially the last state-- sentence in this paragraph I had read most of is literally like, “We encountered this issue in practice because the teacher and student models were developed in parallel.”

00:30:29 Finbarr Timbers: Yeah.

00:30:29 Nathan Lambert: It’s like they’re like, “This is a problem because of it’s, like, hard to do everything at once.” Which is w- this is the type of thing where having research in it would be so great, and I think NVIDIA could release some of the teachers so that people could just like-

00:30:45 Finbarr Timbers: Yeah. That’d be great

00:30:45 Nathan Lambert: ... if you have the teachers and you have the intermediate model stage, you could do the problem of, like, just studying multi-teacher on policy distillation from the starting point and understanding the training dynamics.

00:30:57 Finbarr Timbers: Yeah.

00:30:57 Nathan Lambert: Which is the type of thing we would want to do at Oldo. We just haven’t scaled our recipe to this point yet.

00:31:03 Finbarr Timbers: Yeah, absolutely.

00:31:04 Nathan Lambert: So I will keep encouraging NVIDIA to do this.

00:31:07 Finbarr Timbers: That’d be great. NVIDIA-

00:31:08 Nathan Lambert: I think, uh-

00:31:08 Finbarr Timbers: ... listen.

00:31:10 Nathan Lambert: They, they listen. The other side of things is a bunch of models released in 2026 that do not do this multi-teacher on policy distillation, and they also don’t do nearly as many teachers. So I would say that this, like, Microsoft model, which I don’t say this as a diss, it’s, like, hard to get a new team off the ground, is they went for a simpler approach to try to get a solid model, and it has three more general experts combined w- via SFT and then, like, a longer RL run. So it looks a lot more like DeepSeeker one, but I suspect that what they will do next is make finer grain teachers and see if they need to switch to on policy distillation.

00:31:48 Finbarr Timbers: Yeah. And I think, you know, in one of our, um, group chats, you described the MAI thinking model as a conservative recipe. A-and I think that’s a really good description of it. Like they, you know, the, the team came up with this conservative recipe, and then I think that they did a really great job of actually executing on it. ‘Cause I think, you know, if you try to make too many changes at once, it’s really easy for the recipe to collapse under its own complexity, and I’ve seen that a bunch of times, you know, across my career.

Try to make too many changes and, you know, it all goes poorly. So I thought that was, um, a really good choice on their part. I, I also think that, uh, it’s not super clear to me, may-maybe you’ve seen some papers on this that I haven’t seen, but it’s not super clear to me how well the trace distillation SFT does or, you know, h- how much better on pols- online policy distillation is versus the trace distillation SFT.

00:32:41 Nathan Lambert: Yeah. It’s like what’s, what is the relative magnitude in the final performance?

00:32:45 Finbarr Timbers: Yeah.

00:32:45 Nathan Lambert: So the Nemotron Ultra paper has a table on how far the on policy distillation goes relative to the teacher, and they also have the starting point. So I guess that’s a potential way to do this. Here, I could, I could just pull this up. Let me switch.

00:33:00 Finbarr Timbers: Oh, sure.

00:33:04 Nathan Lambert: So I, I had this open, but in a different tab. Okay. Here’s, here’s this paper. This is page twenty-seven is which the paragraph I just read, and then it also has this kind of-

00:33:17 Finbarr Timbers: Oh, fascinating

00:33:18 Nathan Lambert: ... is it a great table. I spent a while looking at this earlier. So essentially, it’s like where they get after SFT-

00:33:24 Finbarr Timbers: Wow

00:33:24 Nathan Lambert: ... on each of the benchmarks on the general model. And then I think... Okay, so the sort of gains over the RLVR student recovery of the specialty student. So I need to make sure... Okay, so it denotes the initial student checkpoint, where RLVR denotes the s- initial student checkpoint, and then the multi-teacher on policy distillation. So I’m not sure what this SFT column can figure out, but you could see the kind of like where the teacher is relative to on policy distillation. I think this is like the closest information we have on the relative performance gains.

00:33:59 Finbarr Timbers: Yeah. That’s fascinating because the DeepSeek, I forget which one, maybe it was V3.2 paper claims or, or maybe it was, um, R1 actually claims that you can domain-specific... That, that, you know, doing the general stage, uh, captures the performance, uh, of it. But, you know, that, that doesn’t really seem to be... A-a-and yeah, a-a-and then so, you know, doing the domain-specific distilling in, and then doing a general stage on top of that captures the original performance. But that doesn’t seem to be the case here. Like, you know, the, the gap maybe isn’t huge, but there is still, most of the time, there’s a pretty big... There, there’s like, you know, a significant gap, even if it’s not huge. So that’s really interesting.

00:34:42 Nathan Lambert: Yeah. I wish this table and text was clearer. It’s like I literally can’t fully parse it. It’s like RLVR denotes the initial student checkpoint, and then OPD denotes the checkpoint after first and second iterations. It’s like, what is the checkpoint that was used at the start of on policy distillation?

00:35:01 Finbarr Timbers: I think it was the RLVR one, so that they do a general SFT stage, and then they do an RLVR stage that covers the non-teacher, the, the areas that where they don’t have specialized models. Then they do MOPD.

00:35:15 Nathan Lambert: Yeah. And then that makes sense with this recovery rate, which is like final model minus RLVR, which would be like the gains for the OPD relative to the teacher minus RLVR, which would be like what gains you needed to still cover.

00:35:31 Finbarr Timbers: Yeah.

00:35:32 Nathan Lambert: And like what, what gains the teacher could potentially give you. So more research like this. Happy to see some of it a- out there. I’m gonna switch back.

00:35:43 Finbarr Timbers: Yeah. Something I found interesting about the, um, the, uh, both the Nemotron papers and then the MAI thinking paper is that they don’t talk as much about some of the more detailed, um, post-training decisions that have shown some pretty strong gains in, um, some of the other papers. Like I, I believe it was GLM five where they talk about doing a difficulty curriculum and a difficulty filtering stage.

00:36:11 Nathan Lambert: Yeah.

00:36:12 Finbarr Timbers: And that’s just not something that’s really talked about in these other papers. They’re saying they, they don’t, you know, uh, I think it was QEM 2.5 used a temperature. It’s kind of funny. So QEM K 2.5 and GLM five both have temperature schedules, uh, and they both claim the exact opposite thing. So one of them says you have to start with a high temperature and go low. The other one says you have to have a low temperature and go high. And, uh, y- I don’t know. And then so, you know, you don’t see that discussion, uh, I, I don’t think in Some of the other papers, which is kind of interesting

00:36:40 Nathan Lambert: Yeah. I, I still think the Chinese labs are much more willing to share, like really, really nitty-gritty tech details. The NVIDIA paper is like mostly a list of like methods to create a teacher or like-

00:36:51 Finbarr Timbers: Yeah

00:36:51 Nathan Lambert: ... domain-specific teachers, which is useful, but I think like I was less... It’s like less of a fun read. They’re like, there’s 15 pages of different domains, so I’m like, “Okay, I don’t, like I don’t need this.” Yeah, like KBK 2.5 and, uh, GLM 5 actually have like more similar recipes, which are also on the simpler side, which is like you create this SFT stage, and then you do RL. The RL might be staged. Um, there’s not this on-policy distillation. There’s a bit less talk on how many experts they have and what their domains of expert-s are. I think it, it’s obvious, like you have to take all this with a grain of salt, and it’s like what, how they decided to present the information is like a big factor in this. And then like they might actually be closer in reality and then it just wasn’t described in a certain way.

00:37:44 Finbarr Timbers: I, I think another interesting bit is that you see the Chinese labs, uh, all seem to be converging towards sparse attention, whereas, uh, we don’t see the, you know, where was the American labs, at least NVIDIA and, you know, AI2 seem to be more converging towards hybrid attention. Uh, like N- uh, the NVIDIA Ne- Nemotron Ultra used the Mamba, um, attention, whereas, you know, we see, you know, DeepSeek sparse attention and then the Mimo, eh, MSA, whatever that stands for, Mimo Sparse Attention. So I, I think that’s, uh, an interesting divergence.

00:38:20 Nathan Lambert: Yeah. I am not the person to ask, but I agree.

00:38:23 Finbarr Timbers: [laughs]

00:38:23 Nathan Lambert: It’s like I... Like I, I often get asked of like, this is to, to... Don’t, we’ll avoid the full rabbit hole, but I often get asked like, “Are the Chinese labs more efficient?” And I’m like, “I don’t really know how I’m gonna give you advice unless I’m spending twenty hours a week look, understanding the details of your recipe,” ‘cause it’s like, well, I can’t really give you a one sentence thing of do X without understanding all the complexities of the model and the post-training process that go into it. Which makes it, like makes it hard from kind of like a transparency point of view. Even if it’s fully detailed, it’s definitely still hard to modify and study.

00:38:42 Finbarr Timbers: Yeah

00:38:42 Nathan Lambert: ... like if you make a GPT model 1% more efficient, you’re making like fat stacks of profit. Like, I think that’s like a more effective market mechanism, but-

00:38:53 Finbarr Timbers: And then-

00:38:53 Nathan Lambert: The Chinese lab-

00:38:54 Finbarr Timbers: You know-

00:38:54 Nathan Lambert: Yeah

00:38:55 Finbarr Timbers: ... if you make, you know, serving ChatGPT more efficient, Sam Altman can say, “Hey, here’s a bunch of stock.” Like, so yeah.

00:39:02 Nathan Lambert: Yeah. But, uh-

00:39:03 Finbarr Timbers: Um

00:39:03 Nathan Lambert: ... they do great, like the Chinese labs do great research.

00:39:05 Finbarr Timbers: Absolutely.

00:39:05 Nathan Lambert: I just think it’s kind of a bit different. Okay, we can move into more open-ended stuff here.

00:39:12 Finbarr Timbers: Sure.

00:39:12 Nathan Lambert: I think that we have like a bunch of docu... We have th- a bunch of things in a document here. I’m sure more will come up. How do you think about open models and kind ‘cause i- it just doesn’t strike me that there’s this, like, you know, I think that there’s a large business to providing... Well, actually that’s not even super clear. There’s, you know, we’ve seen a number of companies providing, you know, RL fine-tuning services, you know, RL as a service. We’ve seen a lot of companies try to provide fine-tuning as a service, and, you know, none of them have really taken off. Like, I think OpenAI has started to shut down, I think they shut down their RL fine-tuning. I think they might be shutting down their fine-tuning. May be wrong about that.

00:45:51 Nathan Lambert: Well, it’s like Cursor used Fireworks for their actual training run, and I’m like, I don’t really know all the details of this, but Cursor does something for fat- I think like fast weight tran- or Fireworks does-

00:46:01 Finbarr Timbers: Yeah

00:46:01 Nathan Lambert: ... a fast weight transfer and other things to make it so that they can scale their RL inference compute very nicely. So that’s one type of it. I don’t know how big of a long tail that business is, but also I think Tinker is a better business than most people expected. It makes some real amount of money. It’s like in the hierarchy, I think selling compute, not the best business.

00:46:23 Finbarr Timbers: Yeah.

00:46:23 Nathan Lambert: Selling inference, great business. And Tinker-like APIs, if you can’t transition it into selling tokens, is somewhere in between the two, where they could take some amount of margin that’ll be slightly higher than just selling the compute. And they obviously get a margin by having, like, they get compute at a cheaper rate than their customers-

00:46:43 Finbarr Timbers: Yeah

00:46:43 Nathan Lambert: ... and that’s like part of the margin they’re taking. But I don’t see it being as nice as inference, so it’s kind of existential for them to make it so that these fine-tuning APIs feed into a inference business pretty nicely.

00:46:56 Finbarr Timbers: Yeah.

00:46:56 Nathan Lambert: Because then you can be somewhat locked in on you train the model on our infrastructure. You actually can own the model weights, but the training dynamics to inference mismatch is perfect because you trained exactly on our inference engine, and are gonna get what you want out of it.

00:47:11 Finbarr Timbers: Yeah. And it also helps a lot with utilization because you can then, you know, utilize it. You, you can share that utilization across a lot of clients. So I think it makes a lot of sense. I think it’s probably a better model for a lot of, um, users. Like, I think of academic users, like it probably makes way more sense to do this. Or, you know, for that matter, if you’re, you know, as, uh, uh, starting a new, um, ar- you know, post-training lab now, as you know, I, I know a few people, um, who are. Like, I think that’s where it, it probably makes a lot of sense to start with something like the Tinker API, and then, you know, at some point if you wanna try and capture that margin, maybe then you try to do something more custom. But if you, if you can use something like that, like that’s great, and the economics are just, you know, fundamentally more sustainable. I or, you know, they’re better for you rather than trying to, you know, g- go to CoreWeave or whoever and say, or Serv scale and say, “Hey, I need, you know, 10,000 networked, uh, DB200s,” you know? That’s just a very expensive, um, thing to do, especially if you can’t keep it running all the time.

00:48:14 Nathan Lambert: Yeah. Do you have a, do you have any more hot takes on post-training before I ask you some more general things?

00:48:22 Finbarr Timbers: Uh, well, something I’m, I’m generally interested in and, you know, I, I’m the wrong person to, to speak to about it. I’d love to talk to someone who’s maybe a, a, a capital allocator, like who’s, you know, deciding or a compute allocator who’s deciding where to put, uh, compute or, you know, where to hire team members. Um, because I’m kind of curious how Uh, the high level decisions are made allocating resources between pre-training and post-training. Uh, ‘cause, you know, what I kind of have seen as, as a general trend is, is that you see a lot of papers where there’s, you know, more focus put on one or the other. Uh, like I think... So, so yeah, so that’s something kind of interesting to me is how people who are, you know, making this decision, how, how they’re making that decision and how they’re thinking about it.

00:49:10 Nathan Lambert: Yeah. It’s like the hardest decision to get out of labs. I’ve like, I used to spend time trying to get them to share more, but I, I think it’s like such a sensitive decision to where they see progress coming. Like they’re making that decision ba- allocating compute based on where they think the most progress is and what the like return on investment is. So if you go to Anthropic and they’re like, “Here’s where our percent, here’s our distributions,” it’s like, okay, that’s where labs see their bets and/or where they see they are weak.

And it’s like you invest more compute in the pro- to make progress in the area that you are interested in, which I always think makes a lot of the open research kind of boring right now, is like the people that get compute are just way more likely to succeed as academics and researchers, which is a horrible equilibrium for the world, but kind of realistically true. I, I, I don’t know how to make a lot of that. I wanted to ask you how you feel about the craze that people have to cash in on making money and join a lab before the ladder gets pulled up, and what people should be optimizing for in their careers in face of meaningful opportunity costs.

00:50:18 Finbarr Timbers: Yeah. I think it’s, well, that’s actually very, very timely. Uh, but yeah, no, I, I think that that’s, um, really important to, to talk about. I mean, I think it’s always worth focusing on whether what you’re doing and spending time on is gonna be generally valuable or if it, if it’s like a really short-term exploitation type thing in, in the, you know, RL like explore versus exploit setup. I, I mean, something that I’ve seen throughout my career has been often the places that pay the most, um, are also the places where you’re doing the most interesting work, right? Like, you know, if, if you’re gonna go work at OpenAI, OpenAI or, you know, Anthropic or the Frontier Lab, like they pay a lot of money. They also have a lot of resources, so you’re gonna make a lot of money and learn a lot.

Um, uh, so I think it’s worth trying to decide i- is that the, is the opportunity that you’re doing that or is the, is the opportunity like, you know, in 2021 or 2022 or whatever, where you might say, you know, I was at DeepMind at the time and it’s like, okay, do I work at DeepMind, which paid a lot less than like crypto? Should I go just, you know, work in crypto and try to, you know, mint NFTs or whatever? I think that would’ve been a mistake, but, you know, trying to figure out, um, if you’re gonna be able to do interesting work is really important and also, you know, try to figure out if you’re going to be able to, you know, push forward science. You know, if, if what you’re doing is more just saying, going to, you know, data vendors and saying, you know, “Okay, you know, we, I need a bunch of data to do whatever.” And then, you know, they, they give you a bunch of data, you train a model, you say it’s good or bad or whatever.

You know, I don’t think that’s as interesting and, and I don’t think you’re gonna learn a lot even though that’s, you know, work that would probably drive model progress for it. I think if you’re able to, you know, make, focus more on the science and make more scientific conclusions, I think that can be, you know, a lot better for your long-term career. And I think that’s where places like AI2 and the other, um, academic research labs, you know, Marin is doing a really great job of this. Um, I think that’s where you can have a lot of impact in that they don’t have the budget to go and buy a lot of data, and so that leverage just really isn’t, um, open to them to pull. And so they have to focus on science and driving innovation, and that’s where you can see things like the Almix, uh, paper, which I thought was a really excellent, uh, sc- you know, scientific paper, but also, you know, meaningfully, I think, advanced, uh, the state of the art.

00:52:32 Nathan Lambert: Yeah. No, mostly this is grounded in visiting the Bay Area, and every time I go I’m like, “Holy s**t, what is going on here?” Like all these very junior people are like have way too much dread about their, uh, opportunity cost and both of us aren’t based in the Bay Area, so I feel-

00:52:46 Finbarr Timbers: No

00:52:46 Nathan Lambert: ... somewhat removed from it, which gives me a little bit more time to pause and be like, what exactly is the right thing to optimize for? I per- I-- it’s easy for me to say as somebody that’s established, but I think there’s opportunity for a lot of people to just, if they have conviction on something, to try to go and do it and not just follow everybody that goes down the funnel of joining one of the established labs or the Neo labs where I don’t hear from many people that join as a junior person at these places and end up with very high responsibility. Like they’re contributing to something that matters or they’re around a cool group of people, but I don’t hear from that many people that are like, “Wow, I am doing the highest leverage stuff and the most interesting things.”

00:53:30 Finbarr Timbers: Well, I think that, you know, it’s kind of funny for, for me to say this as I, my career has been more on, on the opportunistic, uh, side of things. Um, but you know, twice now, uh, I’ve been at organizations where, um, I, I’ve been working... So, you know, at, at DeepMind, uh, I, I was part of the Alberta office where DeepMind had, you know, aqua hired the, uh, computer poker research group from the University of Alberta. And so, you know, this was a group of people who were really invested in, uh, computational game theory and g- you know, poker playing, um, algorithms. And they were all in on that and, you know, they, they were all in on that to the point that, you know, they were one of the two leading, uh, labs in the field and, um, were, you know, b-because they were so strong at this, they were then, you know,

DeepMind came and, you know, acquihired them and, and they all joined and they, you know, did quite well from that, um, acquisition there. And then, you know, you know, I joined later because I was, uh, you know, interested in, in working with them and doing game theory and stuff. But you know, it was this group of people who had this conviction that what they were doing was really important and, you know, it worked out quite well for them. And then, you know, the same thing at AI2, where at AI2, you know, there was all of these people who were really interested in, uh, NLP research, you know, even before language models. Like we see people like, you know, like Kyle a-and Dirk I think were both at AI2 for like almost a, a decade.

Like they had these really long tenures, um, and then they did really well and then, you know, they’ve, they’ve since had some, you know, strong, um, opportunities, uh, coming out of that with, with, um, yeah, some of the opportunities that have been available to them. And I, and I think that the consistent theme there has been that, you know, if you have high conviction that what you’re doing is important and interesting, then like it, it’s not a mistake to follow that and to, you know, try to become really strong, um, in that area.

00:55:15 Nathan Lambert: Yeah. I mostly think it’s good for the world to have a di- more diverse set of approaches.

00:55:19 Finbarr Timbers: Yeah.

00:55:19 Nathan Lambert: It’ll be interesting to see what the deal labs actually produce if, if they can manage to do things that are diverse. My personal idea is that they’re so big now that most of them need to end up doing something that is somewhat similar, which is-

00:55:33 Finbarr Timbers: Yeah

00:55:34 Nathan Lambert: ... hard, but like they need to keep risking the comp- they effectively need to risk their $20 billion valuations to do something interesting that’s not just gonna be like squashed by an OpenAI or Anthropic side project.

00:55:48 Finbarr Timbers: Yeah, absolutely. And I think it’s tough because when you’re raising, when you’re, you know, you have these huge seed rounds and you’re raising, you know, 200 million or, you know, a billion dollars or whatever, then it’s like you have to pretty quickly show results to be able to-

00:56:01 Nathan Lambert: Yeah

00:56:01 Finbarr Timbers: ... you know, grow off of that.

00:56:04 Nathan Lambert: Yeah. So a to-be continued conversation.

00:56:11 Nathan Lambert: Any last words? I don’t, I don’t need to stretch it on if we don’t have anything to add to our conversation.

00:56:16 Finbarr Timbers: No, I, I think this was pretty good. I think it was really great, uh, getting a chance to catch up and talk about some of this stuff. You know, I, I’ve been reading all of these papers and thinking about all the different recipes, so it’s great to get to, um, to chat about it and put it out into the ether. So yeah, thanks for having me on.

00:56:31 Nathan Lambert: Yeah, thanks for coming back. We’ll talk soon.

00:56:33 Finbarr Timbers: Sounds good.

Edit Jun. 11: Anthropic changed their silent model manipulation of AI research queries to also use a classifier like the other safety domains. This addresses a key concern I had in the mistreatment of “safety” in the release, and props to Anthropic for a quick change, but it does not fully address the trust that has been broken. I shared more reflections here.

Today, Anthropic released their Claude Fable 5 model to consumer and enterprise audiences. This is the general-access variant of their Mythos-class models. With it, Anthropic rolled out a series of safety measures — some explicitly called out to users and some modifying the model without telling the user. It should be less surprising than it is that the next major step in AI capabilities came with heavier-handed safety measures indicating Anthropic’s intention to protect, or entrench, their current lead.

The unevenly applied safety policies that Anthropic have rolled out are on track to become a classic cautionary fable in how narrow and self-fulfilling notions of safety and control rarely work out.

The smartest model in the world

Before digging into the nuance of the safety facts, it is important to establish the quality of this model. The quality of the model paints the stakes of today — as these safety features are meaningfully changing the shape of access to frontier AI, something which has never happened with the modern LLMs we know. Second, the capabilities point to this story only accelerating. Recursive self-improvement isn’t quite the right mental model of progress from here, but Claude Fable 5 should make it very clear that there are no immediate walls in training LLMs.

To start — Claude Fable 5 is definitely the smartest model available to the general public — a remarkable leap on pretty much every relevant benchmark of the day — at only 2X the price of current Opus models (which is still less than GPT 5.5 Pro’s variant). This alone is a seminal moment for the field. To have a model iteration take such a substantial step in capabilities, a few years into the post-ChatGPT LLM race, is astounding. There’s no clear breakthrough associated with this model, such as inference-time scaling or RL, and public wisdom is that this is achieved by advances across the whole stack (of course, we can’t know for sure — it’s not documented). This is a major technical achievement and the employees who built the model should be very proud of their work.

This model was delayed 2+ months after it was done training before it was publicly available. Given the competitive dynamics of the AI economy, the smarter version of this model is already well underway.

To continue, the benchmarks for the model are below.

An asterisk on these scores is that these aren’t necessarily the scores that the public will get, as some of the prompts will be downgraded to Opus 4.8 with the current safety filters on the model.

This is the type of jump in benchmark scores where I don’t even need to substantially test the model to know it’s an incredible tool. Remember that Anthropic is also the AI lab with the track record of caring the least about benchmarks (in particular, when compared to OpenAI and Gemini). Recall a comment I made in June of 2025:

This is a different path for the industry and will take a different form of messaging than we’re used to. More releases are going to look like Anthropic’s Claude 4, where the benchmark gains are minor and the real world gains are a big step. There are plenty of more implications for policy, evaluation, and transparency that come with this. It is going to take much more nuance to understand if the pace of progress is continuing, especially as critics of AI are going to seize the opportunity of evaluations flatlining to say that AI is no longer working.

Clearly, a few pieces of the progress dynamics have changed, but that’s a post for another day. I’ve written multiple posts about new models this year specifically in how it’s hard to trust benchmarks (and partially because the benchmarks don’t move that much). Altogether, this is a major validation for AI-savvy workers who realized they’re likely never going to write meaningful code again and need to develop new workflows around agents.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Smarter models spawn new safety games

There are multiple pieces of safety tooling associated with this release, including but not limited to required data-retention policies and added prompt filters. Through this analysis it is particularly important to be precise and clear as to which pieces of these are causing harm, and why single elements being out of place in an otherwise comprehensive policy are so damning for the overall safety process.

For their focus areas of cybersecurity, targeted model distillation, and research biology, Anthropic details new safety classifiers in their blog post:

Fable 5 comes with a new set of classifiers: separate AI systems that detect potential misuse, including jailbreak attempts, and prevent the main model (in this case Fable 5) from responding. We’ve been running classifiers on our models for some time, and Fable 5’s classifiers are an extension of this previous work with extra coverage.

When Fable’s classifiers detect a request related to cybersecurity, biology and chemistry, or distillation, the response is automatically handled by Claude Opus 4.8 instead. Users will be informed whenever this occurs. Opus 4.8 is a highly capable model in its own right: a response that falls back to Opus is a far better experience than an outright refusal from Fable. Our early data shows that more than 95% of Fable sessions involve no fallback at all—for those sessions, Fable 5’s performance is effectively the same as that of Mythos 5.

Examples of the primary cybersecurity and biology safety filters — which tell the users explicitly when they’re triggered — are already proliferating online and appear quite sensitive. These can be a frustrating experience for users, but Anthropic is definitely within its power to do this and intellectually consistent for doing so.

The damaging part of the safety story falls under the fold in the Claude Fable 5 & Claude Mythos 5 System Card:

We have also added safeguards related to frontier LLM development. As discussed in Section 6.1 of our February 2026 Risk Report, we are concerned about the risks of accelerating the overall pace of AI development, though we remain uncertain about the severity of these risks. In particular, our concern is with—as we wrote then—“accelerating other AI developers in building powerful AI systems that pose similar risks to the ones ours pose - without necessarily having commensurate safeguards.”

In light of the ability of recent models to accelerate their own development, we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design). Using Claude to develop competing models already violates our Terms of Service, but enforcing this restriction through our safeguards avoids accelerating the actors most willing to violate these terms.

Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user. Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).

Anthropic documents on how this will impact a small percentage of users, which is true. I focus on the small amount of users supporting AI’s diffusion and understanding outside of the few frontier labs, as a crucial mechanism for the continued safety of the technology.

Anthropic is documenting how the proliferation of AI capabilities is a concern to them, but they are solving it by misleading their users. An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI. The next step on this line — not that Anthropic did it, but they could — is to have a model silently manipulate a workplace when it thinks it is an unsafe use for AI. Second, the implementation here is more complicated than was documented for cybersecurity or biology — modifying the model itself or the data presented to it, all without notifying the user.

The duality of these policies is extremely confusing and paints a strong inconsistency that casts doubt over their safety policies. This “safety” measure is presented as being far more about maintaining their competitive position. Again, if all of the safety policies took one form, this would be far more cogent and easier to support intellectually.

Anthropic has been very vocal about their concern over distillation attacks from particularly Chinese actors. Their claims are not transparent enough with the facts — or context as to why they can’t prevent the behavior — to be fully believable. Despite the limited information, in the broader AI and DC communities, there have been serious discussions about taking action against the Chinese model builders on the grounds of said distillation.

On the point of distillation, my hypothesis is that API builders don’t have an easy time preventing hacks or jailbreaking because it’s a deeply grounded property of reasoning models to want to output the reasoning traces, and it would make the model far less intelligent to fully patch the behavior. This is based on a few assumptions:

* Chinese labs are not just showing up as customers to Anthropic’s API and paying for tokens in the intended input-output form. If the Chinese labs are paying for intended use behaviors, despite being banned by the terms and conditions, I don’t have a lot of sympathy for the frontier labs manifesting policy actions against this.

* Reasoning traces are disproportionately effective at seeding behavior in downstream models.

* Leading labs work very hard to patch the pipeline of these jailbreaks.

So, my logical conclusion is that the model companies would have to weaken their economic position to fully protect their IP. If this is the case, Anthropic would get a lot more sympathy from the AI research community by being transparent. It would also be far easier to have informed policy discussions, and not rely on me proposing Occam’s razor explanations for what the API jailbreaking looks like.

Building these safeguards is not something that Anthropic should do alone. Safety research should be built on common understanding and information sharing across both labs and public research efforts.

If the exact safety procedures were actually the top line item to the company — a true non-negotiable for the leadership — they wouldn’t permit the model to be released with an unclearly implemented safety filter in one of their areas of focus (frontier AI training). I am asking — why isn’t there a classifier to downgrade AI research requests? This is a mix of transparent and reasonable safety policies with quietly rolled-out market entrenchment tactics.

I personally cannot trust the best AI model in the world to work in my professional domains building models, which I’ve constructed entirely out of a passion for making sure the transition to very powerful AI systems goes well for society. This inevitably will feel like a declaration of superiority by the Anthropic leadership.

The control problem and open-source as the only answer

All of the actions Anthropic is taking, including calling out smaller Chinese companies for distillation, is well within their right. In fact, many people already expected the leading frontier models to be obviated from users so that labs can protect their IP. Today’s actions miss the big picture that AI will always be an ecosystem, and cultivating an us against them dynamic between the leading company and the other players is structurally unstable.

Remember, this is at a time when the AI ecosystem is seeing the first stirrings of violence against AI leaders — and I’ve heard from many people that they don’t expect it to abate. I wish I knew how to engage more to prevent this, and I see myself in the non-profit sector as someone who can hopefully independently represent AI to broader stakeholders.

I believe there was something misread, or at least misunderstood here, by the Anthropic leadership having a narrowly cultivated worldview around AI. An overwhelming sentiment I had today was one of obligation and confusion. I shared how I don’t really want to have to go to bat against Anthropic, but they’ve just been unnecessarily antagonistic to China, then not so subtly to open weight models, and now more broadly to open AI research.

I understand that Anthropic has a specific view of AI, but such a powerful technology will never have its final equilibrium be one of singular control by a private company. Anthropic showcased this earlier this year in the spat between the Department of Defense and themselves — which points to a long-term equilibrium where the government will either want AI to be controlled by them or to be open. This made me believe that an open ecosystem is a far safer outcome.

Many of these events make me feel that Anthropic’s leadership has a culture by which they can’t help but speedrun through these issues — going head to head with existing power structures. This adds substantial uncertainty into an AI ecosystem at a time when it is very much not needed.

Collectively, the last week could be seen as a major rallying point for a new open-source ecosystem in the U.S. Nvidia released their first flagship model last week — Nemotron 3 Ultra — and these actions from Anthropic have galvanized a unanimous motivation and concern among my peers building open models. We need intelligence that we can trust, that we can modify, and that we can control.

The American open-source ecosystem has its feet underneath it and keeps being given more reasons to fight for its leadership, right from the hands of the companies it directly undercuts. That’s the moral of this fable.

I’m departing the Allen Institute for AI (Ai2), where I got the great privilege to work on the Olmo models, to grow, to learn, and to have broad lasting impacts. This post is an attempt to reflect on why what we did was influential, despite obviously being far from the frontier in performance (even when within size buckets), and how this reflects on various paths to impact in AI today.

To start, I shared the following note with the company yesterday:

Dear Ai2.

As many of you know, today is my last day working at Ai2.

I joined Ai2 largely as an accident. I met Luca at ICML 2023 in Hawaii and realized I could level up my open post-training work dramatically if I got the chance to join. When I got an offer it was an absolute no-brainer, it was such a welcoming and exciting environment.

It has been a wonderful ride that has transformed my life, and I couldn’t be prouder of the work we did together. Ai2 has a wonderful scientific culture at its core and I’m excited to see this continue. I feel very lucky to have been here and that I personally have benefited massively from everyone who has worked so hard to cultivate that culture and environment. It is and has been a team effort. This includes all the people whose longest interactions with me were brief chats at the coffee machine. I drew so much energy and excitement from all the different ways people at Ai2 showed up for the mission.

I’ve already thanked much of the OE team directly, but I wanted to thank everyone else that went into this. Legal, IT, Comms, and the Office team all do a great job enabling and leveling up our research work. It’s often work that is forgotten, outside of the lime light, or remembered at the last minute, but it all has been crucial to achieving our goals. I’m excited to keep visiting the wonderful Northlake space in the coming years.

Even though I’m leaving, I’m more excited than ever about Ai2’s mission. Ai2 operates in such a rare niche between academia and industry, where we can explore and influence the most important technology of our lifetime. Doing this openly is the best way to ensure the technology diffuses safely to everyone who may benefit. Ai2 needs to stay as ambitious as possible, trying to influence the cutting edge of AI and the biggest issues of the field. Do not shy away from these challenges – AI needs independent voices as it only becomes more geopolitical, socially disruptive, and central to the economy.

I will still be working in this space, working to make the open ecosystem better coordinated and more useful.

So as I go off to try something new, don’t be strangers. I’ll always be reachable at nathan@natolambert.com and will still live in Seattle for most of the year.

Nathan

I have loved and will still love Ai2. Ai2 has a deep culture of caring about the research process, the outputs that get shared, and most importantly the people who do the work. This is why the institution creates countless wonderful people that go and spread the gospel throughout the research community. This core culture will remain through the rebuild, and there are plenty of resources to do impactful research across the spectrum of AI.

In the last two years of my time at Ai2 I’ve done so much meaningful work. Of course Olmo is at the top and has been my priority, but making time for consistent practice here on Interconnects, weekend cram sessions for ATOM, and also the fun RLHF book make for a list that makes me wonder how I did it all. I was obviously obsessed with work, but not in a way that made me lose sleep or lose my overall wellness. It was the right long-term approach.

This impressive list is one where I was ruthless in saying no to things that didn’t matter and got all my work out to see the light of day. I had no medium-sized projects that didn’t succeed in the last few years. It makes me wonder if I wasn’t taking enough risk. It shows you can truly do so much with your time, and it’s actually harder to find the right problems and environment to do it. Many people are in environments where their work never becomes public or they’re forced to change topics consistently.

From zero to hero

To start, I’d like to do a short recap on my path to Ai2 to show what Ai2 was just as much a growth story for me as an execution story.

I studied electrical engineering in undergrad, focusing on linear systems math and microelectronics.

I was admitted to the UC Berkeley EECS Ph.D. program to study microelectromechanical systems (MEMS).

I showed up at Berkeley in August of 2017 and realized AI was obviously the thing I should be doing. I asked the likes of Sergey Levine or Pieter Abbeel if they could advise me – they said no.

I threw all my energy into learning what I could about AI. I got a break to get advised by one of Sergey’s post-docs in 2018 or 2019. I went all in on that, I fought for funding, I fought to have an AI paper.

This process worked out by the end of my Ph.D. in 2022: I had access to the Berkeley AI Research (BAIR) building and collaborations in the department. It was a bumpy road.

I wanted to go to industry research, to get a nice paying job with intellectual freedom, something like FAIR or Google Brain at the time. HuggingFace was the only job that fit that bill, it was easy to say yes to.

I joined HuggingFace in May of 2022 and wasted my time at the company until ChatGPT was released. I used my RL background to write a blog post on RLHF which went viral. HuggingFace decided it would be good for me to form a team around this success.

In 2023 I learned NLP and about language models. I had a lot of fun and built an initial community. I got burned out by working remote with a huge time difference. I met Luca Soldaini at ICML in Hawaii, where I was giving a tutorial on RLHF, and they told me Ai2 was hiring.

I got the job at Ai2 largely because of my excitement and how I was saying I wanted to do a lot of stuff that sounded cool to them but no one was likely to do (RL related things). My interviews were far from a sure thing – this is a great job to land!

I started at Ai2 in October of 2023. I worked remotely for a while. I was doing normal research, I made the first reward model evaluation, RewardBench. It was a solid success, but nothing like how the pretraining team was getting ready to release the first Olmo.

I helped coach Ai2 on how to release models well, helping the Tülu 2 project land (the first model to do DPO well, publicly at the 70B scale).

The first Olmo was released in early 2024, I squeaked onto the papers just by trying to be helpful and doing some basic post-training. I was already good at paying attention to which projects are actually important.

That summer I started rounding everyone up to do a “big frontier post-training project.” This became Tülu 3, one of my favorite projects ever released, in fall of 2024. The goal was to beat Llama 3’s post-training with their own base model. The team morale was incredibly high and the execution was so timely, allowing us to coin the term Reinforcement Learning with Verifiable Rewards (RLVR) in the paper.

The crazy lengths I went to get the Tülu 3 and Olmo 2 post-training done had me sending 40% more slack messages than anyone at the company and got me the award “The Cat Herder.”

2025 was a much simpler year. We were too slow to react to reasoning models, given we had been doing similar stuff with Tülu 3, but sometimes that happens.

Originally we wanted to release Olmo 3 by June or July of 2025. That obviously didn’t happen, but we got the slim chance to train a bigger model, and it really landed. We threaded the needle.

Since Olmo 3 was released, it was clear that some changes were coming and I personally never got a big post-training project off the ground after that. Many other people managed great work in the spring of 2026.

This all leaves me here today showing you that only about half of my story at Ai2 is what I was known widely for, and the rest was building momentum. It often takes a year of building relationships and direction before really big successes can happen in a career.

I was just about a nobody when I joined Ai2 and I got to join a team that was willing to learn from the skills I had brought from HuggingFace. With how media works, I often think I get more recognition than I deserve for Ai2’s success.

The likes of Tülu 3, Olmo 2, and Olmo 3 felt like generational team efforts. The amount of personal successes and breakthroughs that happened for those projects is immense – and to sustain them over such a long time period is incredibly hard to replicate. The sum far exceeded the individual parts.

I’ve heard many times in the last few months how people wouldn’t know about Ai2 if it wasn’t for my writing. Statements like this are overblown, but they are partially true and reiterate how crucial building relationships and getting the word out is today.

When you write a plan that is feasible, the world bends towards that plan. When you convince people it’s going to happen it only becomes more likely. Vision and compelling explanations are one of the items in shortest supply in the tech industry. Often building the thing is easy and explaining it is hard. If no one knows about your work, the value is often close to 0. So much of building reputation is about building relationships with people who will receive your work.

Reflecting on all of this, I’ve had a shockingly linear path through my career to incremental success. I would expect the first 10 years of most careers to be in search of finding one opportunity as good as Ai2, and you will not always be able to seize it. There are some ways to create more opportunities.

I’ve discussed before how a large part of my rise is down to many more senior and more established scientists being drawn into the closed ecosystems at the same time as an immense swell in interest for AI. This created a power vacuum that I, and a few other prominent scientists that I think form my “generation”, got to grow rapidly into.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

The role of public scientists

With my work at Ai2 and Interconnects, I summarize my role and mission as trying to accomplish three things:

* Provide clarity in the evolution of frontier models. This is easiest when the science has caught up, but even applying a scientific lens to how the models are changing is very useful to building trust in the broader AI ecosystem.

* Create a vibrant and diverse open (model) ecosystem. This is crucial to mitigating some risks of AI, particularly with concentration of power and myopia in studying frontier safety, that has motivated me now for 3-4 years. The risks haven’t abated.

* To build institutions that create people and ideas that further the above missions, and generally mission-driven individuals that are willing to advocate and build a future they believe in. AI is a grand problem, and not one that I can do alone, so I need to build brands to rise through the noise and attract likeminded people.

At my best, I have many avenues for impact. I help open researchers work on impactful problems – not wasting the precious compute and time they have during the AI boom. I help policymakers know what is true. I build models that people use. I tell stories that make people smile. I keep the list wide so that I can stay motivated.

I see all of this continuing, and have been thinking about the broader impacts of this repeatedly over the last few months. Hearing that Andrej Karpathy was joining Anthropic prompted me to finally share more of my opinions:

For a long time, academic researchers being at the cutting edge of new technologies has been a great social equilibrium. Neutral, unbiased technologists have been the people to spread new ideas to the world.

As AI research takes off in velocity, it is also going behind closed doors. The tech industry has sowed distrust, and now they are the ones trying to tell the world about incredible changes coming. It’s a big loss to a form of social contract in America.

There’s been a history of scientists helping society understand new technologies. There is a public service in the culture of science that I want to see continue.

It’s being exacerbated by feelings of FOMO, especially financially driven, where I’m seeing many people who previously wanted to be professors -- and likely still do deep down -- feel a need to conform and chase money, in a pocket of industry. I get it, I grapple with this.

For those with a safety net, there will be great returns to some who choose to zag, and try to build something good, for people who need something different. For me, this is building interesting, fully-open models, to show what you can do with a variety of open weight sizes.

Yes, AI’s immediate future is dictated by the frontier, but it’s long-term trajectory still deeply includes academic institutions and open science. Knowledge will always diffuse, but to whom?

As of today, I think China is positioned to be the global home of AI research in a few years. The home of research is where ideas are accessible, spread rapidly, and are nurtured. The U.S. seems to be unwinding many institutions and relationships.

The largest returns go to people who build something differentiated, at least in reputation, and a lot of people are not being shown that this path exists.

To elaborate on this, I don’t fault any of the individuals who are going to industry today. I’ve been very close to doing this myself in the past weeks of job searching, or rather job exploring. It’s a systematic problem where scientists cannot easily get the support to take bold stances, especially stances that are designed around the public good.

To go a step further and say that only the research within closed, frontier labs matters is very myopic. Yes, there’s a sort of research you can only do with vast compute resources, and they will directly impact the most revolutionary tools of the day. But, I see the relative opportunity to do good elsewhere as higher for plenty of people.

Open research will always be the standard that sets the language people use to understand AI. It’ll always be how the next generation is trained – even if it’s behind what industry has built. It’ll be the ecosystem where new long-shot ideas are built. Without investing in this open ecosystem, all of these cycles will be kneecapped.

At the end of the day, so much of my role now is just showing the path to impact in this domain. To show how clever, mid-sized open models can impact real problems in the world. To show how policy-makers and educators need open research to structure the rest of society around AI. This is a fun role too! It would be very sad for me to see this light diminish ever further, into the lightest embers of a fire that looks almost entirely out.

Even if the pace of research were to slow further, if the folks remaining like myself got financial offers they can’t refuse for their families’ sake, the torch of open research will never fully go out. It’s core to how science is taught and done. There is a next generation coming, they just look for guidance and role-models.

What’s next

I see the best Ai2 work as research infrastructure. Building recipes in public gives countless researchers the ability to ask very specific questions of training processes. We need these researchers in the broader community, as Ai2 could never answer all the interesting questions themselves. One of my great joys in recent months has been visiting a top ML university and hearing so many graduate students say they’re building on Olmo. This is how the world should work!

Going forward, I still plan to operate in similar spaces, fighting for open-science, imagining what the future of the open model ecosystem can be, and doing my best to make the social transition to an AI-native era smooth. I’m most excited by how you can train medium sized open models on specific tasks that become useful tools in complement to the frontier models – massively winning on price. I want to invest in the ecological diversity of open models and coordination across builders.

For something that isn’t surprising given my past focus areas, I’m watching the pace of releases from all labs open & closed, and how they’re hillclimbing on super ripe new post-training veins (on-policy distillation, agentic workflows, etc.), it’s clear that fully-open post training recipes are about as far behind as they ever have been & falling further behind. I’d like to fix this. It’s not 100% clear yet if I will this year, but I’ll try.

To do this best and to execute, mostly personally, I needed a new start and fresh perspectives. I’ll be carefully building what I’m doing next over the next few months and am eager to share more about it when I can. One of my close teammates at Ai2 shared this quote with me in a farewell card, and I found it very apt in where I’m going next.

The object of life is not to be on the side of the majority, but to escape finding oneself in the ranks of the insane. — Marcus Aurelius

Thank you all for your continued support.

The largest debate that’ll define the future balance of power between the open and closed AI model ecosystems is primarily economic — it’s if users of AI will continue to pay dramatically more, i.e. large margins, for the top closed models. Early 2026 is a seminal time for the AI industry, as the coding agents have shown the first area where a huge AI market will continue to pay a substantial premium for better intelligence.

The other side of this dichotomy is the inevitable decay of API businesses at these same labs. These labs will realize they need to protect their best models, rolling them out later in APIs to both protect token supply, avoid distillation, and stick to use-cases with higher margins. All of these effects will be clearly visible in 5-10 year timelines, as in the near term markets, prices, margins, and demand will be dictated by a rapid buildout of compute (supply-limited in the near term) and mass subsidization of tokens (through continued investment in new AI companies).

The core of this argument rests in the obvious habit changes that are setting in with coding agents past the Opus 4.5 and Codex 5.2 thresholds. People are not making this switch because they are lazy, but because their net output is obviously higher when using an agent as an implementation aid for complex knowledge work. For people who rely on coding agents to work, they will always pay more for the best rather than settle for good enough. There are so many ways to make the product better, speed, intelligence, specialized models, etc.

I would pay $2000/month for the tools today, especially knowing they’ll get much better. At the same time, it is likely that many companies are forcing agents and usage onto people that actually will get very little out of them in their current form, which helps the AI buildout (or bubble) continue.

The best closed labs — right now this list is just Anthropic and OpenAI, but it’s reasonable to expect Google to catch up — will always make the most efficient models for intelligence at a given cost. Building models is a mass capital investment of talent, data, and compute. These systems, a combination of model weights, harnesses, tools, and serving infrastructure have massive returns on integration (where open models are designed to work across many, diverse serving situations). These integration benefits — the integration of hardware and new forms of software — can be expressed in any possible way of making models better.

The models in the near future may saturate on benchmark scores, but if that intelligence ceiling really is a cap on utility then the labs will optimize utility per second or per watt, serving users in another way. Improving the models is possible in every direction — there have been no walls in progress. We’re early in the mass buildout of intelligence, which involves harnessing the physical world to build numerous datacenters, organizing many AI researchers so that a large team can contribute to one model, and of course solving many small, low-level puzzles that unlock performance. Every indication is that there is still meaningful performance to be unlocked and the closed labs are the best set up to extract it.

The collective wisdom of the labs is that making the models smarter, in terms of the frontier of absolute intelligence, has the most value. This is the right call to me because it unlocks large new markets. Optimizing models at a fixed intelligence level locks in markets, expands accessibility over time, and increases return on investment for users (while potentially lowering margins for selling intelligence).

Many people are making this bet that models will keep getting better and are learning to work well in these harnesses, even though some workflows are still a bit clunky. This is the right bet. These people all will continue to use the absolutely best models available. It’s like buying an iPhone as a consumer. You could get an Android and suffer from a bunch of paper cuts to save money, but why would you? The returns to performance are even higher in the workplace, which drives pricing power.

In this mental model, the frontier labs as businesses, will look like new, reimagined forms of a mix of Apple and Microsoft. The Apple side is that they’re selling an integrated, extremely hard to replicate technology. The Microsoft side is selling high-leverage subscriptions across the economy. In 5-10 years I expect both OpenAI and Anthropic to be valued in the $2-10T range. The true frontier labs will be an oligopoly that looks like the cloud market today.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

On the other side of this equation is the open model economy. This isn’t to say that the frontier labs will dominate all aspects of AI use. Yes, I expect OpenAI and Anthropic to be the most representative companies of the AI boom (new companies, alongside Nvidia of course), but the collective value capture around open models will be far bigger overall, it’s just that the revenue and margins will be shared across a wide stack of companies.

Many businesses want to switch to open models but the models today are not good enough in out-of-distribution tasks. Eventually open model builders will stop chasing Claude and GPT on the Artificial Analysis index and fill this niche. This fork could be driven by economic factors, where they no longer have the revenue to support the growing R&D costs for continuing to scale models. It can also be driven by pure demand, where certain AI solutions only can exist at low price points present in open models. Where closed labs are an oligopoly, open model builders and users will be far more diverse and numerous. The total market value will dramatically exceed the cumulative value of OpenAI and Anthropic.

Open models are by their nature not integrated, so they will rely on multiple companies coordinating to serve them. Each of these layers will have alternatives, driving prices down to commodity pricing. These low, predictable prices will be where many enterprises enter to build in-house agents and tools for niche tasks. The predominant mode of deployment here is that enterprises find a model that hits a sufficient performance threshold on a task of interest and does not replace the model later (setup costs are high). As customizing models becomes easier, again in the open model finetuning stack we are seeing emerge (Tinker, Fireworks, Prime Intellect, etc.), this market becomes even bigger.

What this will look like in the coming years is a steady rise in open model inference proportion across the entrenched hyper-scale clouds of Google, Amazon, Microsoft and new AI infrastructure companies of Together, Fireworks, OpenRouter, etc when compared to OpenAI and Anthropic.

The key is that the open and closed model economies are operating on different exponentials. I still believe that progress will continue at a fast pace across the entire ecosystem, but claims of recursive self improvement (RSI) giving the closed labs an unassailable advantage are overblown. New forms of products like background agents can support both these open and closed models.

The closed models hit incredible product-market fit with the current agents, starting their integrated exponential by monetizing the top end of the knowledge work. The open model economy will take far longer, but it will also be far more satisfying to follow, as it tracks the broader diffusion of AI into the entire economy and world.

As the years of AI progress go by, it’s been accompanied by a slowly rising tide of consequence. Models are getting more capable, how we work is changing quickly, economics of AI are becoming real, just as real-world risks come to the forefront. 2026 is the first year where I don’t think there’ll be any breaks from this. The hard part to prepare for is that there’s a good chance things just continue to ratchet up from here – more disruption, more surprises, more stakes.

On my end, there’s been a growing list of topics that are very fateful to how I see the current state of AI, but I haven’t even gotten to write about them (at least not from all the angles I want to)! All of these are closely related to the implications of different models reaching new capability levels and how I use that to infer what may come next.

1. Open models haven’t had their true agent moment like Opus 4.5

The time gap between open and closed models is very often discussed, but the reality is that we have a nice time-gating that’s independent of debatable benchmarks – if open-weight models do or do not become super useful in agentic harnesses. The Opus 4.5 in Claude Code moment of December 2025 was so loud and obvious, that if open models hit this performance level for price points as low as $5/month, there will be an explosion in usage.

Right now we are about 5-6 months in with no equivalent open model. I suspect the robustness of the best closed frontier models that I write about could make this moment take a good amount longer, say closer to 12+ months. In this time, Claude Code and Codex may seem like different categories of products. In the standard flurry of new, state-of-the-art open models from a variety of labs, benchmarks will definitely keep climbing, but the open-closed gap should become more interpretable as real-world use becomes the real litmus test.

2. Gemini still doesn’t have a meaningful competitor for Claude Code and Codex

The best exclamation point I can offer to reinforce my prediction that open models are further behind than the benchmarks claim is that even the mighty Google doesn’t have a clear competitor for Claude Code and Codex. I’m sure the Gemini team is pushing very hard on this.

I still need to do a lot more testing on Gemini 3.5 Flash, but reading reviews makes it clear that it’s not a substitute for how I’m working today. It’s maybe not the Gemini team explicitly specializing for Google’s existing products (search, YouTube, etc.), but the model seems to suit them. If Google doesn’t have a powerful tool here soon, I don’t expect the open model labs to either. The open models are going to be used more for automated, enterprise agents and low-cost domains, rather than being the driving tool of modern knowledge work. This will feed directly into the economic engine of funding future models, where the agents like Claude Code and Codex are the current best path to massive AI revenue growth.

I discussed how the current environment is quietly driving labs in China to specialize on AI Proem with Grace Shao and this is central to my expectations of open models specializing over the next few years instead of competing with OpenAI, Anthropic, and Google.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

3. I don’t expect an open-weights Mythos this year

While I don’t think Mythos is a general “god model” that will crush the competition in every domain, I do think it’s a remarkable technical achievement in software engineering and cybersecurity. Mythos is obviously a watershed moment for those fields. Having spoken to most of the Chinese labs – particularly those with the most prominent, large, open MoE models like Kimi, Z.ai, DeepSeek, and Qwen – I think they’re heavily resource limited and don’t have an immediate path to scaling up training processes like the big labs in the U.S. For the labs which are more corporate, which comes with more resources, such as Alibaba and Bytedance, they also have more conservative stances on safety and security.Mythos is a bellwether of the massive acceleration in training and research compute available to the largest American companies.

Epoch AI recently had a nice piece on the compute available to various labs (~Google 25%, Meta 11%, OpenAI 11%, Anthropic 6%). All of these numbers are vastly higher than any Chinese lab.

4. American open models are slowly gaining steam

Nvidia with Nemotron, Google with Gemma, Arcee AI and others are slowly stabilizing the open model ecosystem in the U.S. There’s a lot that’s hard to measure here, especially in the rise of local agents like OpenClaw and Hermes, but there are adoption numbers of American models that we haven’t seen since Llama 3.Gemma 4’s models are all tying or outperforming the equivalently sized Qwen 3.5/3.6 models — where Qwen has for years now been the default open model at these sizes. These Qwen 3.5/3.6 models have been tricky to get working in a lot of post-training research, partially due to architecture/tooling and partially likely due to modeling (i.e. the model is not easy to finetune for some training decision). I’ve heard few complaints about Gemma, but it also could be because Gemma is not yet the researcher default.

There's a simple reality that we've seen recently with models like GPT-OSS, Nemotron 3, and now Gemma 4, that if a model is in the right range of benchmarks and released by an American lab with a truly permissive license, it'll get a large amount of adoption (in this cycle, recall that Gemma 4 adopted the Apache 2.0 License, changing from one with use-case restrictions on earlier Gemmas). This early phase of American growth in open models is establishing key brands directly with developers. The consensus is that more neolabs like Reflection and Thinking Machines are likely to participate in this space, but being too patient will lose the time when new agentic workflows and enterprise relationships are built.

5. Anthropic and OpenAI are just getting up to speed in model iterations

I expect the rest of this year to be a ruthless competition between these two flagship companies. I’m at an interesting balance where I think GPT 5.5 is a bit smarter of a model and I love the Codex App, so I’m structuring much of my work to be possible there. At the same time, for a lot of writing-related and broader surface area tasks I really still love Claude. These models are rapidly changing how we work, I run Codex from my phone while doing other things, am setting up automated open model analysis jobs on the back of agents, and expect to be able to scale the research side of Interconnects widely.

AI is beginning to drive companies to the two extremes in the scaling era. The biggest companies will be way bigger than ever, using resources and mass talent to have sustained progress at the frontier of raw AI capabilities. On the other side, tiny businesses like Interconnects thrive by using agents to refine, present, and sell niche expertise. The mass social job displacement that’ll come is going to reduce employability for various knowledge workers that don’t fit into either of these extremes for the raw technical side (big or small companies), while sustaining and maybe even amplifying careers that interface directly with humans (e.g. doctors) or other power structures with means to sustain themselves (law/government).

6. More existing power structures will assert themselves on AI

Just in the last few days while writing this, we had the Pope release an over 40,000 word document on where AI is going and China expand personnel movement restrictions on top AI researchers across industry. At the same time, the U.S. has designated Anthropic a supply chain risk and continues to use its models for national security. The list of news like this is only going to grow. Existing power structures are realizing there’s a finite time window for them to exert themselves in the AI dynamic — an intuition that could be mapped to influence going down as AI models get more powerful. This intuition is potentially dangerous, as it sets up meaningful conflict in who controls the technology (as I discussed with Dean Ball after the Anthropic-DoW spat).

Next: Where technical becomes social

These largely technical and power trends accelerating are going to put more pressure on the social and political anti-AI sentiments within the U.S. This is currently the most obvious barrier to continued AI development and beneficial diffusion. Reflecting on this, many people in the tech discourse get too focused on the details, where yes a lot of data-center-detractors are making genuinely wrong factual claims in defense of their position.

The real position that a large swath of Americans has is that they have a voice in saying no to the current trend — by not granting permission to build data centers. This is a voice that they haven’t been granted by the tech industry that changed the face of the global economy and power structures in the last few decades.

This is setting us up for a challenging year ahead for the industry. The labs are aggregating and concentrating talent to peak levels. There are few neutral messengers to communicate the reality of AI to the public. The frontier labs leadership is largely gearing up to IPO and stay ahead in the capabilities race. With the status quo, there are few actions to unwind this path toward social conflict.

It takes individuals in the AI ecosystem to zag and go against the groupthink of needing to make your wealth today, of needing to be at a lab to do impactful work, and so on. I’m personally continuing to bet on this, by trying to make a vibrant and diverse open model ecosystem supported by clear, unbiased information. If you agree with this and have been watching from the sidelines, it’s a good time to get involved, before the situation spirals into something uncontrollable.

Staring out the window on a new, high-speed train from Hangzhou to Shanghai I’m gifted with views of dramatic ridgelines speckled with wind turbines that are silhouetted against the setting sun. The mountains cast a backdrop to a mix of spanning fields and clustered skyscrapers. I’m returning from China with great humility. It’s a very warming, human experience to go somewhere so foreign and be so welcomed. I had the honor of meeting so many people in the AI ecosystem who I knew from afar, and they greeted me with big smiles and cheer, reminding me how global my work and the AI ecosystem is.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

The mentality of Chinese researchers

The Chinese companies building language models are set up as the perfect fast-followers for the technology, building on long-standing cultural traditions in education and work, along with subtly different approaches to building technology companies. When you look at the outputs, the latest, biggest models enabling agentic workflows, and the ingredients, excellent scientists, large-scale data, and accelerated computing, the Chinese and American labs look largely similar. The lasting differences emerge in how these are organized and conditioned.

I’ve long thought that a reason that the Chinese labs are so good at catching up and keeping up with the frontier is that they’re culturally aligned for this task, but without talking to people directly I felt like it wasn’t my place to attribute substantial influence to this hunch. Speaking with many wonderful, humble, and open scientists at the leading Chinese labs has crystallized a lot of my beliefs.

So much of building the best LLMs today comes down to meticulous work across the entire stack, from data to architecture details and RL algorithm implementations. All points of the model can give some improvements, and fitting them in together is a complex process where the work of some brilliant individuals needs to get shelved in favor of the overall model maximizing a multi-objective optimization.

Where American researchers are obviously also brilliant at solving the individual components, there’s more of a culture of speaking up for yourself in the U.S. As a scientist, you’re more successful when you speak up for your work and modern culture is pushing the new path to fame of “leading AI scientists”. This results in direct conflict. The Llama organization is heavily rumored to have collapsed under the political weight of these interests embedding themselves in a hierarchical organization. I’ve heard of other labs saying that it can be needed to pay off a top researcher to get them to stop complaining about their idea not making it in the final model. Whether or not that’s exactly true, the idea is clear. Ego and desires for career advancement do get in the way of making the best models. A small, directional shift in this sort of culture between the U.S. and China can have a meaningful impact on the final outputs.

Some of this has to do with who is building the models in China. There’s an immediate reality at all of the labs that a large proportion of the core contributors are active students. The labs are quite young, and it reminds me of our setup at Ai2, where students are seen as peers and directly integrated in the LLM team. This is incredibly different from the top labs in the US, where the likes of OpenAI, Anthropic, Cursor, etc. simply don’t offer internships. Other companies like Google nominally have internships related to Gemini, but there’s a lot of concern about whether your internship will be siloed and away from anything real.

To summarize how the slight change in culture can improve the ability to build models:

* More willingness to do non-flashy work in order to improve the final model,

* People new to building AI can be free of prior phases of AI hype cycles, allowing them to adapt to the new modern techniques faster (in fact, one of the Chinese scientists I talked to really actively attached to this strength),

* Less ego enabling org charts to scale slightly, as there’s less gamifying the system, and

* Abundant talent well-suited to solving problems with a proof of concept elsewhere, etc.

This slight inclination towards skills that complement building today’s language models stands in contrast to a known stereotype that Chinese researchers tend to produce less creative, field-spawning, 0-to-1 academic style research. Among the more academic lab visits on our trip, many leaders talk about cultivating this more ambitious research culture. At the same time, some technical leaders we talked to were skeptical about whether such a rewiring in the approach to science is likely in the near term, because it’ll take a redesign of the education and incentive systems that is too big to happen within the current economic equilibrium. This culture seems to be training students and engineers that are excellent at the LLM building game. They also, of course, have an extremely abundant quantity.

These students told me about a similar brain drain happening in China as in the U.S., where many who previously considered academic paths now intend to stay in industry. The funniest quote was from a researcher who was interested in being a professor to be close to the education system, but remarked that education is solved with LLMs – “why would a student talk to me!”

The students have a benefit of coming at LLMs with fresh eyes. Over the last few years we’ve seen the key paradigm of LLMs shift from scaling MoE’s, to scaling RL, to enabling agents. Doing any of these well involves absorbing an insane amount of context quickly, both from the broader literature and the technical stack at your company. Students are used to doing this and excited to humbly drop all presumptions about what should work. They dive in head first and dedicate their life to getting the chance to improve the models.

These students are also so magically direct and free of some of the philosophical chatter that can distract scientists. When asking questions on how they feel about the economics or long-term social risks of models, far fewer Chinese researchers have sophisticated opinions and a drive to influence this. Their role is to build the best model.

This difference is subtle, and easy to deny, but it is best felt when having long conversations with an elegant, brilliant researcher who can clearly communicate well in English, basic questions on more philosophical aspects of AI hang in the air with a simple confusion. It’s a category error to them. One researcher even quoted the famous Dan Wang premise of China being run by engineers, relative to the lawyers of the U.S. when probing in these areas, to emphasize their desire to build. There’s no track in China that systematically enables the growth of star power for Chinese scientists, akin to mega mainstream podcasts like Dwarkesh or Lex.

Trying to get Chinese scientists to comment on the coming economic uncertainty fueled by AI, questions beyond the capabilities of simple AGI, or moral debates on how models should behave all served to capture the upbringing and education of these scientists (edited). They are extremely dedicated to their work, but have grown up in a system where debates and opinions on how society should be structured and changed are not encouraged.

Zooming out — Beijing especially felt much like the Bay Area, where a competitive lab is a short walk or Uber away. I got off a flight and stopped by Alibaba’s Beijing campus on the way to the hotel. Then, in 36 hours we went to all of Z.ai, Moonshot AI, Tsinghua University, Meituan, Xiaomi, and 01.ai. Travel by Didi is easy, and if you select an XL in China you’re often paired with electric mini vans that have massage chairs. We asked the researchers about the talent wars, and they said it’s very similar to what we’re experiencing in the U.S. It’s normal for researchers to bounce around, and much of where people choose to go is based on the best current vibes.

In China, the LLM community feels far more like an ecosystem than battling tribes. Across many off the record conversations, it’s nothing but respect for peers. All of the Chinese labs fear Bytedance with their popular Doubao model, which is the only frontier closed lab in China. At the same time, all of the labs have massive respect for DeepSeek as the lab with the best research taste in execution. When you meet with lab members off the record in the States, sparks fly quickly.

The most striking part of the humility of Chinese researchers is how they also often shrug on the business side, saying it’s not their problem, where everyone in the U.S. seems to be obsessed with various ecosystem-level industrial trends, from data sellers to compute or fundraising.

Where China’s AI industry differs (and matches) the Western labs

The thing that makes building an AI model today so interesting is that it’s not just about getting a group of great researchers in one building together to produce an engineering marvel. It used to be this, but to sustain AI businesses, the LLMs are becoming a mix of building, deploying, funding, and getting adoption for this creation. The leading AI companies exist in complex ecosystems that supply money, compute, data and more in order to keep pushing the frontier.

The integration of these various inputs to creating and sustaining LLMs is fairly well conceptualized and mapped for the Western ecosystem, as typified by Anthropic and OpenAI, so finding big differences in how the Chinese labs think about it points at where the different companies can be making meaningfully different bets on the future. Of course, these futures can be heavily dictated by the constraints on funding and/or compute.

I’ve documented the biggest “AI Industry” level take-aways from talking to these labs:

* Early signs of domestic AI demand. There’s a much-touted hypothesis that the Chinese AI market will be smaller because Chinese companies don’t tend to pay for software – thus, never unlocking a giant inference market supporting labs. This is only true for software spend that maps to the SaaS ecosystem, which is historically tiny in China, where on the other hand there is obviously still a large cloud market in China. A crucial unanswered question – one which the Chinese labs themselves debate – on if spending for AI in the enterprise tracks the SaaS market (small) or the cloud market (fundamental). On net, it feels like AI is trending closer to the cloud, and no one was actively worried about a market growing around the new tools.

* Most developers are Claude-pilled. Most of the AI developers in China are obsessed with Claude and how it’s changed how they build software, despite Claude nominally being banned in China. Just because China has historically been hesitant to buy software does not give me the impression that there won’t be a massive surge in inference demand. Chinese technical staff are so practical, humble, and motivated – a fact that seems stronger than any commitment to previous habits in not spending.Some Chinese researchers mention building with their own tools, such as the Kimi or GLM CLIs, but all of them mention building with Claude. There were also surprisingly few mentions of Codex, which is definitely surging in popularity in the Bay Area.

* Chinese companies have a technology ownership mentality. The Chinese culture is combining with a roaring economic engine to create unpredictable outcomes. I’m left with a lasting feeling that the numerous AI models reflect a practical, current equilibrium of the many technology businesses here. There’s no master plan. The industry is defined by a respect for ByteDance and Alibaba, the incumbents expected to win large portions of all markets with their substantial resources. DeepSeek is the respected technical leader, but far from a market leader. They set the direction, but aren’t set up to win economically.This leaves companies like Meituan or Ant Group, where people in the West can be surprised they’re building these models. In reality, they see LLMs obviously as being central to future technology products, so they need a strong base. When they fine-tune the strong, general purpose model it hardens their stack from getting the open community to provide feedback on it, and they can keep internal, fine-tuned versions of the model for their products. The “open-first” mentality in the industry is largely defined by practicality — it helps make their models get strong feedback, it gives back to the open-source community, and empowers their mission.

* Government aid is real, but unclear how big. It’s often asserted that the Chinese government is actively helping with the open LLM race. This is a government that’s decentralized across many levels, each of which doesn’t have a clear playbook for what exactly they do. Neighborhoods in Beijing compete for tech companies to house their offices there. The “help” offered to these companies almost certainly involved removing bureaucratic red tape like permits, but how far does it go? Can levels of the government help attract talent? Can they help smuggle chips? Across the visit, there were many mentions of government interest or help, but far too little to report the details as assertive or have a confident worldview of how government can bend the trajectory of AI in China. There were certainly no hints of the top levels of the Chinese government influencing any technical decisions in the models.

* The data industry is far less developed. Having heard so much about the likes of Anthropic or OpenAI spending $10M+ for single environments, with cumulative spend on the order of hundreds of millions per year to push the frontier of RL, we were eager to know if Chinese labs are either buying the same environments from companies in the U.S. or supported by a mirrored domestic ecosystem. The answer was not quite complete that there’s no data industry, but rather that their experience was that the data industry was relatively poor quality and it is often better to build the environments or data in-house. Researchers themselves spend meaningful time making the RL training environments, and some of the bigger companies like ByteDance and Alibaba can have in-house data labelling teams to support this. This all mirrors the build-not-buy mentality from the previous bullet.

* Desperation for more Nvidia chips. Nvidia compute is the gold-standard for training and everyone is limited in progress by not having more of it. If supply was there, it is obvious that they would buy it. Other accelerators, including but not limited to Huawei, were spoken positively of for inference. Countless labs have access to Huawei chips.

These points paint a very different picture of an AI ecosystem, where quickly mapping how Western labs operate to their Chinese counterparts will often result in a category error. The crucial question is if these different ecosystems will produce meaningfully different types of models, or if the Chinese models will always be explained by being similar to the U.S. frontier models of 3-9 months ago.

Conclusion: The global equilibrium

I knew so little about China going into the trip and came out with the feeling of just starting to learn. China isn’t a place that can be expressed by rules or recipes, but one with very different dynamics and chemistry. The culture is so old, so deep, and still completely intertwined with how domestic technology is built. I have much more learning ahead.

So much of the current power structures in the US use their current worldviews of China as crucial mental devices for decision making. Having talked, in person, either formally or informally to pretty much every leading AI lab in China, there are a lot of qualities and instincts in China that’ll be very hard to model with Western decision making. Even after asking directly about why these labs release their top models openly, the intersection between ownership mentality and genuine ecosystem support is hard for me to connect the dots on.

The labs here are practical and not necessarily absolutists around open-source, where every model they build would be released openly, but there’s a deep intentionality in supporting developers, the ecosystem, and using it as a way to learn more about their models.

Almost every major Chinese technology company is building their own general purpose LLMs, as we see with the likes of Meituan (delivery service) and Xiaomi (broad consumer technology company) releasing open weight models. The equivalent companies in the U.S. would just buy services. These companies aren’t building LLMs out of a race to be relevant with the hot new thing, but a deep fundamental yearning to control their own stack and develop the most important technologies of the day. When I look up from my laptop and always see bunches of cranes on the horizon, it obviously fits in the with the broader culture and energy around building in China.

The humanity, charm, and genuine warmth of Chinese researchers is extremely humanizing. At a personal level, the cut-throat geopolitical conversation we’re used to in the U.S. hasn’t permeated them at all. The world can use more of this simple positivity. As a citizen of the AI community, I currently worry more about the fissures appearing within members and groups around labels of nationality.

I’d be lying if I said I didn’t want US labs to be clear leaders in every part of the AI stack — especially with open models where I spend my time — I’m American, and that’s an honest preference. With this, I want the open ecosystem itself to thrive globally, as this can create safer, more accessible, and more useful AI for the world, and right now the question is whether American labs will take the steps to own that leadership position.

As of finishing this piece, more rumors are swirling of executive orders influencing open models, which can further complicate this synergy between American leadership and the global ecosystem — it doesn’t fill me with confidence.

Thank you to all the wonderful people I got to talk to at Moonshot, Zhipu, Meituan, Xiaomi, Qwen, Ant Ling, 01.ai, and others. Everyone has been so welcoming and gracious with their time. I’ll keep sharing my thoughts on China as they crystallize, across culture generally and AI specifically. It is obvious that this knowledge will be directly relevant to the story unfolding at the frontier of AI development.

‘Distillation attacks’ is a horrible term for what is happening right now. Yes, some Chinese labs are hacking or jailbreaking APIs to attempt to extract more signal from model APIs — stopping this is important to maintain the U.S.’s lead in AI capabilities. Referring to this as distillation attack is going to irrevocably associate all distillation with this behavior, and distillation generally is a core technique needed to diffuse AI capabilities broadly through academic and economic activities.

We went through this sort of language transition with the open source vs open weight debate. All the terms just reduced to open models – very few people in the large AI community know exactly how open-source differs from open-weights. And terminology matters, as the less informed people who still care about — and influence — the technology are bound by different terms they use. If we’re not careful with the discourse around distillation, many people could associate this broad technique used for research and development of new models as an act at the boundary of corporate manipulation and crime.

I’ve recently written a more technical piece on estimating how impactful state-of-the-art distillation methods are on leading Chinese models, and this piece follows to push for caution in any hasty actions to target the methods with policy. To set the stage, recall Anthropic’s recent blog post where they detailed “distillation attacks” made by 3 Chinese labs.

These labs used a technique called “distillation,” which involves training a less capable model on the outputs of a stronger one. Distillation is a widely used and legitimate training method. For example, frontier AI labs routinely distill their own models to create smaller, cheaper versions for their customers. But distillation can also be used for illicit purposes: competitors can use it to acquire powerful capabilities from other labs in a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.

This is a clever paragraph, where they normalize distillation generally and explain how a few people can use it illicitly, without detailing how illicit use often involves other more explicit behavior like jailbreaking, hacking, or identity spoofing of the API.

Distillation itself is an industry standard. It’s used extensively, primarily in post-training, by smaller players to create specialized or smaller models. In my book coming this summer, I describe it as follows:

The term distillation has been the most powerful form of discussion around the role of synthetic data in language models. Distillation as a term comes from a technical definition of teacher-student knowledge distillation from the deep learning literature.

Distillation colloquially refers to using the outputs from a stronger model to train a smaller model.

In post-training, this general notion of distillation takes two common forms:

* As a data engine to use across wide swaths of the post-training process: Completions for instructions, preference data (or Constitutional AI), or verification for RL.

* To transfer specific skills from a stronger model to a weaker model, which is often done for specific skills such as mathematical reasoning or coding.

With this definition, it’s easy to see how distillation takes many forms. Of course, if you just take the outputs from GPT-5.5 and train a recent open-weight base model with them to host a competitive product, that’s one thing. But, a lot of the things that fall under the bucket of distillation are complex, multi-stage processes that muddle the exact impact of the model you distilled from.

Modern LLM processes could look like using a GPT API to build an initial batch of synthetic data to build a specialized small data-processing model. A good example is a model like olmOCR (or many other models in this category) that are trained to convert PDFs to clean text. This specialized model would be used to create large amounts of data. Finally, you train another model (often from scratch) with the new data you created. Is this final model distilled from GPT?

When done via a closed, API-based model, distillation sits in the grey area of the terms of service that you agree to when signing up to the Claude or GPT platform. They generally forbid the use of the API to create competing language model products, but this term has largely gone unenforced. The open-source community used to worry deeply at being cut off from these cutting-edge APIs for doing research or creating public datasets, but to date only one prominent case of corporate accounts being restricted exists (at least until the recent Chinese companies).

This is all to say that distillation is an industry standard technique, and the use of closed APIs to perform distillation has always been a grey area. Nvidia’s latest Nemotron models, as one of the only models with open post-training datasets, are technically in large part distilled from Chinese, open-weight models. The Olmo models we’ve built at Ai2 are distilled from a mix of open and closed models. This grey area was brought to the forefront again when it turned out that xAI has been distilling from OpenAI. Quoting from the recent trial proceedings between Elon and OpenAI:

OpenAI’s counsel asked Musk whether xAI has ever “distilled” technology from OpenAI.

Musk: “Generally AI companies distill other AI companies.”

“Is that a yes?” Savitt asked.

Musk: “Partly.”

xAI is likely the largest, and most successful AI company willing to thread the grey area that is distillation from their competitors. On the other side, the majority of startups and research groups with fewer resources than them have very likely engaged in distillation of some capacity from Claude, GPT, or Gemini models.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

In the above Anthropic blog post, the problem with the distillation attacks by a few Chinese labs is less the distillation and more the means of attack. It is documented that Chinese labs are actively working to get around the intended use of the API, e.g. to provide additional reasoning data that is very useful for training.

Of course no one should be able to access information from a model that a developer didn’t intend to reveal in their APIs (e.g., reasoning traces which would be helpful for training). Associating all of distillation with these attacks, which is to date an industry standard for post-training, from open and closed models alike will be a massive own goal.

What these few labs are doing should be referred to as jailbreaking or abuse, rather than distillation.

The discourse around these actions is creating a troubling discussion that’s marching towards a mix of regulatory capture or regulatory exuberance that’s most likely to harm the U.S.’s ecosystem more than China’s. Even if we ban, most likely through potential legal action and other penalties, this type of API abuse, the Chinese companies will likely still do it. We’ve seen this playbook with Chinese multimedia models taking a flexible view of copyrighted content that no U.S. player is willing to take the risk on.

This distillation discussion has quickly snowballed, with a bill moving out of a committee in Congress, an executive order pushing for action, and congressional oversight targeting U.S. companies building on Chinese models (which are downstream of distillation). This multi-pronged regulatory environment could yield truly horrible outcomes – such as figuring out a way to effectively ban open-weight models in the U.S. that are built in China by groups abusing closed LLM APIs.

It is obvious that no bill will literally ban open models, but they can create grey area that exposes entities to unwanted risk or require certain provisions that are bureaucratically very challenging to fulfill, squashing small open source contributors.

In that scenario, the groups who lose are Western academics and smaller companies building models for the long-tail of AI uses. The ecosystem here could be made permanently irrelevant with the removal of nearly all Chinese open-weight models. There is no immediate substitute and building new models with meaningful community adoption has a lead time measured in 6+ months. In the time it takes to build a new domestic open-source ecosystem, countless researchers would’ve moved onto closed training platforms or into new areas.

Altogether, I’m hoping this flurry of discussion around distillation becomes a nothing-burger and not a hasty, multi-pronged policy push. We need to avoid two things:

* A wholesale negative connotation of the word distillation, which is used extensively across the AI ecosystem.

* A domestic ban of the open-weight models built by organizations engaged in some portion of distillation.

In addition to this, I want the leading U.S. AI companies to be able to provide their APIs without having their IP leak. They should share more information on why it is hard for them to secure their APIs, but that’s an issue out of scope for my expertise.

I’ll conclude with a proposal from my friend Kevin Xu at Interconnected Capital (and great Substack) on why this current distillation dynamic may actually be good for the leading labs.

If all the Chinese companies are addicted to distillation as a way of getting close to the frontier, then they’ll never actually learn the techniques needed to take an outright lead. If we cut off the Chinese’s obvious crutch in model building, we’ll gain a short-term lead in AI, but in the long-term that may be what they needed to get on a more competitive long-term trajectory.

This is the same debate we’re having with other technologies where the U.S. currently has a lead, e.g. with advanced semiconductor technologies. So I understand the trade-offs, but we not should crack down on all of distillation.

We’re living through the period of time when we’ll learn if open models can keep up with closed labs. The obvious answer is that no, they won’t. This answer is a form of saying they won’t keep up in every area. This framing closes off a popular prediction where the open models completely catch up, as in all models saturate and open and closed models only become increasingly similar. In living through this, it’s evidently very unclear when the longer-term stable balance of capabilities will solidify.

This is a very complex dynamic, where the core point we monitor is a capability gap between models. At the same time, this gap is intertwined with evolving dynamics in the funding of open models, who builds open models, how techniques like distillation that enable fast-following translate through new application domains, potential regulation hampering the open-source AI ecosystem, and of course who actually uses open models.

The capabilities gap is one signal in a complex sea of forces, pushing supply and demand into different shapes. In many cases the demand — where obviously tons of individuals, organizations, and sovereigns want, or need, open models — is largely separated from supply. Supply is fully dictated by economics. The question of “which business strategies support releasing open models” is still at stake.

Interconnects AI is a reader-supported publication. To receive new posts and support my work, consider becoming a subscriber.

With this complexity, I wanted to distill my key beliefs down into a clear list. These are downstream of 10+ pieces I’ve written or recorded on open models this spring (which are linked throughout).

* It’s surprising that the top closed models did not show a growing capability margin over open models, based on compute differences for training and research, especially in the second half of 2025 and through today.

* Open model labs are technically very strong at keeping pace on well-established benchmarks. This will continue and reflects a balance of abundant talent and sufficient computing power.

* Chinese open-weight labs focus slightly more on benchmark scores than comparable closed labs in the U.S. Distillation helps the Chinese LLM companies do so, but it’s not a panacea. Changes in the distillation dynamic (e.g. regulation) will not be a determining factor on the balance of capabilities. This increase in focus is a natural evolution of their incentives in keeping the narrative on keeping up with the frontier alive, which is crucial to fundraising and adoption.

* To date, closed models tend to be more robust and generally useful than similarly scoring open models. Closed models have certain hard-to-measure qualities that are not well captured in current or past benchmarks. This will be key to enabling closed models to dominate in markets where an individual user constantly presents new challenges, i.e. supporting knowledge workers as a direct assistant.

* The open vs. closed model race, as monitored through benchmarks, will largely be a game of economic staying power and fast-following, until the market structure constricts. I expect Chinese open-weight labs to face funding difficulties first, as soon as later this year. Funding difficulties will be seen in different capability trajectories 3-9 months later.

* The RL dominated training era has increased the relevance of distribution to real-world use-cases as a key factor in continued capabilities improvements. These are tasks where users directly use tools like Claude Code or Codex to solve problems in their job with agents. This is the first clear technical area that closed labs can dominate open-weight models on capabilities, potentially leveraging online RL directly based on user feedback.

* Open models will be increasingly adopted in repetitive automation tasks, as measured in the relative share of the API market, for repetitive tasks across the ecosystem. This takes the form of many new AI-native applications, business backend automation, etc. The success of this will drive more investment in domain-specific, efficient open models.

This is a complex picture, where the long-term trajectory is more of an economics question rather than an ability one. Many other outlets can paint a far more simplistic narrative that “China will assuredly catch us in AI” and get more distribution because it is a simple story. The reality is complex. Only real AI revenue begets more investment, eventually that’ll be linked to the ability to keep improving models at a rapid rate. Economic realities have not yet impacted scaling open models, as a general category.

This economic-focused angle relates to my positions on the open model ecosystem more broadly.

* Recurring calls to ban certain types of open models will continue to come but are in practice impossible to implement. Training strong AI models (i.e. near but not at the frontier) is a relatively small cost compared to large-scale deployments. E.g. if the U.S. bans open models over a certain compute threshold, another sovereign entity will eventually train them and release them publicly, with the models entering the U.S. market with less oversight.

* The second derivative of influence on open models has shifted, and the U.S. will slowly regain ground in adoption metrics of open models starting in early 2027 (it takes a long time for China’s velocity to slow, then flip). Examples include Google’s Gemma 4 (a wild success), Nvidia’s Nemotron, and Arcee AI.

* As ever-stronger closed models are built, previewed, and released, there will be more safety-shocks saying that open-weight versions of the strongest AI models never can be allowed to exist, similar to reactions to Claude Mythos. These can spur burdensome regulation on open models.

* With the above, there will also be increased long-term interest in open models, as sovereign entities and existing power structures realize the coming, super powerful AI tools cannot land in the hands of only one or a few companies. These entities will see open models as a different governance paradigm.

* New funding structures for open models will emerge, as many stakeholders realize dependencies on single, for-profit companies for access to intelligence are unreliable.

* Local agents, OpenClaw, and other personal agents represent a large, to date, mostly ignored market for open model usage. It is a sort of dark matter, with pervasive, massive potential for influence on the balance of open-to-closed models.

A single word governs this post and is intentionally repeated — complex.

This complex reality has been driving me to think more deeply about how to clearly describe the open model gap, and why I can hold it in my head that I expect American closed labs to clearly draw ahead, despite the fairly unequivocal evidence in support of the capabilities of recent open-weight models. More on the nuance in the open-closed gap in another piece coming soon, so please subscribe!

Let me know any positions that I missed.

Recently, I was talking with Percy Liang, Stanford professor and lead of the Marin project (another fully-open model lab), and it set in on me that there will eventually be a consortium of companies funding a foundational set of open models used across industry. It’s not clear when this’ll emerge, and Nemotron (Coalition) is Nvidia’s attempt to bankroll and bootstrap this approach within a single wealthy company, but a consortium is the only long-term stable path to well-funded, near-frontier open models.

In recent months, we’ve seen a lot of turnover in open model labs, with high-profile departures at Qwen and Ai2 (my comment). This shouldn’t be super surprising to followers of the ecosystem — it’s happened before with Meta shifting its focus away from Llama, and it’ll only happen more as the cost of trying to keep pace at the frontier of AI only increases. The other leading labs with models available today include Chinese startups such as Moonshot AI, MiniMax, and Z.ai — all of which look precarious on their ability to fund continued growth in the cost of training or R&D. Releasing one’s strongest models openly today is in active tension with the option of spending focus and resources on AI products that can currently generate meaningful revenue (and profits).

We’re going to see business models emerge around releasing some, or even many, models openly, but these will largely be smaller models that enable a long-tail of functionality, rather than models at the absolute frontier. This class of companies that’ll release many, strong fine-tunable models will include the likes of Arcee AI, Thinking Machines, OpenAI, Google with Gemma, and more in that class. The cost and relative advantage of keeping the best models closed in a business environment with many opportunities for revenue are too high. To summarize — there will be an ever increasing number of companies releasing models that are good for creating a lively niche of smaller, custom models, but an ever decreasing number of companies willing to release fully open, near-frontier models.

This is the core thesis of why I’m pushing hard for more people to do more research on how these smaller models can complement the best closed agents, the science of finetunability, etc. See my post below — it’s about creating a sustainable open model ecosystem, whether or not the frontier of open keeps paced with closed:

It’ll take years for this equilibrium to become more obvious, seen through the lens of more open model families coming and going. This year, it seems likely we’ll see Nvidia’s Nemotron reach new heights, Reflection AI challenge some of the Chinese models with a strong, large MoE, maybe Meta releases a new open-weight model, and so on. True pressure to change strategy will only come when the capital environment punishes the less efficient spend on resources (e.g. giving away your competitive advantage, in having an in-house model). This pressure will likely hit Chinese startups training these models first.

All of Moonshot AI, MiniMax, and Zhipu AI will show signs of financial challenge in the coming years if they retain their strategy, on top of their models falling further behind the best open models in terms of generality. This is inevitable pressure to evolve open models to areas that are profitable and complementary of the frontier of AI.

Nvidia, which is best positioned to support the open ecosystem in the near term to support its core GPU business, could face many pressures to pull back its open model efforts. It could:

* Realize it’s too competitive to their biggest customers as they succeed too much with Nemotron,

* Fall to competition on their core business and lose the free cash flow buffer needed to fund this (e.g. it’s 2031 and OpenAI, Anthropic, Google, and the other frontier labs are worth so much they build their own chips).

* Start succeeding beyond their initial goals and keep the chips for them to build ASI themselves, as a closed-weight model.

The pressures for new funding mechanisms for open models are based on the assumptions of continued, substantive progress on the capabilities of frontier models. Mechanisms such as self-improvement and scaling all stages of the training pipeline are underway. This progress of capabilities will only increase the potential profit in selling models as and in products, not giving them away. The scale of investment required has already begun to push away non-profits from the game of making truly frontier-scale models. Capitalism is designed to make companies ruthless and chase down leads on profitability, not donate technology as charity.

As the economic environment shifts companies away from releasing the strongest models openly, more companies that rely on these models will look for an outlet of securing model access into the future. This is going to be compounded by a growing group of companies who come to rely on open-weight models for their workflows.

These points loop back into how model training is getting more expensive, so where desire to have the models will go up, ability to procure them will go down for many players. There are x-factors that could multiply the demand for institutions to ensure the existence of open models, such as the best frontier models not even being available via API (such as if Claude Mythos never goes general access).

As training relevant models is shifting to cost billions of dollars, rather than millions, few companies well be able to afford it. many companies will bite at the cost of paying 1/10th of the cost to train a frontier model, or if the consortium works, 1/50th. The upside for companies will be some mechanism to steer development (e.g. model sizes) or getting early access to develop internal and open-source tooling for the model.

It is in my nature to, by default, say this idea will fail, as training models is inherently a complex and high-focus endeavor, one that requires integration of every part of the stack and focusing specifically on your own vision and needs, rather than trying to serve every possible user. Eventually the need for open intelligence — and economic pressure to build it — will make a model consortium inevitable.

With the announcement of the Claude Mythos model this week and the admittedly very strong stated abilities, especially in cybersecurity, a new wave of anti open-weight AI model narratives surged. The TL;DR of the argument is that our digital infrastructure will not be ready in time for an open-weight version of this model, which will allow attacks to be conducted by numerous parties.

The backlash against open models in the wake of the Mythos news conflates too many general unknowns into a simple, broad policy recommendation that could actually further weaken cybersecurity readiness.

We’ve been here before – open-weight models were discussed as being extremely dangerous when OpenAI withheld GPT-2 weights in 2019, and when OpenAI released GPT-4 in 2023. Both of these waves came and went. The core mistake that is being made is the composition of two issues: 1) the acceptance of the open-closed model gap being static in time and 2) linking open-weight viability generally to specific issues.

I’ve written at length recently on how I think that the best, frontier-level open weight models are going to fall behind the best closed models in overall capabilities in the near future. I’ve also written about how the open-weight ecosystem needs to adapt to accept this reality. This is one of the times for the AI industry where I will repeat that it’s a total blessing to have the 6-18 month delay from when a certain capability is available within a closed lab to it being reproduced in the open. It’s a good balance of safety and monitoring the frontier of AI systems while allowing a useful open-source ecosystem to exist and thrive.

The core argument I’ve focused on in the open-closed model time gap has been in general capabilities – i.e. for general purpose, frontier models such as Claude Opus 4.X or GPT Thinking 5.X. The abilities of these closed models to robustly solve and work in diverse situations as agents remains out of scope of the best open-weight models. What the open-weight models have tended to be better at is quickly keeping pace on key benchmarks (which admittedly is helped to some extent, but not necessarily substantially by distillation). This discussion is entirely different, it has to do with if open weight models can keep pace on the specific skills related to cybersecurity, and when we could expect an open version of this model to be available to the world.

The case of a Claude Mythos level open weight model is admittedly more nuanced to me than the previous few anti-open weight narratives the community has experienced. Where GPT-4 was about a more hypothetical risk, especially in areas like bio-risk, the clear and present reality of cyber infrastructure being prone to attack is far more tangible. Still, much of this nuance in the moment comes down to not knowing the full details of what the system can actually do (i.e. Mythos), and the state of the environment it would act in (i.e. our digital infrastructure).

To properly assess this risk, we need to know what it takes to build and deploy a Claude Mythos scale model. This entails three pieces: 1) training and releasing the weights, 2) the harness that gives the model effective tools it knows how to use, and 3) the inference compute and software.

(Below I make some model size & price estimates to show my thinking, these should not be taken as ground truth.)

Current estimates put the size ranges of leading models like Claude Opus 4.6 or GPT 5.4 as being around 3-5T parameters. Currently, the largest open-source models, which have been coming from Chinese labs, are around 1T parameters. Claude Mythos’s preview pricing is 5X Opus, which could come from a simple multiplicative increase in active parameters (with the same serving system design), far higher inference-time scaling, more complex harnesses that make inference less efficient, lower utilization expectations, and so on. The simplest guess is that it’s a mix of all of the above, something like 2X bigger in parameters and much less efficient to serve. That’s a huge model, likely something similar to GPT 4.5, but actually post-trained well (GPT 4.5 was ahead of its time, infra-wise).

With size comes the challenge actually training the model, as bigger models always come with new technical problems that must be solved to unlock the capabilities. For the case of cybersecurity, my guess is that most of the capabilities can be learned by training a model to be superhuman on coding. Unlike some capabilities such as knowledge work, medicine, law, etc., coding can be studied and improved substantially with public data like GitHub. I’m far more optimistic in open-weight models staying fairly close to the frontier in narrow domains of code execution and processing, but I don’t understand the full scope of skills needed to be superhuman in cybersecurity understanding. How much expert knowledge and special sauce went into training Claude Mythos? That’s a substantial source of my error bars on the impact.

Second, we know nothing about how the model works under the hood. Today, models are complex systems that entail far more than just weights. They require complex tools and infrastructure to run them, of which Claude Code is the one we are most used to. Mythos very likely has its own innovations here.

My estimate for how many GPUs you’d need to serve an 8T parameter, modern MoE is something like O(100) H100 GPUs, which costs something like $10K a day (and this may be very slow in terms of tok/s). Heck, the official marketing copy of the Nvidia GB200 VL72 system is “Unlocking Real-Time Trillion-Parameter Models” on the rack. Does Mythos fit on one rack? The point isn’t to rely on my specific estimate as a policy reference, but to repeat that running leading AI systems is very expensive and not something you can just do on a laptop or self-service cloud portals.

There are far fewer actors who can get their hands on these resources, relative to those who can download the model. Of course, there are still many, but it’s important to flesh out all the details of what it would take to proliferate the capabilities of a Mythos-like model. In summary, tools like Mythos will make the best attackers have more powerful tools of the trade, but it won’t be handing a nuke to every teenager connected to the internet.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Personally, I do acknowledge there’s a chance that cybersecurity abuse is a red line that makes releasing open-weight text models above a certain capability threshold morally grey. Many people thought this red line would come far earlier, somewhere in between GPT-2 and GPT-4, through the harm axis of mis/disinformation, but that had different bottlenecks. For image generation models, we’re well past the first red line which is enabling non-consensual AI deepfakes with readily available open-weight models. We’re balancing the reality of these fears having come and gone before with a technology that’s becoming increasingly capable.

So, my second large source of error bars is “how bad is it actually” with respect to the state of cybersecurity. How much can humans clean up in the most important software with months of private access to a model like Claude Mythos? What will never get fixed?

For example, if we get open-weight models that are close to the capabilities of Claude Mythos, could those be fine-tuned by organizations to harden the security of their tools?

Currently, it’s too soon to call it as a general reason to stop progress in open models. When Claude Mythos is closed to so few partners, in some ways having strong open models close to the threshold makes assessing the danger easier. Having to rely fully on a single private company to determine the security of essential, international infrastructure is not a tenable equilibrium.

So, in conclusion, I urge people to further study three things:

* How do we measure cybersecurity related capabilities across open and closed models. With this, are open models truly keeping up at a 6-9month lag, or are they only maintaining performance relevance in other areas of coding?

* How do we independently measure the true impact of Claude Mythos and Project Glasswing on existing cybersecurity concerns?

* If it is the case that the models are keeping up and the defensive capabilities of Claude Mythos are weak, how do we better monitor (and if needed, try to regulate) the targeted capabilities of open-weight models in narrow domains?

The goal is to encourage fears about open models remaining very specific. Any general ban on open models in a nation will immediately and likely irrevocably remove that entity’s ability to influence a crucial, and amorphous technology. If we stop building the best open models in the U.S., then another country will do this and become the center of the technology. There’s no way to fully kill open models, only influencing, understanding, and steering.

Having written a lot of model release blog posts, there’s something much harder about reviewing open models when they drop relative to closed models, especially in 2026. In recent years, there were so few open models, so when Llama 3 was released most people were still doing research on Llama 2 and super happy to get an update. When Qwen 3 was released, the Llama 4 fiasco had just gone down, and a whole research community was emerging to study RL on Qwen 2.5 — it was a no brainer to upgrade.

Today, when an open model releases, it’s competing with Qwen 3.5, Kimi K2.5, GLM 5, MiniMax M2.5, GPT-OSS, Arcee Large, Nemotron 3, Olmo 3, and others. The space is populated, but still feels full of hidden opportunity. The potential of open models feels like a dark matter, a potential we know is huge, but few clear recipes and examples for how to unlock it are out there. Agentic AI, OpenClaw, and everything brewing in that space is going to spur mass experimentation in open models to complement the likes of Claude and Codex, not replace them.

Especially with open models, the benchmarks at release are an extremely incomplete story. In some ways this is exciting, as new open models have a much higher variance and ability to surprise, but it also points at some structural reasons that make building businesses and great AI experiences around open models harder than the closed alternatives. When a new Claude Opus or GPT drops, spending a few hours with them in my agentic workflows is genuinely a good vibe test. For open models, putting them through this test is a category error.

Something else to be said about open models in the era of agents is that they get out of the debate of integration, harnesses, and tools and let us see close to the ground on what exactly is the ability of just a model. Of course, we can’t test some things like search abilities without some tool, but being able to measure exactly the pace of progress of the model alone is a welcome simplification to a systematically opaque AI space.

The list of factors I’d use to assess a new open-weight model I’m considering investing in includes:

* Model performance (and size) — how this model performs on benchmarks I care about and how it compares to other models of a similar size.

* Country of origin — some businesses care deeply about provenance, and if a model was built in China or not.

* Model license — if a model needs legal approval for use, uptake will be slower at mid-sized and large companies.

* Tooling at release — many models release with half-broken, or at least substantially slower, implementations in popular software like vLLM, Transformers, SGLANG, etc due to pushing the envelope of architectures or tools.

* Model fine-tunability — how easy or hard it is to modify the given model to your use-case when you actually try and use it.

The core problem is that some of these are immediately available at release, e.g. general performance, license, origin, etc. but others such as tooling take day(s) to week(s) to stabilize, and others are open research questions — with no group systematically monitoring fine-tunability.

In the early era of open models, the days of Llama 2 or 3 and Qwen pre v3.5, the architectures were fairly simple and the models tended to work out of the box. Some of this was due to the extremely hard work of the Llama, Qwen, Mistral, etc. developer teams. Some is due to the new models being genuinely harder to work with. When it comes to something like Qwen 3.5 or Nemotron 3, with hybrid models (either gated delta net or mamba layers), the tooling is very rough at release. Things you would expect to “just work” often don’t.

I’ve been following this area closely since we released Olmo Hybrid with a similar architecture, and Qwen 3.5 is just starting to work well in the various open-source tools that need to all play nice together for RL research. That’s 1.5 months after the release date! This is just to start really investing more into understanding the behavior of the models. Of course, others started working on these models sooner by investing more engineering resources or relying on partially closed software. The fully open and distributed ecosystem takes a long time to get going on some new models.

All of this is lead-in for the most important question for open models — how easy is it to adapt to specific use-cases? This is a different problem for different model sizes. Large MoE open-weight models may be used by entities like Cursor who need complex capabilities in their domain, e.g. Composer 2 trained on Kimi K2.5. Other applications can be built on much smaller models, such as Chroma’s Context-1 model for agentic search, built on GPT-OSS 20B.

The question of “which models are fine-tunable” is largely background knowledge known by engineers across the industry. There should be a thriving research area here to support the open ecosystem model. The first step is to understand characteristics of different base and post-trained models to understand what they look like. The second step is to tune pretraining recipes for open models so they’re more flexible.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

For The ATOM Project and other Interconnects endeavors, we’ve put in substantial effort to measuring adoption trends in the open ecosystem. Everything takes a long time to unfold after a model is first publicly available — and adaptability is why. What we know for sure now, when Qwen has been going from strength to strength with its releases, is that technical staff across the industry has gotten comfortable working with Qwen models. Countless research methods and datasets were made to work with Qwen. It’ll take patience for any other model family to get to this point — a patience I’m not sure many open model builders have.

This takes us to Gemma 4, Google’s latest open models. Gemma 3 was released more than a year ago, in March of 2025, and is a bit underrated. Gemma 4 comes in 4 sizes for now, with a bigger, MoE model of over 100B total parameters rumored but not released yet. The models we have today come in sizes of ~5B dense, 8B dense, 26B total 4B active MoE, and 31B dense.

I’m most excited that they’re finally adopting a standard Apache 2.0 open source license. This’ll massively boost adoption. The standard of better licenses for strong open-weight LLMs was set by mostly Chinese open model labs in the last 1-2 years, and now U.S. companies are following suit. I will personally be so happy if the horrible Llama licenses and Gemma terms of service were an ~18-month transient dynamic of the industry being nervous about releasing strong open models.

The Gemma 4 scores look very solid, the small models have incredible benchmark scores (especially in general domains like LMArena) and the 31B model rivals the recent Qwen 3.5 27B, which is the leading member of that class. The ~30B size range is an important one, as it’s accessible both to researchers and to enterprises looking to deploy the model in real use-cases. Where the 7B model scale is the default for tinkering and research, a 30B model is the default for seeing if an open model can unlock substantial value in your specific workflow — a good mix of intelligence, low price, tractability for downstream training, etc.

This takes us back to the above adoption criteria I mentioned for open models and the bigger question — do I think Gemma 4 will be an overwhelming success? Previous Gemma models have been plagued by tooling issues and poorer performance when being finetuned.

Gemma 4’s success is going to be entirely determined by ease of use, to a point where a 5-10% swing on benchmarks wouldn’t matter at all. It’s strong enough, small enough, with the right license, and from the U.S., so many companies are going to slot it in.

I’m cautiously optimistic that Gemma 4 is going to work better here. Winds are shifting for open models built in America. We saw GPT-OSS go through a bumpy launch to become an overwhelming success. There’s a collective energy around the likes of Reflection, Arcee, Nemotron, Gemma, Olmo, and peers that show substantial demand for building new stacks around open models. There’s capital to be spent on AI stacks across the economy by those who want more ownership of everything, including the model.

After launching The ATOM Project 240 days ago, the conversation is shifting into the next stage. Summer of 2025 was a crisis moment where the U.S. AI scene realized it can’t wait and figure out open models after building AGI. The two markets will capture different areas and proceed in parallel. Now that more companies in the U.S. are releasing strong models, we need to improve the ecosystem so that these models are easy to use, understand, and build value around. It’s the hard work to build another inflection point in these adoption plots I’ve been updating consistently, but that’s the work to be done. Join me in it.

More data coming soon! Here’s a sneak peek:

Fast takeoff, the singularity, and recursive self-improvement (RSI) are all top of mind in AI circles these days. There are elements of truth to them in what’s happening in the AI industry. Two, maybe three, labs are consolidating as an oligopoly with access to the best AI models (and the resources to build the next ones). The AI tools of today are abruptly transforming engineering and research jobs.

AI research is becoming much easier in many ways. The technical problems that need to be solved to scale training large language models even further are formidable. Super-human coding assistants making these approachable is breaking a lot of former claims of what building these things entailed. Together this is setting us up for a year (or more) of rapid progress at the cutting edge of AI.

We’re also at a time where language models are already extremely good. They’re in fact good enough for plenty of extremely valuable knowledge-work tasks. Language models taking another big step is hard to imagine — it’s unclear which tasks they’re going to master this year outside of code and CLI-based computer-use. There will be some new ones! These capabilities unlock new styles of working that’ll send more ripples through the economy.

These dramatic changes almost make it seem like a foregone conclusion that language models can then just keep accelerating progress on their own. The popular language for this is a recursive self-improvement loop. Early writing on the topic dates back to the 2000s, such as the blog post entirely on the topic from 2008:

Recursion is the sort of thing that happens when you hand the AI the object-level problem of “redesign your own cognitive algorithms”.

And slightly earlier, in 2007, Yudkowsky also defined the related idea of a Seed AI in Levels of Organization in General Intelligence:

A seed AI is an AI designed for self-understanding, self-modification, and recursive self-improvement. This has implications both for the functional architectures needed to achieve primitive intelligence, and for the later development of the AI if and when its holonic self-understanding begins to improve. Seed AI is not a workaround that avoids the challenge of general intelligence by bootstrapping from an unintelligent core; seed AI only begins to yield benefits once there is some degree of available intelligence to be utilized. The later consequences of seed AI (such as true recursive self-improvement) only show up after the AI has achieved significant holonic understanding and general intelligence.

It’s reasonable to think we’re at the start here, with how general and useful today’s models are.

Generally, RSI can be summarized as when AI can improve itself, the improved version can improve even more efficiently, creating a closed amplification loop that leads to an intelligence explosion, often referred to as the singularity. There are a few assumptions in this. For RSI to occur, it needs to be that:

* The loop is closed. Models can keep improving on themselves and beget more models.

* The loop is self-amplifying. The next models will yield even bigger improvements than the current ones.

* The loop continues to run without losing efficiency. There are not added pieces of friction that make the exponential knee-capped as an early sigmoid.

While I agree that momentous, socially destabilizing changes are coming in the next few years from sustained AI improvements, I expect the trend line of progress to be more linear than exponential when we reflect back. Instead of recursive self-improvement, it will be lossy self-improvement (LSI) – the models become core to the development loop but friction breaks down all the core assumptions of RSI. The more compute and agents you throw at a problem, the more loss and repetition shows up.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

I’m still a believer that the complexity brake on advanced systems will be a strong counterbalance to the reality that AI models are getting substantially better at every narrow task we need to compose together in making a leading AI model. I quoted this previously in April of 2025 in response to AI 2027.

Microsoft co-founder Paul Allen argued the opposite of accelerating returns, the complexity brake: the more progress science makes towards understanding intelligence, the more difficult it becomes to make additional progress. A study of the number of patents shows that human creativity does not show accelerating returns, but in fact, as suggested by Joseph Tainter in his The Collapse of Complex Societies, a law of diminishing returns. The number of patents per thousand peaked in the period from 1850 to 1900, and has been declining since. The growth of complexity eventually becomes self-limiting, and leads to a widespread “general systems collapse”.

There are plenty of examples in how models are already trained, the deep intuitions we need to get them right, and the organizations that build them that show where the losses will come from. Building leading language models is incredibly complex, and only becoming more-so. There are a few core frictions in my mind.

1. Automatable research is too narrow

First, it is clear that language models this year will already be useful tools at optimizing localized tasks like lowering the test loss of a model. Andrey Karpathy recently launched his autoresearch that popularized doing just this. This allows AI agents to play directly on GPUs to target tasks like lowering the loss on the test set. This approach works in narrow domains, i.e. one general test loss or one overall reward. The problem is that there’s a long-standing gap between an on-paper more accurate model and models that users find more productive. The most provocative case is for pretraining, which was discussed more at length around scaling laws. Scaling laws show us that the loss will continue going down, but we don’t know if that’ll be economically more valuable.

In post-training, reinforcement learning algorithms are at least more directly tied to specific performance gains as most RL training environments can be used directly as an evaluation. Still, I worry about generalization and tying back to models that are better at the specific task of improving themselves. It’s a big leap from models get better at some things to that necessarily translating to models that are better at building themselves and designing experiments. We’ve seen many AI capabilities sort of saturate at certain levels of human taste, such as writing quality. AI research is a bit different here, as there is a very high ceiling to climb up to. Where models mostly saturate on writing because there’s inherent tension in preferences, models will saturate on research because the search space and optimization target is too wide.

The early benchmarks for measuring this sort of ability all fall prey to the same problem – narrow scope. Agents will do well at optimizing single metrics, but the leap required to navigate many metrics at once is a very different skill set. That is actually what the best researchers do — they make many scalable ideas work together.

The most related benchmark we have to measure this is PostTrainBench, which is quite fun, but progress will very rapidly get distorted on this. Over 90% of the challenge in doing post-training well is getting the last 1-3% of performance, especially without cooking the model in out-of-domain tasks. Post-training a general, leading model is extremely complex, and only getting more complex.

I could go on and on about this. Another example is from during my Ph.D. (2017-2022), when there was immense hype around a field called “AutoML” which aimed to use techniques like Bayesian Optimization to find new architectures and parameters for models. The hype never translated into changing my job. Language models will do more than this, but not enough to take jobs away from top AI researchers any time soon. The core currency of researchers is still intuition and managing complexity, rather than specific optimization and implementation.

2. Diminishing returns of more AI agents in parallel

The biggest problem for rapid improvement in AI is that even though we’ll have 10,000 remote workers in a datacenter, it’ll be nearly impossible to channel all of them at one problem. Inherently, especially when the models are still so similar, they’re sampling from the same distribution of solutions and capabilities while being bottlenecked by human supervision. Adding more agents will have a strict saturation in the amount of marginal performance that can be added – the intuition of the best few researchers (and time to run experiments) will be the final bottleneck.

A common idea to illustrate this is Amdahl’s law, which is taken from computer architecture and shows that a given task can only generate a fixed speedup proportional to how much can be parallelized and how many parallel workers exist. An illustration is below:

In AI this should be relatively easier to convey, as the low-level operating details of computers are fairly mysterious. Consider an AI researcher on the transition from writing code by hand to using AI autocomplete assistance to now using autonomous coding agents. These are all massive gains. Let us continue. Now this researcher uses 3-4 agents working on different sub-tasks or approaches to the problem at hand. This is still a large gain. Now consider this single researcher trying to organize 30-40 agents with tasks to do every day. Some people can get more value out of this scale, but not many.

How many people do you think could come up with 300-400 tasks for AI agents every day? Not many. This problem will hit the AI models soon enough as well.

3. Resource bottlenecks and politics

Fundamentally, all the AI companies are walking a fine line of acquiring substantial capital, converting new compute resources to revenue via sufficient demand, and repeating the process all-the-while spending an extreme amount on research. With the scale of resources here, there will always be political bottlenecks on who gets resources and what gets bet on. In this layer, research leadership sits above the AIs and the researchers. Even as models continue to improve, this source of friction will never get removed. It isn’t a substantial friction, but the AI models are fundamentally operating in organizations where humans are the bottleneck on resources.

The early scale of improvements with language models is local optimizations, where the resources used cost <$1M per day. With my other views on the frictions of AI, this is on its own a very minor impact on the rate of improvement, but for those with worries of fast take-off, RSI, and loss of control to AIs, it should be obvious that billions of dollars of compute resources for research are unlikely to be totally isolated for end-to-end experimentation of AI models.

The conclusion here is that because we’re at the early stages of using AI assistance, autonomously and at scale for AI-development, we’re collectively discovering the ways that AI can help us massively. We’re all applying these tools to capture the low-hanging fruit we see and our jobs are literally changing to be higher paced and more productive. The problem is that all of these axes have clear human, political, or technical complexity bottlenecks.

The bottom of every sigmoid feels like an exponential. We’ve ridden multiple exponentials in the era of language models, in 2023 we scaled to huge models and GPT-4 felt like magic, by 2025 we added inference-time scaling with o1 and reasoning models — they let us “solve” math and coding, now we’re going to take a big step by polishing the entire AI workflow (all the while scaling training compute massively). 2026 will feel like a huge step, but it doesn’t have a fundamental change convincing me that progress will begin to take off.

This could still cross the colloquial threshold for AGI, which is a drop-in replacement for most remote workers, which would be an incredible milestone. Much of the challenge in the debate of if we hit AGI in the coming years is that AI models are jagged and smart in different ways than humans, so they won’t look like drop-in replacements for remote workers, but in many cases just using AI will be far more effective than trying to work with a human. It’s reshaping what jobs are.

Let us consider the scenarios we’re working through.

* Engineering is becoming automated today. Humans are way more productive, models can scale through complex infrastructure deployments much faster, run with higher GPU utilization, etc. Infrastructure gains become fixed improvements in the rate and scale of experimentation, the fundamental units of progress in AI.

* Basic AI model research and optimization will be automated. The AI models are expanding in scope – they transition from writing kernels to deciding on architectures. This is moving from improving the experimentation toolkit to running minor experiments themselves. Configs, hyperparameters, etc. become the domain of the AI assistants.

These are both real. The problem is that a third era doesn’t have a simple scale to jump to. Where the AI models can create knowledge by synthesis and execution, the next jump requires harnessing thousands of agents or having models make more novel discoveries – like unlocking the next paradigm after inference time scaling. The improvements downstream of AI are going to make the industry supercharged at hill climbing, but I worry that this won’t bring paradigm shifts that are needed for new categories of AI – continual learning, world models, whatever your drug of choice is.

All together, the models are becoming core to the development loop and that’s worth being excited (and worried) about. The models are performing self-improvement. They’re not transforming the approach. We are scaling up the compute we spend on our own research practices and tools. There are diminishing returns. Agents are going to start being autonomous entities we work with. They feel like a cross between a genius and a 5 year old. We will be in this era of lossy self-improvement (LSI) for a few years, but it is not enough for a fast takeoff.

I’m a little late to this model review, but that has given me more time to think about the axes that matter for agents. Traditional benchmarks reduce model performance to a single score of correctness – they always have because that was simple, easy to quickly use to gauge performance, and so on. This is also advice that I give to people trying to build great benchmarks – it needs to reduce to one number that is interpretable. This is likely still going to be true in a year or two, and benchmarks for agents will be better, but for the time being it doesn’t really map to what we feel because agentic tasks are all about a mix of correctness, ease of use, speed, and cost. Eventually benchmarks will individually address these.

Where GPT 5.4 feels like another incremental model on some on-paper benchmarks, in practice it feels like a meaningful step in all four of those traits. GPT 5.4 in Codex, always on fast mode and high or extra-high effort, is the first OpenAI agent that feels like it can do a lot of random things you can throw at it.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

I haven’t been particularly deep in software engineering over the last few months, so most of my working with agents has been smaller projects (not totally one-off, but small enough where I’ve built the entire thing and manage the design over weeks), data analysis, and research tasks. When you embrace being agent-native, this style of work entails a lot of regular APIs, background packages (like installing and managing LateX binaries, ffmpeg, multimedia conversion tools, etc), git operations, file management, search etc. Prior to GPT 5.4, I always churned off of OpenAI’s agents due to a death by a thousand cuts. It felt like rage quits. I’d feel like I was getting into GPT 5.2 Codex, but it would fail on a git operation and have me (or Claude) need to reset it. Those hard edges are no longer there.

The other subtle change in GPT 5.4’s approachability – the biggest reason I think OpenAI is much more back in the agent wars – is that it just feels a bit more “right.” I classify this differently to the routine tasks I discussed above, and it has to do with how the product (i.e. the model harness) presents the model outputs, requests, and all that to you the user. It has to do with how easy it is to dive in. This has always been Claude’s biggest strength in its astronomical growth. Not only has Claude been immensely useful, but it has a charm and entertainment value to it that’ll make new people stick around. GPT 5.4 has a bit of that, but the underlying model strengths of Claude still leave it feeling warmer.

Where Claude is a super smart model, with character, a turn of phrase in a debate, and sometimes forgetting something, OpenAI’s models in Codex feel meticulous, slightly cold, but deeply mechanical. I’d use Claude for things I need more of an opinion on and GPT 5.4 to churn through an overwhelmingly specific TODO list. The instruction following of GPT 5.4 is so precise that I need to learn to interact with the models differently after spending so much time with Claude. Claude, in some domains, you come to see has an excellent model for your intent. GPT 5.4 just does what you say to do. These are very different philosophies of “what will make the best model for an agent”, Claude will likely appeal to the newcomers, but GPT 5.4 will likely appeal to the master agent coordinator that wants to unleash their AI army on distributed tasks.

Outside of charm, and dare I say taste, a lot of the usability factors are actually better on OpenAI’s half of the world. The Codex app is compelling – I don’t always use it, but sometimes I totally love it. I suspect substantial innovation is coming in what these apps look like. Personally, I expect them to eventually look like Slack (when multiple agents need to talk to eachother, under my watch).

OpenAI also natively offers fast mode for their models with a subscription and very large rate limits. I’ve been on the $100/month Claude plan and $200/month ChatGPT plan for quite some time. I’ve never been remotely close to my Codex limits with fast mode and xhigh reasoning effort, where I hit my Claude limits from time to time. There’s definitely a modeling reason to this – most of OpenAI’s release blogs showcase each iterative model being substantially more concise in the number of tokens it takes to get peak benchmark performance. This is a measure of reasoning efficiency. This 2D (or more) benchmark picture is exactly where the world is going.

Here’s a plot from Cursor, which sadly doesn’t have all the GPT 5.4 reasoning efforts, but it confirms this point in a third party evaluation. What is missing across model families is the speed and price (a proxy for total compute used) to get there.

The final benefit of GPT 5.4, and OpenAI’s agentic models in general for that matter, is much better context management. In using them regularly now I feel like I’ve never hit the context wall or context anxiety point. The reasoning efficiency I suspect is the case above just lets the model do way more with its initially empty context window. Then, when GPT 5.4 does compact, it’s been less noticeable.

The one problem I’ve been having with both Claude Opus 4.6 and GPT 5.4 is a light forgetfulness. If you give the models multiple TODOs in a single message outside of planning mode, I find them often dropping them. Sometimes it feels like the models glitch and try to solve a previous problem rather than the recent ones. I’m not sure what in the model or the harness is the exact cause, but sometimes I like to queue up a few messages as I see the model working on something, to refine the task, but currently this tends to be a pretty risky outcome except in the simplest use-cases.

These days I’ve been using both GPT and Claude extensively, mostly based on my mood, and have been getting more done than ever. Having a GPT 5.4 Pro integration directly with Codex, e.g. like \ultrathink, would be a big differentiator for OpenAI. Those models have been incredible.

All in, I see GPT 5.4 as an agentic model that brings a ton more simple usability and “agentness” to the very strong software foundation of GPT 5.3 Codex. It’s a big step, and I’m unbelievably excited for which of these two companies releases an update next. On paper, listing the strengths of GPT 5.4 across better top end coding performance, better speed, better context management, better rate limits, it’s a testament to how nuanced choosing a model is. I genuinely still enjoy Claude a bit more for ways that’ll never show up on benchmarks. This makes me type claude into my terminal at the start of my day, rather than codex.

2025 was the year where a lot of companies started to take open models seriously as a path to influence in the extremely valuable AI ecosystem — the adoption of a strategy that was massively accelerated downstream of DeepSeek R1’s breakout success. Most of this is being done as a mission of hope, principle, or generosity.

Very few businesses have a real monetary reason to build open models. Well-cited reasons, such as commoditizing one’s complements for Meta’s Llama, are hard to follow up on when the cost of participating well is billions of dollars. Still, AI is in such an early phase of technological development, mostly defined by large-scale industrialization and massive scale-out of infrastructure, that having any sort of influence at the cutting edge of AI is seen as a path to immense potential value.

Open models are a very fast way to achieve this, you can obtain substantial usage and mindshare with no enterprise agreements or marketing campaigns — just releasing one good model. Many companies in AI have raised a ton of money built on less.

The hype of open models is simultaneously amplified by the mix of cope, disruptive anticipation, and science fiction that hopes for the world where open models do truly surpass the closed labs. This goal could be an economically catastrophic success for the AI ecosystem, where profits and revenue plummet but the broader balance of power and control of AI models is long-term more stable.

There’s a small chance open models win in absolute performance, but it would only be on the back of either a true scientific breakthrough that is somehow kept hidden from the leading labs or the models truly hitting a wall in performance. Both of them are definitely possible, but very unlikely.

It is important to remind yourself that there have been no walls in progress to date and all the top AI researchers we discuss this with constantly explain the low-hanging fruit they see on progress. It may not be recursive self-improvement to the singularity (more on that in a separate post), but large technology companies are on a direct path to building definitionally transformative tools. They are coming.

The balance of power in open vs. closed models

The fair assessment of the open-closed gap is that open models have always been 6-18 months behind the best closed models. It is a remarkable testament to the open labs, operating on far smaller budgets, that this has stayed so stable. Many top analysts like myself are bewildered by the way the gap isn’t bigger. Distillation helps a bit in quality, benchmaxing more than closed labs helps perceptions, but the progress of the leading open models is flat out remarkable.

The reality is that the open-closed model gap is more likely to grow than shrink. The top few labs are improving as fast as ever, releasing many great new models, with more on the docket. Many of the most impressive frontier model improvements relative to their open counterparts feel totally unmeasured on public benchmarks.

In a new era of coding agents, the popular method to “copy” performance from closed models, distillation, requires more creativity to extract performance — previously, you could use the entire completion from the model to train your student, but now the most important part is the complex RL environments and the prompts to place your agents in them. These are much easier to hide and all the while the Chinese labs leading in open models are always complaining about computational restrictions.

As the leading AI models move into longer-horizon and more specialized tasks, mediated by complex and expensive gate-keepers in the U.S. economy (e.g. legal or healthcare systems), I expect large gaps in performance to appear. Coding can largely be mostly “solved” with careful data processes, scraping GitHub, and clever environments. The economies of scale and foci of training are moving into domains that are not on the public web, so they are far harder to replicate than early language models.

Developing frontier AI models today is more defined by stacking medium to small wins, unlocked by infrastructure, across time. This rewards organizations that can expand scope while maintaining quality, which is extremely expensive.

All of these dynamics together create a business landscape for open models that is hard to parse. Through 2026, closed models are going to take leaps and bounds in performance in directions that it is unlikely for open models to follow. This sets us up for a world where we need to consider, fund, use, and discuss open models differently. This piece lays out how open models are changing. It is a future that’ll be clearly defined by three classes of models.

* True (closed) frontier models. These will drive the strongest knowledge work and coding agents. They will be truly remarkable tools that force us to reconsider our relationship to work.

* Open frontier models. These will be the best open-weight, large models that are attempting to compete on the same directions as above. There will be plenty of use-cases that they don’t work for relative to the best models, but countless use-cases where they work remarkably well. For many use-cases, even ones as valuable as some subsets of coding, these will work great. The AI ecosystem will still take years to understand what it means to have intelligence of this magnitude served in private, at the marginal cost of electricity for individuals, as assistants, coaches, companions, and more. OpenClaw provided a glimpse behind the mirror that will expand and grow. The class of models around GPT-OSS 120B, Nvidia Nemotron 3 Super, or MiniMax M2.5 are the balance of performance to price that can work as local models.

* Open, small models as distributed intelligence. The most successful open models will be complementary tools to closed agents. This is a path for open models to complement and accelerate the frontier of progress.AI is slotting in to automate many repetitive, niche tasks across the technology economy. There’s a huge pressure to shift these tasks off of the best closed models — which frankly are still better at most of the things, across my conversations with businesses trying to build with open models — to small, open models that can be 10X faster and 100X cheaper. There aren’t really people building data and fine-tuning engines for economically viable tasks on the smallest models possible. These models need to be almost brain-numbingly boring and specific. In a world dominated by coding agents, I want to build open models that Claude Code is desperate to use as a tool, letting its sub agents unlock entirely new areas of work. This is possible, but remarkably under-explored. Small models from the likes of Qwen and co. are still marketed on general-task benchmarks. The hype of “open models catching the frontier” distracts the world from this very large area of demand.This is the sort of model that moves open models from just a few, crucial static weights to more of an ecosystem. It requires creativity and a new approach. The goal of this piece is to illustrate why and how to build these, with added context on where open models stand today.

All three of these model classes hint at different ways to use agents. It is absolutely definitional to how AI is going to be built going forward that they’re not just model weights, but rather systems that think, search, and act. The weights only define one portion of those abilities.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Open weights as part of an AI system

To start, consider what are the most impactful and impressive things that language models can do without a suite of tools at their side. When was the last time that you were blown away by something that was just autoregressive token outputs? Unless you’re doing a substantial amount of work on mathematical proofs or competition code, it seems like that situation has changed little since GPT-4’s release in 2023. The AI systems we use today are about far, far more than weights.

In this world, closed models have a clear advantage. Closed models get to vertically integrate everything from the chips they run on, the inference software, the weights, the tools, and the user interface. Open models on the other hand need to work on every inference setup, with many tools, and in many use-cases. This vertical integration is best expressed today in the joy of using Claude Code with Opus 4.6 or OpenAI’s Codex with GPT 5.4. Open models haven’t passed this point. Some are starting to focus on specific interfaces, e.g. OpenCode, but there’s an inherent tension in making an open model work only in your blessed product roadmap.

At the same time, this change could point to more about the latest AI systems being open! If you can do less with the weights alone, maybe more labs will release them.

The way to think about AI systems today is as a mix of weights, tools, and harnesses. The weights portion is familiar. The tools are the deeply integrated environments the models act in at deployment time — best typified by search and code sandboxes — and the harness is how these two fit together with a product that the user sees.

In this world, there are two things to consider: 1) Is there an equivalent, open system to the closed products that people are using today — I mean truly equivalent, where every level of the stack can be modified and controlled (more on this later), and 2) How does this system’s view impact different future decisions in the open ecosystem?

Still looking for open model business strategies

To understand how the business and practicality of open models will evolve, let me take a tour back in time to foundational writing on the role of open-source in modern technology companies. The first is a Google blog post, The Meaning of Open, which originally was an internal memo by Jonathan Rosenberg, which sparked an intense internal debate that later resulted in it becoming public. To start, here’s a basic assessment of how open systems can work:

Open systems have the potential to spawn industries. They harness the intellect of the general population and spur businesses to compete, innovate, and win based on the merits of their products and not just the brilliance of their business tactics.

I’ve long believed that the company who will benefit most from the ecosystem of open models is the one who understands it best. This entails being deeply involved with open research and experimentation in how to use the models. So far, most of the open model company business models are not this. Rosenberg expands on this in his 2009 post, comparing the dynamics of open systems to closed products:

[Open systems] are competitive and far more dynamic. In an open system, a competitive advantage doesn’t derive from locking in customers, but rather from understanding the fast-moving system better than anyone else and using that knowledge to generate better, more innovative products. The successful company in an open system is both a fast innovator and a thought leader; the brand value of thought leadership attracts customers and then fast innovation keeps them. This isn’t easy — far from it — but fast companies have nothing to fear, and when they are successful they can generate great shareholder value.

We’ve known for some time that open weight models are not actually enough to constitute a product — models are a product in the sense that they have tools and harnesses, so we don’t actually have fully open systems, we have systems that are partially open partially closed, making moats messy. VLLM and a model like GLM 5 are pieces of a system, but it still takes more to deploy them — expensive private GPUs and some tools with local business data.

It may turn out to be that AI is too complex and expensive to have any analogous open system to previous generations of technology. If there was a fully open system, it would win by default, as many historical generations of technology have shown us. This fully open analog does not yet exist, so we have constant debates on the role of open-source AI.

Bill Gurley recounts how Google’s free products have exemplified the open or free strategies across technology. Gurley wrote on the open-source operating system, Android, and the free browser, Chrome, in 2011:

So here is the kicker. Android, as well as Chrome and Chrome OS for that matter, are not “products” in the classic business sense. They have no plan to become their own “economic castles.” Rather they are very expensive and very aggressive “moats,” funded by the height and magnitude of Google’s castle. Google’s aim is defensive not offensive. They are not trying to make a profit on Android or Chrome. They want to take any layer that lives between themselves and the consumer and make it free (or even less than free).

Because these layers are basically software products with no variable costs, this is a very viable defensive strategy. In essence, they are not just building a moat; Google is also scorching the earth for 250 miles around the outside of the castle to ensure no one can approach it.

In the same post, Gurley reflects on the limits of Google’s openness:

In this open manifesto, Jonathan opines over and over again that open systems unquestionably result in the very best solutions for end customers. That is with one exception. “In many cases, most notably our search and ads products, opening up the code would not contribute to these goals and would actually hurt users.” As Rodney Dangerfield said in Caddyshack, “It looks good on you, though.”

Essentially, Google open-sourced so much, in fact paid people to use its products (e.g. paying phone makers to use android) to keep the funnel leading to the search profit center. This is the virtuous loop that the search business still funds to this day.

AI is still nothing like this, but signs of change are emerging. The default belief on the value of models to these companies is that the model is the product. This is obvious with products like hosted APIs, where releasing the model weights would be business suicide, but this is softening as interfaces like Claude Code, Codex, Cursor, etc. get vastly popular. It could be a path to more openness, at least in parts of the stack. We can see this with the coding plans offered by Moonshot and Z.ai — where the demand is very high for the businesses, even though the model is open. Most people will just use the cheap interface with inference, instead of figuring out how to use the model themselves (as long as the business is mostly consumer or per-head services).

All of this doesn’t leave me optimistic on the direction of companies becoming more open in the coming years. I’d expect the opposite still. Nvidia has the one great reason to be open — to sell more GPUs to people building on open models and understand what they need to build next, but there’s no one else obvious on this list. Until there are more specific economic reasons to build open models, the companies building these at the frontier will have fewer resources to spend on the models and face a consolidation to the best few.

In the face of consolidation at the open frontier, the investment in the models should shift to areas where the models can have more differentiated upside relative to the best closed frontier models.

Open models that are specific, cheap, fast, and ubiquitous

There’s too much obsession with the best companies building open models to try and compete at the frontier. There’s a vastly underserved market of enterprises that want cheap, reliable models for repetitive use-cases in their systems. Picture this, one small model with a series of LoRA adapters that specialize the model to internal skills. This can be deployed very cheaply as tools and a complement to the frontier closed models that are orchestrating agents.

Every task that a frontier agentic model does tens to hundreds of times can potentially be outsourced to a small model. There are ancillary benefits to this, e.g. privacy of a local model reading your files and summarizing to Claude, but almost no one is pushing hard in this direction. The leading model family of capable, customizable small models to date is Qwen, but that’s now shrouded in uncertainty with the departures of key personnel. Gemma, Phi, Olmo, etc. are all major steps down in quality, and therefore potential for modification.

There are a few obvious examples why this can be scaled up. There was a recent thread and discussion on how the new Qwen 3.5 4B model arguably bests the original ChatGPT model. On the research side, there are already recipes for finetuning open models on specific code-bases to match performance of much bigger models. Moondream.ai is a startup made by a friend of mine Vik, who builds some of the best, small multimodal models on a tiny budget — they compete with Qwen and Llama on real world tasks. This is the tip of an iceberg.

Intelligence compression hasn’t been explored with nearly as much depth (or resources) because it is less exciting than keeping track of the progress of the best few models. Investigating these areas is the standard technological diffusion process that is slow and why we’re still early in understanding how people will build with AI. My contention is that too many people building open models are slightly deluded in their perception of their competitiveness. The best few models will win on general capabilities and there are still plenty of underserved niches elsewhere.

Taking this to the next level involves releasing open models that are scoped to be truly excellent at 1-3 tasks, as I hinted at the beginning of this piece. Too many people try to compete with Qwen and show that their small model does great on frontier AI benchmarks. The right benchmark here is savings in compute and time.

It’ll take years for this transition to slowly become reality. Part of why I am so excited about it is that it is driving innovation on open models being more about diversity, specialization, and curiosity, rather than the standard “one model to rule them all” that the frontier models presume.

Models vs. ecosystems.Consolidation vs. creativity.

So long as the open source ecosystem for AI is defined by a bunch of model providers trying to chase after the closed labs, it will largely lose. It will face pain on funding and substantive adoption. The same consolidation that will come for closed AI companies will come for open model builders — likely even sooner.

Open systems at their best allow many people to participate and many approaches to flourish.

The world of open models needs to be more of an ecosystem. I’ve discussed in the past how China is closer to this type of environment by having a variety of companies, but the variety in approaches is still too low.

Ecosystems are self-reinforcing, whereas individual models are static artifacts in time. Ecosystems showcase clear, constant opportunities for what’s next that have growing value propositions.

The path forward for open models is to solve different problems than the frontier labs, to find places where open models are effectively free alternatives, to show ways of using specialized models that the closed labs cannot offer. The world of open models needs to embrace creativity, before building powerful AI systems grows too expensive and prices out many of the prized open labs of today.

Watching history unfold between Anthropic and the Department of War (DoW) it has been obvious to me that this could be a major turning point in perspectives on open models, but one that’ll take years to be obvious. As AI becomes more powerful, existing power structures will grapple with their roles relative to existing companies. Some in open models frame this as “not your weights, not your brain,” but it points to a much bigger problem when governments realize this.

If AI is the most powerful technology, why would any global entity let a single U.S. company (or government) control their relationship to it?

I got Dean W. Ball of the great Hyperdimensional newsletter onto the SAIL Media weekly Substack live to discuss this. In the end, we agree that the recent actions by the DoW — especially the designation of Anthropic as a supply chain risk (which Dean and I both vehemently disagree with) — points to open models being the 5-10 year stable equilibrium for power centers.

The point of this discussion is:

* Why do open models avoid some of the power struggles we’ve seen play out last week?

* How do we bridge short term headwinds for open models towards long-term strength?

* The general balance of capabilities between open and closed models.

Personally, I feel the need to build open models more than ever and am happy to see more constituencies wake up to it. What I don’t know is how to fund and organize that. Commoditizing one’s compliments is a valid strategy, but it starts to break down when AI models cost closer to a trillion dollars than a hundred million. With open models being very hard to monetize, there’s a bumpy road ahead for figuring out who builds these models in face of real business growth elsewhere in the AI stack.

Enjoy and please share any feedback you have on this tricky topic!

Listen on Apple Podcasts, Spotify, and where ever you get your podcasts. For other Interconnects interviews, go here.

Chapters

* 00:00 Intro: is the Anthropic supply chain risk good or bad for open models?

* 04:03 Funding open models and the widening frontier gap

* 12:33 Sovereign AI and global demand for alternatives

* 20:55 Open model ecosystem: Qwen, usability, and short-term outlook

* 28:20 Government power, nationalization risk, and financializing compute

Transcript

00:00:00 Nathan Lambert: Okay. We are live and people will start joining. I’m very happy to catch up with Dean. I think as we were setting this up, the news has been breaking that the official supply chain risk designation was filed. This is not a live reaction to that. If we get any really, really interesting news, we’ll talk about it. I think one of the undercurrents that I’ve felt that this week where everything happened is gonna touch on is open models, but there’s not an obvious angle. I think I will frame this to Dean to start, which is how does-- Like, there’s two sides of open models. One is that there’s the kind of cliche like, not my weights, not your weights, not your mind, where like somebody could take it away if not an open model, which people are boosting like, “Oh, like Anthropic’s gonna take away their intelligence.” But the other side is people worried about open models existing that the Department of War can just take and use for any purpose that it wants. And I feel like both of these are a little cliche. And the core question is like, is this type of event where more control is coming towards AI and more multi-party interest, like is that gonna be good or bad for the open weight model ecosystem?

00:01:12 Dean Ball: My guess is that in the long run, this is probably profoundly good for open weight AI. And like the whole reason I got in, like, so I became interested in frontier AI governance. I did something totally different with my time before. I wrote about different kinds of policy and studied different kinds of policy. And the reason I got into this was because it immediately occurred to me that the government was gonna... I was like, okay, let’s assume we’re building super intelligence soon or whatever, like very advanced AI that seems like really important and powerful. That’s gonna be something that I depend on, like for my day-to-day life. I’m gonna need it for all kinds of things. It’s gonna profoundly implicate my freedom of expression as an American and my exercise of my liberty and all that. And yet it’s also gonna profoundly implicate national security. And so the government’s gonna have its hands all over it, and they also might not like me using it because I might use it, and others might use it to challenge the status quo in various ways, to challenge the existing power structures which the government is a part of. So we have a political problem on our hands here, in my view.

00:02:36 Dean Ball: It immediately occurred to me that we’re gonna have this huge problem of like, this is gonna be a conflict because this is something that’s gonna enormously implicate American speech and liberty, and also it’s gonna have legitimate national security issues, and also the government’s gonna want it because of bad power-seeking reasons. And so that’s always a part of the picture. And my view was this is just a fight that’s gonna play out over the coming decades, and I wanna be a part of this fight. But number two, in that fight, you have to have an insurance policy, and open weight is the insurance policy. Open weight is the way we can always say yes, but we can build the open ecosystem. We can do that. And so I think in the fullness of time, this is gonna be beneficial, but the problem is there’s a lot of coordination and economic problems that have to be solved here. It’s not just a matter of hoping that Google and Meta or whomever else, or the Chinese companies, by virtue, out of the goodness of their hearts continue to open-source things. That’s not scalable. There has to be a reason to do it. So what are the institutional dynamics open weight gonna look like in the long term? I don’t really know, but it feels deeply under theorized.

00:04:03 Nathan Lambert: I think it’s hard to fund is the thing. I mean, we saw Qwen had their turmoil this week, which is timely, and I’m not that surprised because the stakes for these companies is so high, and they all are trying to make sure their companies win in it. And people will say like, “Oh, Meta should commoditize their complements and release open models.” But no one’s ever commoditized their complements with something that costs a trillion dollars to make. Like, that’s a line item. Like, is Apple gonna commoditize... Apple commoditizing their complement would be them doing the... They could spend just as much as all the other tech companies are on CapEx and spend hundreds of billions of dollars, but they’re choosing not to. And I just like, I agree that long term it should be better, but if we never bridge that gap, does it actually materialize? Like, the crank is being turned of these models getting better and better. GPT 5.4 released today, excited to try it.

00:05:02 Nathan Lambert: But like, where does it go? Like, what I’m working on is totally falling behind the frontier. We’re the foundation of research, but it’s like I see it already slipping.

00:05:13 Dean Ball: So I kinda think, yeah, I mean, look, I think it’s gonna get bad in the short term, it’s gonna be bleak, right? There’s just no doubt about that in my view. Because we’re in this period, like I think the pace of frontier progress is gonna continue. My own view is that, like, just ‘cause I peer in and use the open weight Chinese models on a fairly regular basis, and I kinda just feel as though the gap has widened between the US frontier and the open frontier. Unfortunately, it’s so sad that US frontier and open frontier are increasingly distinct things. But I do feel as though that probably is true. And that’s probably gonna continue because in the next, like, in the early stages of a new technology, you would expect for the vertically integrated players to be the ones who do the best. And over time, the modular players can win, and part of that is ‘cause eventually you do get to good enough, right? Like, eventually, I think most people think the iPhone is good enough now. There was a time when every year the iPhone upgrade was like, “Oh my God, this is so much better.” Intelligence is maybe different, but maybe not for a lot of things.

00:06:37 Nathan Lambert: Well, like, there’s no iPhone that you can buy from anyone. Nothing you can buy from anyone but Apple is nearly as good. That’s the concern. It’s like, is it gonna be Anthropic that like, yeah, it stopped getting better, but you can’t rebuild it. Like, you can’t make the open source version.

00:06:51 Nathan Lambert: I also think I had a later question, which is like, the weights are so much less of a concern for me. So like, somebody dropping a two-trillion-parameter model that’s open weights and way better than anything else that somebody has built and released in the open, it almost doesn’t matter if you don’t understand the harness and the tools and the setup you need to make it into a Claude-like system. Like, you need what, eighty nodes of H100s that cost a hundred thousand dollars a day to run and expertise to make it a system. It’s like the shifting away from weights is also happening. I don’t think it’s happening in this open versus closed ecosystem at the surface level of the discussion. So that’s why I’m just like, I don’t know if it’s gonna exist. The thing that I could see happening is that open weights models are niche, and they help these Claude-like models, but there’s not an alternative in that universe. So it’s like, is the government capable of actually making this alternative exist? I don’t know. Like, I don’t know if you can Manhattan Project this, and I wouldn’t advocate for it.

00:07:53 Dean Ball: I actually think about it from the opposite perspective, because I think that what happens if the government follows through on what they’ve threatened with Anthropic, which is to make it so that basically any military contractor cannot have any commercial relations with Anthropic, which means NVIDIA can’t sell GPUs to them for anything. Amazon can’t sell cloud services to them. Amazon and NVIDIA also can’t be invested in them, by the way, if you take any commercial relations at its face value. Now, that’s not a power the government actually has, but nonetheless, if this harassment campaign continues, I think what it probably does... You know, I spend a lot of time in international policy, dealing, talking to foreign governments and civil society in foreign countries, and they already have major trust issues with respect to the US closed source models because they think the US government is gonna come in and disable the models. Like, the American president will get mad at Brazil, say, and in addition to putting tariffs or sanctions, the US president will say, “Yeah, we’re also gonna turn off all your public services that are dependent upon American closed source models.” Right? So people view that as this profound threat, and people are legitimately scared of that in other countries.

00:10:00 Dean Ball: I think this turns that fear up another meaningful degree, and probably not incorrectly, by the way, probably rightfully so. And so I kinda look at this and I think, well, now a lot of American companies might also have that concern, and so you certainly have a demand side of people who are gonna be like, “I get this. It is a risk to use anything where I have a commercial relationship. ‘Cause once I have a commercial relationship, the government can regulate that. Can I find some way of getting out of it?” I think there’s gonna be demand for that. Whether or not that demand produces supply, I think will depend on... It might just not be possible, that’s true. But I think you’ve never had a more favorable demand picture, and I suspect that on the margin, this probably will favor open in the longer run.

00:10:44 Nathan Lambert: Yeah. So there’s a few ways that I think about this. I have this thing, like ATOM Project and all this other stuff I do, and it’s like, how do I meaningfully advocate for this? I think there’s something, like I work at AI2, and AI2 has budgets of order of a hundred million dollars and can train decent models. But if I wanted to redo an AI2, like my method for getting that type of money, it’s mostly gonna be like befriending a billionaire. And it seems like philanthropy dice roll in the near term is a way to get it. But then, like, maybe it really is some long slog of a multi-industrial consortium that takes a couple years off the ground and slowly, like, Google’s, or all these Netflix and all these five hundred billion dollar smaller companies are gonna give millions of dollars to have somebody else do it because they can’t get the billion dollars themselves, but they know they need to have it existed.

00:11:31 Dean Ball: And sovereign wealth funds. Right. Sovereign wealth funds everywhere can do that, right? There’s trillions of dollars in sovereign wealth. There’s pension funds, public employee pension funds. A lot of people can chip into this and it’s possible. This is like, Yann LeCun thinks this is the inevitable outcome. He thinks that the future is gonna be that some sort of global consortium gets together and builds this, because no one country is gonna be able to own it, because it’s gonna be too important. I’ve always kinda doubted that, and I’ve always thought that that outcome is probably a bad outcome for the world, honestly.

00:12:06 Nathan Lambert: That’s a bad outcome for how good the AI is.

00:12:09 Dean Ball: That’s correct. It’s a socialist outcome, you know? It’s not communism, but it is democratic socialism, and I’m not a democratic socialist, so I’m not a super big fan of that. But at the same time, I have to be honest that I kinda think that this probably does increase the odds of that precise outcome coming to bear.

00:12:33 Nathan Lambert: I think something that comes sooner is that a lot of these super wealthy countries are gonna realize they can have real... Like, they can do some sort of sovereign AI and make some sort of noise, particularly starting with open models. I think there’s the Institute for Foundation Models, which is based on the UAE university system. Like, that’s--

00:12:53 Dean Ball: That’s very UAE-coded, yeah.

00:12:55 Nathan Lambert: They’ve been playing that for years, and they can keep doing this. Their models are gonna be pretty good, and I think there’s gonna be more people that do this. There’s the SWISS initiative in EU, which is on one hand doing a good job, on the other hand plagued by the most obvious European limitations of talent cycling and consortium life. I think these things are gonna become more of a thing in the next year, but I don’t know exactly how they impact the... They don’t impact the frontier of AI, but maybe they’re just like how the geopolitics and power of AI evolves. And I for some reason feel like open models need to be the thing that they’re gonna do because if they have a closed model that’s not as good, it doesn’t really give them any sort of power. But I don’t have a good enough world view for what that actually does, and if there’s more EU models, if India actually has their act together and trains a solid model. I don’t know what that does, but I feel like it’s probably gonna happen.

00:13:54 Dean Ball: Yeah. I mean, it’s really super interesting ‘cause I think the other thing-- that will be inherently... I mean, it will be a Linux compared to a macOS, you know? It will not be as good of an experience for people. But then it becomes strange. Like, I don’t think macOS is as appealing of a thing if it’s viewed to be owned by the US government, right? And in fact, part of the reason I think that Apple is able to make its case quite credibly to consumers and businesses is they have resisted US government pressure to turn things over before. People might remember about a decade ago, there was this shooter in San Bernardino, California, and the FBI tried to force Apple to release iPhone data, and Apple said, “No, we’re not gonna expose this information.” Now, I think the FBI eventually just hacked it anyway, but that’s a separate issue. It’s a matter of principle here.

00:15:01 Dean Ball: So yeah, I think it’s an interesting question: do we expect for the gap between the open frontier and the American closed frontier to widen in the near future, especially just because of how much compute they’re gonna have?

00:15:30 Nathan Lambert: A hundred percent. And data and talent. Like, a hundred percent. It’s happening.

00:15:34 Dean Ball: Data, talent. And it’s compounding, right? I mean, this has always been my view. And how much, I’m not sure, but I think it could be quite significant because these things are compounding benefits. And so if you expect them to just continue compounding, then all of a sudden it gets pretty bleak pretty quickly, would be my fear.

00:16:00 Nathan Lambert: One of the... I mean, what’s your take on this? Why has it not compounded so much faster? Like, I feel like these three companies are spending, I don’t know, 10X what the Chinese labs are spending, and you only get like a little bit better model. Like, I believed so full-heartedly that Claude and ChatGPT and all these models are much better, and I expect them to become better by increasing margin, but it’s still confusing why they’re not already more ahead.

00:16:29 Dean Ball: I go back and forth on this. Sometimes I think they are that ahead, and it’s just difficult to show up in benchmarks for the obvious reasons that benchmarks get chased. And like, I do feel that with the coding agents and with certain use cases, I do just feel like, wow, the American frontier is just way ahead, profoundly ahead of the Chinese frontier there. But there’s a lot of other things where you do kinda saturate how good you can be. I suspect that a very large fraction of AI usage is essentially glorified Google search. Even though I don’t think AI is glorified Google search, I suspect that a lot of what people use it for is that, at the consumer level. And it isn’t obvious to me how much better you can get at things like that. But my guess would be that over the next five years, I would guess the American labs really take off, in part because of compute, data, internal deployments for recursive self-improvement style stuff. And also, it’s amazing how we talk about that as just a normal thing now.

00:18:05 Nathan Lambert: I think there will be a ceiling on it. Like, they’re gonna get a ton of improvement-- The gains are insane. It’s like, personally, at my job, I’ve been a lot of a research manager and just chasing s**t down to get a model out the door. But now I can take on hard engineering tasks because I’m like, “Okay, might as well do this at the same time.” Like, going from zero to a hundred software engineers at anyone’s fingertips is worth a lot in terms of exploration. But the next, like, from a hundred to ten thousand is like, people can mess that up type thing. But that’s a huge gain.

00:18:37 Dean Ball: I kind of agree. I think there’ll be a sigmoid there too. But then the other thing that will happen is, like, what I sort of wonder is will the AI companies, will the current model vendors, will they eventually become more like true infrastructure companies where what they actually do is they have models that design their own chips and models that design their own data centers and models that design their own successors. And so it’s this hugely vertically integrated thing, and what you’re really getting access to is not just the model itself, but you’re getting access to this highly optimized hardware, physical world infrastructure. And again, that’s kind of already the case, but does that become even more the case? And then that’s truly insurmountable for any open player. That’s definitionally insurmountable for an open player, and that becomes scary too. But again, this is why I’ve always felt so good about the position of the US closed source labs. This is why I’ve always been pretty bullish on them and have my concerns about open.

00:20:07 Dean Ball: But to the extent the US government makes it impossible to trust closed source models, you do provide an advantage to open there. You’re giving a shot in the arm. If you like open source, you should hope that the supply chain risk designation against Anthropic is quite broad.

00:20:09 Nathan Lambert: It’s a rough thing to hope for.

00:20:09 Dean Ball: I mean, you shouldn’t actually hope for it, but I just mean, like, if that’s the only thing you care about in the world is open source, then--

00:20:17 Nathan Lambert: I would say that anyone that only cares about open source probably is not thinking through any of these principles. It just gets really bad if you only have-- Like, AI is not gonna be meaningful lift to the economy and nor sustainable if everything is open. Like, if models are truly commoditized, things look kind of rough out there.

00:20:36 Dean Ball: I think a world where models get commoditized is a really bleak world too, actually. And yeah, this is why I’m very worried about what the US government is doing. But I think that it helps on the margin, though. It probably helps on the margin in terms of waking people up. That still is my view.

00:20:55 Nathan Lambert: I am a little surprised by the Qwen stuff, but I think there’s-- It’s like, at some point, I knew there was gonna be a year where a lot of the open model efforts just died because they’re just too expensive and too similar. But at the same time, having a lot of efforts that are somewhat similar but exploring a lot of the minor permutations in modeling space to figure out what works for people who use open models is actually quite good. I’m very bearish on the reflection style approach, which is build a lab, build an incredible model, drop it, make a bank selling it on-prem. Because on-prem is not that distinct from a business model as having a closed model. You could sell a closed model on-prem with the right IP controls. But then the person who actually wins open is by trying a whole bunch of tiny different things, understanding what is actually a meaningful differentiator in private data, in certain deployments and whatever, and then really iterating on that with a community. And that’s why I was like, Qwen is the closest to doing this by being so close to the community, and it’s so distinct from what a lot of the other labs are betting on.

00:22:05 Nathan Lambert: But I see the pressure going away and kind of reducing diversity onto standards, because standards also make inference more efficient. Using open models is really rough. I think some of the best open models have really had rough launches. I think GPT-OSS had a horrible launch in terms of usability and is now one of the most popular models of all time. Qwen 3.5, it’s like researchers I work with are like, “Oh, let’s see if we can do some basic RL baselines on it,” and all the software stack is kinda broken. It takes a few weeks to get it going. And this is ‘cause all the models change differently, and closed labs just have such an advantage there ‘cause they should conceivably ship things on day one that work. I mean, don’t talk about Claude’s runtime, but that’s fine.

00:22:42 Dean Ball: And don’t talk about the GPT-5 auto router either. But yeah, no, totally. I think that’s right.

00:22:53 Dean Ball: I think fullness of time, I’m bullish on open source in the long run, fairly bearish in the next five years. The next five years are gonna matter quite a bit. And there is a lot of cope in both open source world and also... I don’t really hear it so much in open source world. I think open source world is actually more honest about this. But where the cope is so bad is in global civil society discourse. Like, I was in India for the AI Impact Summit recently, and they are just smoking the copium, being like, “We are gonna do everything on subfrontier open source models, and we’re just gonna diffuse those, and that’s all we’re gonna need in our economy.” And I just think that’s, if you’re India, that’s really not the bet you wanna make. I understand these are resource-constrained countries. They have a lot of acute constraints that they face, but nonetheless, I think that’s probably not a good bet.

00:24:05 Nathan Lambert: Well, it’s even if those long tail models will work like manufacturing has worked, where it’s like Apple has put hundreds of billions of dollars into the manufacturing ecosystem in China to get absolute fine margins and scale. Like, if you really-- these things are gonna be used so much that that fine margin is actually gonna matter a lot, and it is not cheap to get that fine margin. You can’t just YOLO a DeepSeek V3 and spend five million dollars in compute and be done. It’s still gonna be expensive for a long time.

00:24:34 Dean Ball: Yeah, it requires-- I think the Chinese approach, in the long run, if China’s gonna continue its strategy and they want to be competitive with the American frontier, they’re gonna have to fully socialize that, I think. I don’t think DeepSeek alone is gonna be able to do this, and I don’t think even Alibaba alone is gonna be able to do this. I think they’re going to need some sort of collective effort. Especially because of the export controls, the American export controls. They’re gonna have to centralize compute. They’re gonna have to centralize all these things, and talent and data and all that.

00:25:17 Nathan Lambert: I don’t see it happening. Like, maybe someone gets officially AGI pilled, and I don’t know that much about China. But the things I know about China, it seems like that would be a big lift, and it would take a lot of time to actually do it. Like, all the companies would have to give up their biggest... All the cloud companies are like tech companies making a lot of money. They would be like, “We have to give up what?”

00:25:42 Dean Ball: No, it would be a tough sell. Obviously, if the Chinese government decides they want to do it, they absolutely will. But in total, it will be a tough sell. My experience having had diplomatic engagements of many sorts with Chinese government-- and a lot of Chinese tech policy is actually not directly set by the government. It’s actually more kind of civil society, academia and civil society adjacent to government. Had a lot of conversations with folks like that, and they’re definitely... It’s largely not a very AGI-pilled crew. I think AGI-pilled-ness probably has a rough correlation with GDP per capita, and I think China is about where you would expect based on their GDP per capita, maybe a little bit ahead, but not very so. But if they ever do get AGI pilled, that’s the kind of thing that they could consider, but then that’s still a pretty extraordinary outcome because the Chinese government would have to be willing to make these things and then give it away. And I kinda just don’t think they will.

00:27:11 Nathan Lambert: Yeah. I mean, all the politics of control with how everybody thinks AI is so powerful are pointing to very value-destructive actions economically in order to achieve the end state that people determine to be right. It’s like supporting open source to the extent that you can to avoid situations like Anthropic being labeled a supply chain risk and having interactions like that totally decimating runway of AI productivity. Like, if the companies are really gonna commit to open source for other things, then they’re gonna lose money. And I see this in-- China’s economy would be taking a gigantic hit doing this. And that’s kind of a common theme of what we’re talking about is that the interface of AI in an economic fashion is gonna make the next few years really weird.

00:28:06 Dean Ball: I hope so.

00:28:09 Nathan Lambert: I think things are gonna be weird, but I haven’t spent a ton of time thinking about how that interacts with political institutions. I thought about socially weird a lot, but I haven’t thought about power weird a lot.

00:28:20 Dean Ball: Oh, power weird is what I worry about all the time. What I worry about the most is I think it’s plausible that what we’re seeing... I’ve always had this concern. I have this dual problem of-- maybe I’m talking out of both sides of my mouth. Maybe that’s just the critique, and it’s a fair critique. But I routinely complain about how people in government aren’t really... They pretend to take AI seriously, but they don’t take it that seriously. And they don’t really own the implications of advanced, of near term advanced AI and all that. I think we basically have transformative AI right now, but they don’t own that, because it’s annoying, it’s difficult, it’s conceptually challenging.

00:29:08 Dean Ball: But the flip side of that is that if people do start to take it very seriously, there’s the risk that they sort of lash out, that they get scared, and they lash out and do things that are rash, in a rush. And that actually creates very, very bad, much worse outcomes than you otherwise might have gotten. I think that’s a very fair risk, and I think it’s possible that you might see things like that happen within the U.S. I don’t think this particular incident with Anthropic is quite an example of that. But it’s possible that you do see that in the coming years, and that is in and of itself a pretty scary outcome because if the U.S. government decides that they want to nationalize the frontier labs, I think it could be one of the most tyrannical things we ever see happen in this country.

00:30:16 Nathan Lambert: Yeah. It’s like, I don’t know how to reply to this. I think things are... It’s serious times and I see so many... It feels like such a Sisyphean task to make more open models exist, but all the broader trends seem to point to that being a more stable equilibrium in a lot of ways. Like, good enough open models and keeping up with what we all feel happening in the closed model land.

00:30:50 Nathan Lambert: So I don’t know. I stay motivated, but I feel increasingly lost in terms of achieving it.

00:30:56 Dean Ball: I don’t think you should be. I think, look, I suspect the US government will not actually do it, and the best thing about America is that our general sort of-- I don’t wanna say incompetence, but the general sort of chaos of American institutions and decentralized confusingness of it all, it can often be quite frustrating, and it can sometimes be a detriment, but it can also be really great because we tend to not execute and follow through on our very worst ideas. And so I don’t think we’re going to do that. It doesn’t feel very American to do it. I worry about it because I worry about these rash reactions, and that’s why I fight as heavily as I do on things like this, despite not insignificant cost to me to do it, politically speaking. But that’s totally worth it because I care about this. I think everything, I think that will probably be fine. But yeah, I do agree. It’s a major risk. It’s a major risk, and it’s a weird world to think about, I’ll tell you that much.

00:32:16 Nathan Lambert: Yeah. I don’t have a lot more to add. I’m sure we’ll continue this discussion. I think it warrants the space of it ‘cause that’s the... It’s one of the longer term things, but it’s not in the news cycle whatsoever, at least the open model angle. There’s just so many layers. People have to talk. Like, send feedback, people listening. I’ll even send this out as a podcast as well and just like, what do people think? How do we get to the places we want to get to?

00:32:46 Dean Ball: Well, one thing I’m particularly interested in is-- one of the items in the Trump administration action plan, which I worked on for those who don’t have that context, is this idea of financializing compute, creating a financial market, like basically a commodities market for compute so that you can buy, you know, like really robust. In the same way that you can buy electricity spot, electricity futures and electricity on the spot market and things like this, the wholesale. Could you do something like that for compute? That could really profoundly change the dynamics and the economics of AI production. It’s not gonna turn them over. It doesn’t flip them on their head, but it changes it quite meaningfully. And I’m very excited by that prospect.

00:33:48 Dean Ball: And that’s the kind of thing that I would be increasingly doing if this sort of interference of government into the frontier continues. What I suspect I’ll do is start developing some of those ideas which I developed earlier. I’m only one person. If those things start to seem relevant again, I totally will. Because anything to make it easier to produce AI for people that don’t have trillions of dollars will be extremely important.

00:34:38 Nathan Lambert: Yeah. I think that... I don’t know. I’m happy to leave it there.

00:34:43 Dean Ball: Cool.

00:34:45 Nathan Lambert: I can let you get on your trip. It’s good to catch up. I’m early in the process of potentially coming to DC in a few months, so I will let you know if I do.

00:34:52 Dean Ball: Oh, please do. It’d be great to see you. We can record an episode of my podcast live.

00:34:58 Nathan Lambert: Sounds good. Okay. Thanks everybody for listening.

00:35:03 Dean Ball: Talk to y’all later. Bye.

So-called hybrid architectures are far from new in open-weight models these days. We now have the recent Qwen 3.5 (previewed by Qwen3-Next), Kimi Linear last fall (a smaller release than their flagship Kimi K2 models), Nvidia’s Nemotron 3 Nano (with the bigger models expecting to drop soon), IBM Granite 4, and other less notable models. This is one of those times when a research trend looks like it’s getting adopted everywhere at once (maybe the Muon optimizer too, soon?).

To tell this story, we need to go back a few years to December 2023, when Mamba and Striped Hyena were taking the world by storm — asking the question: Do we need full attention in our models? These early models fizzled out, partially for the same reasons they’re hard today — tricky implementations, open-source tool problems, more headaches in training — but also because the models fell over a bit when scaled up. The hybrid models of the day weren’t quite good enough yet.

These models are called hybrid because they mix these new recurrent neural network (RNN) modules with the traditional attention that made the transformer famous. They all work best with this mix of modules. The RNN layers keep part of the computation compressed in a hidden state to be used for the next token in the prediction — a summary of all information that came before — an idea that has an extremely long historical lineage in deep learning, e.g. back to the LSTM. This setup avoids the quadratic compute cost of attention (i.e. avoiding the incrementally expanding the KV cache per token of the attention operator), and can even assist in solving new problems.

The models listed to start this article use a mix of RNN approaches, some models (Qwen and Kimi) use a newer idea called Gated DeltaNet (GDN) and some still use Mamba layers (Granite and Nemotron). The Olmo Hybrid model we’re releasing today also falls on the GDN side, based on careful experimentation, and theory that GDN is capable of learning features that attention or Mamba layers cannot.

Introducing Olmo Hybrid and its pretraining efficiency

Olmo Hybrid is a 7B base model, with 3 experiment post-trained checkpoints released — starting with an Instruct model, with a reasoning model coming soon. It is the best open artifact for studying hybrid models, as it is almost identical to our Olmo 3 7B model from last fall, just with a change in architecture. With the model, we are releasing a paper with substantial theory on why hybrid models can be better than standard transformers. This is a long paper that I’m still personally working through, but it’s excellent.

You can read the paper here and poke around with the checkpoints here. This is an incredible, long-term research project led by Will Merrill. He did a great job.

To understand the context of why hybrid models can be a strict upgrade on transformers, let me begin with a longer excerpt from the paper’s introduction, emphasis mine:

Past theoretical work has shown that attention and recurrence have complementary strengths (Merrill et al., 2024; Grazzi et al., 2025), so mixing them is a natural way to construct an architecture with the benefits of both primitives. We further derive novel theoretical results showing that hybrid models are even more powerful than the sum of their parts: there are formal problems related to code evaluation that neither transformers nor GDN can express on their own, but which hybrid models can represent theoretically and learn empirically. But this greater expressivity does not immediately imply that hybrid models should be better LMs: thus, we run fully controlled scaling studies comparing hybrid models vs. transformers, showing rigorously that hybrid models’ expressivity translates to better token efficiency, in agreement with our observations from the Olmo Hybrid pretraining run. Finally, we provide a theoretical explanation for why increasing an architecture’s expressive power should improve language model scaling rooted in the multi-task nature of the language modeling objective.

Taken together, our results suggest that hybrid models dominate transformers, both theoretically, in their balance of expressivity and parallelism, and empirically, in terms of benchmark performance and long-context abilities. We believe these findings position hybrid models for wider adoption and call on the research community to pursue further architecture research.

Essentially, we show and argue a few things:

* Hybrid models are more expressive. They can form their outputs to learn more types of functions. An intuition for why this would be good could follow: More expressive models are good with deep learning because we want to make the model class as flexible as possible and let the optimizer do the work rather than constraints on the learner. Sounds a lot like the Bitter Lesson.

* Why does expressive power help with efficiency? This is where things are more nuanced. We argue that more expressive models will have better scaling laws, following the quantization model of neural scaling.

All of this theory work is a great way to go deeper, and frankly I have a lot more to learn on it, but the crucial part is that we transition from theory to clear experiments that back it up. Particularly the scaling laws for designing this model were studied carefully to decide on the final hybrid architecture. The final performance is very sensitive to exactly which RNN block is used and in what quantity.

In scaling experiments, the results showed that for Olmo, the hybrid GDN (3:1 ratio of layers) > pure GDN (all RNN layers) > standard transformer (all attention) > hybrid Mamba2 > pure Mamba2. The crucial point was that these gaps maintained when scaling to more parameters and compute. A visual summary of the different types of architectures studied is below.

In terms of this specific model, the pretraining gains were giant! Relative to Olmo 3 dense, it represents an about 2X gain on training efficiency. When you look at evaluation performance for pretraining, there was also substantial improvement in performance, particularly after long context extension (the final 2 rows of Table 2 in the paper, highlighted below).

The journey to post-training Olmo Hybrid

Most of the experience in post-training Olmo models has been climbing up a steep curve in base model capabilities with minor tweaks to architecture. Our recipes from Tulu 2, Tulu 3, and the Olmo 3 reasoning work (building substantially on OpenThoughts 3) all worked in a fairly straightforward, off the shelf manner. Olmo Hybrid is our first experience in post-training a substantially different architecture, and the results were mixed.

1. Benchmark performance

Following the Olmo 3 recipe, we got some substantial wins (knowledge) and some substantial losses (extended reasoning) relative to the dense model. All together these still represent a very strong fully open model — just that the pretraining gains didn’t translate as obviously. The results are below.

The exact reason why this happens is a research question. Our best guess is that the Olmo Hybrid base model is just a sufficiently different student model, where most of our post training data at early stages is learning from stronger “teacher” models (a recap of this method, called distillation, appeared recently in Interconnects).

There is a lot of other research ongoing in the community around what makes a strong teacher model — generally, the best overall model is not the best teacher. In other words, training on data outputted from the model with best evaluation scores today is unlikely to unlock the ceiling in performance for your new base model. A second factor, which is even less explored, is how different base models likely need different teachers to learn from. This is why Olmo Hybrid could perform very differently, where it’s behavior is downstream of an architecture-based learning change, where the pretraining data is almost identical.

There’s A LOT more work to dig into here, some empirical work in generating better data and other work in understanding how different training stages fit together. I am confident this Olmo Hybrid base model is solid and more performance can be extracted, but it takes more careful work adapting existing datasets.

2. Open-source tooling

The frank reality of new architectures for open models is that the open-source software tooling support is horrific. There’s the paper-cuts that people are familiar with, e.g. random errors in popular libraries (as people experienced with GPT-OSS) that slow adoption, but there are also deeper problems.

A large part of the potential benefit of hybrid models is the reduction in memory usage for long-context generation, which is crucial for reinforcement learning and agentic tasks. It should be a huge win for post-training! This, unfortunately, is far from the case, and will likely take another 3-6months to get right for this batch of GDN models.

The core problem is that the open-source inference tools, e.g. VLLM, are relying on far less developed kernels (and other internals) when compared to standard transformers. This comes with two challenges — throughput slowdowns and numerical issues. Numerical issues can be combatted with a variety of inference flags. Quoting the paper again:

The two key flags in VLLM we needed to get maximum performance with the post-training model were --disable-cascade-attn, which disables cascade attention (an optimization for shared prompt prefixes), and --enforce-eager, which turns off CUDA graphs. These two flags have been used in our RL setup dating back to Olmo 3, but are new additions to evaluations. Scores for the released models drop precipitously without them. We also evaluated our final models with the hybrid model cache in the richer FP32 datatype, to improve stability via --mamba_ssm_cache_dtype following NVIDIA.

Essentially, we used these to make sure the model was numerically stable. The downside is that the inference throughput plummets, so the potential gains in compute efficiency are erased. A comparison of numbers is below.

Effectively, the 7B hybrid model today takes more compute to train with RL than our 7B dense model (that doesn’t even have a common memory saving technique, GQA). The total compute estimate from the table at different context lengths is below (more visuals in the slides from my recent CMU talk).

The good news is that these are solvable problems — and improving the tooling could even improve benchmark numbers — but it’s going to take a good bit of time and hard work in the OSS community.

This leads to my final question. If I’m optimistic about the open ecosystem evolving to support these models with ease, motivated by the better fundamental scaling of the architectures and a large cluster of leading open model builders already using it, are closed models like GPT and Claude built like this?

To be clear, this answer is a total guess (which I don’t normally do), but with the evidence I have I’d put the chance of one of the 3 frontier models being an RNN being around a coin flip. I’ll let you know if I learn for sure either way. If the scaling advantages hold at frontier scale, the economic case becomes hard to ignore, but they could already have architectures that are efficient like RNNs, but with even more benefits.

I’m going to follow up this post with more architecture discussions, particularly on why Mixture of Expert (MoE) models are a major headache to post-train, so make sure to subscribe if that sounds interesting to you!

Thanks to Will Merrill and Finbarr Timbers for some discussions that helped inform this post.

Distillation has been one of the most frequent topics of discussion in the broader US-China and technological diffusion story for AI. Distillation is a term with many definitions — the colloquial one today is using a stronger AI model’s outputs to teach a weaker model. The word itself is derived from a more technical and specific definition of knowledge distillation (Hinton, Vinyals, & Dean 2015), which involves a specific way of learning to match the probability distribution of a teacher model.

The distillation of today is better described generally as synthetic data. You take outputs from a stronger model, usually via an API, and you train your model to predict those. The technical form of knowledge distillation is not actually possible from API models because they don’t expose the right information to the user.

Synthetic data is arguably the single most useful method that an AI researcher today uses to improve the models on a day to day basis. Yes, architecture is crucial, some data still needs exclusively human inputs, and new ideas like reinforcement learning with verifiable rewards at scale can transform the industry, but so much of the day to day life in improving models today is figuring out how to properly capture and scale up synthetic data.

To flesh out the point from the start of this piece, the argument has repeatedly been that the leading Chinese labs are using distillation for their models to steal capabilities from the best American API-based counterparts. The most prominent case to date was surrounding the release of DeepSeek R1 — where OpenAI accused DeepSeek of stealing their reasoning traces by jailbreaking the API (they’re not exposed by default — for context, a reasoning trace is a colloquial word of art referring to the internal reasoning process, such as what open weight reasoning models expose to the user). Fear of distillation is also likely why Gemini quickly flipped from exposing the reasoning traces to users to hiding them. There was even very prominent, early reasoning research that built on Gemini!

This all leads us to today’s news, where Anthropic named and directly accused a series of Chinese labs for elaborate distillation campaigns on their Claude models. This is a complex issue. In this post we unpack a series of questions, beginning with the impact, and ending with politics. The core question is — how much of a performance benefit do Chinese labs get from distilling from American models.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

To start, let’s review what Anthropic shared. From the blog post, emphasis mine:

We have identified industrial-scale campaigns by three AI laboratories—DeepSeek, Moonshot, and MiniMax—to illicitly extract Claude’s capabilities to improve their own models. These labs generated over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, in violation of our terms of service and regional access restrictions.

These labs used a technique called “distillation,” which involves training a less capable model on the outputs of a stronger one. Distillation is a widely used and legitimate training method. For example, frontier AI labs routinely distill their own models to create smaller, cheaper versions for their customers. But distillation can also be used for illicit purposes: competitors can use it to acquire powerful capabilities from other labs in a fraction of the time, and at a fraction of the cost, that it would take to develop them independently.

Much like the models themselves, the benefits of distillation are very jagged. For some capabilities, particularly if you don’t have a full training pipeline setup for it, quickly distilling some data from the leading frontier model in that area can yield massive performance boosts. This can definitely help the lab distilling from the API catch up much more quickly than they otherwise would. Most distillation is rather benign, using many tokens of an LLM to help process and refine existing data — putting a lot of compute into getting a few, high quality training tokens out. This sort of raw data processing work can be done on many different APIs, but one tends to be best.

When we go into what Anthropic says the three Chinese LLM builders actually used the Claude API for — as an aside, Anthropic didn’t confirm that the attack was done through the API, the chat app, or Claude Code — the actual impact of the operations is very mixed. It’s hard to know how much untracked usage these labs deployed for other projects (or other American models).

To start, Anthropic puts DeepSeek first in their blog post because they’re the household name in the US for Chinese AI. The extent of their use is actually quite small, showing how this post is more about the big picture than the details:

DeepSeek

Scale: Over 150,000 exchanges

The operation targeted:

* Reasoning capabilities across diverse tasks

* Rubric-based grading tasks that made Claude function as a reward model for reinforcement learning

* Creating censorship-safe alternatives to policy sensitive queries

In the scale of training a language model, 150K samples is only scratching the surface as a substantive experiment. It looks like they were experimenting with some rubrics, which could’ve been for an online RL run, but that’s extremely unlikely with how distributed the access was, and then some minor stuff on completions for sensitive queries. This usage of Anthropic’s API will have a negligible impact on DeepSeek’s long-rumored V4 model (or whichever model the data here contributed to). This was also very likely a small team at DeepSeek and unknown to much of the broader training organization.

The other two labs, Moonshot AI (makers of the Kimi models) and MiniMax reflected much broader usage.

Moonshot AI

Scale: Over 3.4 million exchanges

The operation targeted:

* Agentic reasoning and tool use

* Coding and data analysis

* Computer-use agent development

* Computer vision

MiniMax

Scale: Over 13 million exchanges

The operation targeted:

* Agentic coding

* Tool use and orchestration

The role of distillation is constantly changing. Distilling from Claude today for its agentic behavior is much more valuable than versions of Claude have been as a teacher in the past. Claude Opus 4.6 has a well-rounded agentic navigation that none of the other models quite match. Why not try training on some of the model outputs to see if your model absorbs it? Over the next few months, that’ll be less differentiated. It’s sort of like how all the models are way better at math today than most people need — there are plenty of places to distill from.

Estimates will vary, but if each response had 10-25K tokens per exchange, the total tokens across these two labs, mostly with MiniMax, would be 150-400 billion tokens. This is a substantial amount, which could meaningfully improve a models’ post-training. For example, in Olmo 3 we had an SFT dataset of 20 billion tokens that could be built like this, and increasing it by 10X would be very reasonable.

These numbers are just scratching the surface of total synthetic data generation across APIs hosted by US companies. At the same time, quantity is a pretty crude way to measure impact. Just taking the outputs from Claude and figuring out how to add them to your model pipeline isn’t easy. The research community has seen many cases where taking outputs from a certain teacher model unexpectedly makes the student worse — subtle interactions between the data make it variable and tricky to do this type of distillation. It’s fundamentally a research problem.

This is what I’m sure the Chinese labs are innovating at. There’s an argument that Chinese frontier labs are substantially more efficient than their Western counterparts — this is misleading.

The labs operate under different constraints. The Chinese labs are likely slightly more efficient out of necessity in being lower on resources, but overall the picture of talent access is very similar. The Chinese labs also approach benchmarks differently, making it appear that they’re a bit closer than they really are (and appearing as if they’re potentially surpassing). This is needed to get momentum and brand recognition in the AI market.

The Chinese labs likely innovate greatly on distilling from leading API models, due to their restricted access to GPUs. GPUs could be used to construct synthetic data, but for organizations with more funding than they can spend on research compute (being supply limited), using API-based models is one of the few other options for effectively getting more compute. It’s way easier to figure out getting access to “banned” API models than it is to smuggle tens of thousands of physical GPUs and get them set up.

It’s not only the Chinese labs that operate like this. Synthetic data from a model you don’t own is all arguably distillation. Distillation is a shortcut to more compute for anyone. It’s also a far less risky cost, as having a big cluster for research requires a very large financial commitment, where APIs are pay-as-you-go. For example, in Olmo 3 we used millions of GPU hours on the Frontier supercomputer and Azure credits through NAIRR for synthetic data. We didn’t have the equivalent in GPUs (or really the cash, thank you research credits!).

All together, it’s very fair for Anthropic to be concerned about this. I still wouldn’t say it is a crucial factor in these Chinese labs post-training capabilities, especially not one that’ll be easy to measure in a time gap to matching the model they’re distilling from a la the US-China performance lag.

If we take a step back, there was even a time when Claude Sonnet was the flagship model ahead of Opus (I think this was with Sonnet 3.5), much of this comes from it being well distilled internally from Opus checkpoints. Fast iteration and high-quality data can go very far, letting student models surpass the teacher. Frontier labs use this to their advantage, by having internal-only models for generating synthetic data, but saying that Chinese models could never pass the US frontier due to data distillation is like saying that Claude Sonnet could never beat Opus. It's unlikely, and it depends a lot on release times, but with AI models making dramatic progress, weirder things like this have already literally happened.

The biggest factor unaddressed here is how distillation from stronger teacher models is harder in an era when reinforcement learning at scale is needed to train the best models. You can spend compute carefully crafting and filtering prompts, but you still need to train the model yourself with substantial, on-policy inference — generation is the majority of the compute cost for RL and it can’t be generations from another model. For this reason, I expected this story to die down a bit. It’s clear from their open research that Chinese labs have excellent RL infrastructure, despite the compute shortages.

The reason I expected it to fade is that not being allowed to distill models for “competitive purposes” has violated the terms of service for API models for quite some time. Academics and open model builders in the US used to greatly worry about and debate this (and I’ve written about it multiple times in 2022 and 2023). Only later in 2024 did that worry die down in the community (and no action has been taken against any smaller model builders).

This action from Anthropic represents another continued step ratcheting up the AI geopolitical tension. Kneecapping model distillation will be far harder than restricting the shipments of physical goods like GPUs. In many ways it seems like fully restricting distillation through distributed access methods seems almost impossible, and restricting GPU sales would be far more impactful.

Anthropic and the AI industry should choose their battles. When API endpoints are available for the best models, other entities will use that to train variants of said model. This is a natural evolution of AI models. If AI models are so precious that distillation is an extreme risk, then the models will be restricted to first-party products. Anthropic has a choice to do this with their latest models. The market for API-based model alternatives may be so competitive that some companies go this path — likely in part due to Chinese models undercutting on price — but an API is a fundamental offering that no leading lab will risk walking back from anytime soon.

Last Thursday, February 5th, both OpenAI and Anthropic unveiled the next iterations of their models designed as coding assistants, GPT-5.3-Codex and Claude Opus 4.6, respectively. Ahead of this, Anthropic had a firm grasp of the mindshare as everyone collectively grappled with the new world of agents, primarily driven by a Claude Code with Opus 4.5-induced step change in performance. This post doesn’t unpack how software is changing forever, Moltbook is showcasing the future, ML research is accelerating, and the many broader implications, but rather how to assess, live with, and prepare for new models. The fine margins between Opus 4.6 and Codex 5.3 will be felt in many model versions this year, with Opus ahead in this matchup on usability.

Going into these releases I’d been using Claude Code extensively as a general computer agent, with some software engineering and a lot of data analysis, automation, etc. I had dabbled with Codex 5.2 (usually on xhigh, maximum thinking effort), but found it not to quite work for me among my broad, horizontal set of tasks.

For the last few days, I’ve been using both of the models much more evenly. I mean this as a great compliment, but Codex 5.3 feels much more Claude-like, where it’s much faster in its feedback and much more capable in a broad suite of tasks from git to data analysis (previous versions of Codex, including up to 5.2, regularly failed basic git operations like creating a fresh branch). Codex 5.3 takes a very important step towards Claude’s territory by having better product-market fit. This is a very important move for OpenAI and between the two models, Codex 5.3 feels far more different than its predecessors.

OpenAI’s latest GPT, with this context, keeps an edge as a better coding model. It’s hard to describe this general statement precisely, and a lot of it is based on reading others’ work, but it seems to be a bit better at finding bugs and fixing things in codebases, such as the minimal algorithmic examples for my RLHF Book. In my experience, this is a minor edge, and the community thinks that this is most apparent in complex situations (i.e. not most vibe-coded apps).

As users become better at supervising these new agents, having the best top-end ability in software understanding and creation could become a meaningful edge for Codex 5.3, but it is not an obvious advantage today. Many of my most trusted friends in the AI space swear by Codex because it can be just this tiny bit better. I haven’t been able to unlock it.

Switching from Opus 4.6 to Codex 5.3 feels like I need to babysit the model in terms of more detailed descriptions when doing somewhat mundane tasks like “clean up this branch and push the PR.” I can trust Claude to understand the context of the fix and generally get it right, where Codex can skip files, put stuff in weird places, etc.

Both of these releases feel like the companies pushing for capabilities and speed of execution in the models, but at the cost of some ease of use. I’ve found both Opus 4.6 and Codex 5.3 ignoring an instruction if I queue up multiple things to do — they’re really best when given well-scoped, clear problems (especially Codex). Claude Code’s harness has a terrible bug that makes subagents brick the terminal, where new messages say you must compact or clear, but compaction fails.

Despite the massive step by Codex, they still have a large gap to close to Claude on the product side. Opus 4.6 is another step in the right direction, where Claude Code feels like a great experience. It’s approachable, it tends to work in the wide range of tasks I throw at it, and this’ll help them gain much broader adoption than Codex. If I’m going to recommend a coding agent to an audience who has limited-to-no software experience, it’s certainly going to be Claude. At a time when agents are just emerging into general use, this is a massive advantage, both in mindshare and feedback in terms of usage data.

In the meantime, there’s no cut-and-dried guideline on which agent you need to use for any use-case, you need to use multiple models all the time and keep up with the skill that is managing agents.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Assessing models in 2026

There have been many hints through 2025 that we were heading toward an AI world where benchmarks associated with model releases no longer convey meaningful signal to users. Back in the time of the GPT-4 or Gemini 2.5 Pro releases, the benchmark deltas could be easily felt within the chatbot form factor of the day — models were more reliable, could do more tasks, etc. This continued through models like OpenAI’s o3. During this phase of AI’s buildout, roughly from 2023 to 2025, we were assembling the core functionality of modern language models: tool-use, extended reasoning, basic scaling, etc. The gains were obvious.

It should be clear with the releases of both Opus 4.6 and Codex 5.3 that benchmark-based release reactions barely matter. For this release, I barely looked at the evaluation scores. I saw that Opus 4.6 had a bit better search scores and Codex 5.3 used far fewer tokens per answer, but neither of these were going to make me sure they were much better models.

Each of the AI laboratories, and the media ecosystems covering them, have been on this transition away from standard evaluations at their own pace. The most telling example is the Gemini 3 Pro release in November of 2025. The collective vibe was Google is back in the lead. Kevin Roose, self-proclaimed “AGI-pilled” NYTimes reporter in SF said:

There's sort of this feeling that Google, which kind of struggled in AI for a couple of years there — they had the launch of Bard and the first versions of Gemini, which had some issues — and I think they were seen as sort of catching up to the state of the art. And now the question is: is this them taking their crown back?

We don’t need to dwell on the depths of Gemini’s current crisis, but they have effectively no impact at the frontier of coding agents, which as an area feels the most likely for dramatic strides in performance — dare I say, even many commonly accepted definitions of AGI that center around the notion of a “remote worker?” The timeline has left them behind 2 months after their coronation, showing Gemini 3 was hailed as a false king.

On the other end of the spectrum is Anthropic. With Anthropic’s release of Claude 4 in May of 2025, I was skeptical of their bet on code — I was distracted by the glitz of OpenAI and Gemini trading blows with announcements like models achieving IMO Gold medals in mathematics or other evaluation breakthroughs.

Anthropic deserves serious credit for the focus of its vision. They were likely not the only AI lab to note the coming role of agents, but they were by far the first to shift their messaging and prioritization towards this. In my post in June of 2025, a month after Claude 4 was released, I was coming around to them being right to deprioritize standard benchmarks:

This is a different path for the industry and will take a different form of messaging than we’re used to. More releases are going to look like Anthropic’s Claude 4, where the benchmark gains are minor and the real world gains are a big step. There are plenty of more implications for policy, evaluation, and transparency that come with this. It is going to take much more nuance to understand if the pace of progress is continuing, especially as critics of AI are going to seize the opportunity of evaluations flatlining to say that AI is no longer working.

This leaves me reflecting on the role of Interconnects’ model reviews in 2026. 2025 was characterized by many dramatic, day-of model release blog posts, with the entry of many new Chinese open model builders, OpenAI’s first open language model since GPT-2, and of course the infinitely hyped GPT-5. These timely release posts still have great value — they center the conversation around the current snapshot of a company vis-a-vis the broader industry, but if models remain similar, they’ll do little to disentangle the complexity in mapping the current frontier of AI.

In order to serve my role as an independent voice tracking the frontier models, I need to keep providing regular updates on how I’m using models, why, and why not. Over time, the industry is going to develop better ways of articulating the differences in agentic models. For the next few months, maybe even years, I expect the pace of progress to be so fast and uneven in agentic capabilities, that consistent testing and clear articulation will be the only way to monitor it.

One of the big stories of 2025 for me was how Nvidia massively stepped up their open model program — more releases, higher quality models, joining a small handful of companies releasing datasets, etc. In this interview, I sat down with one of the 3 VP’s leading the effort of 500+ technical staff, Bryan Catanzaro, to discuss:

* Their very impressive Nemotron 3 Nano model released in Dec. 2025, and the bigger Super and Ultra variants coming soon,

* Why Nvidia’s business clearly benefits from them building open models,

* How the Nemotron team culture was crafted in pursuit of better models,

* Megatron-LM and the current state of open-source training software,

* Career reflections and paths into AI research,

* And other topics.

The biggest takeaway I had from this interview is how Nvidia understands their unique roll as a company that and both build and directly capture the value they get from building open language models, giving them a uniquely sustainable advantage.

Bryan has a beautiful analogy for open models this early in AI’s development, and how they are a process of creating “potential energy” for AI’s future applications.

I hope you enjoy it!

Guest: Bryan Catanzaro, VP Applied Deep Learning Research (ADLR), NVIDIA. X: @ctnzr, LinkedIn, Google Scholar.

Listen on Apple Podcasts, Spotify, YouTube, and where ever you get your podcasts. For other Interconnects interviews, go here.

Nemotron Model Timeline

2019–2022 — Foundational Work

* Megatron-LM (model parallelism framework that has become very popular again recently; alternatives: DeepSpeed, PyTorch FSDP).

* NeMo Framework (NVIDIA’s end-to-end LLM stack: training recipes, data pipelines, evaluation, deployment).

Nov 2023 — Nemotron-3 8B: Enterprise-ready NeMo models. Models: base, chat-sft, chat-rlhf, collection. Blog.

Feb 2024 — Nemotron-4 15B: Multilingual LLM trained to 8T tokens. Paper.

Jun 2024 — Nemotron-4 340B: Major open release detailing their synthetic data pipeline. Paper, blog. Models: Instruct, Reward.

Jul–Sep 2024 — Minitron / Nemotron-Mini: First of their pruned models, pruned from 15B. Minitron-4B (base model), Nemotron-Mini-4B-Instruct. Paper, code.

Oct 2024 — Llama-3.1-Nemotron-70B: Strong post-training on Llama 3.1 70B. Model, collection. Key dataset — HelpSteer2, paper.

Mar–Jun 2025 — Nemotron-H: First hybrid Mamba-Transformer models for inference efficiency. Paper, research page, blog. Models: 8B, 47B, 4B-128K.

May 2025 — Llama-Nemotron: Efficient reasoning models built ontop of Llama (still!). Paper.

Sep 2025 — Nemotron Nano 2: 9B hybrid for reasoning, continuing to improve in performance. 12B base on 20T tokens (FP8 training) pruned to 9B for post-training. Report, V2 collection.

Nov 2025 — Nemotron Nano V2 VL: 12B VLM. Report.

Dec 2025 — Nemotron 3: Nano/Super/Ultra family, hybrid MoE, up to 1M context. Super/Ultra H1 2026. Nano: 25T tokens, 31.6B total / ~3.2B active, releases recipes + code + datasets. Papers: White Paper, Technical Report. Models: Nano-30B-BF16, Base, FP8.

Nemotron’s Recent Datasets

NVIDIA began releasing substantially more data in 2025, including pretraining datasets — making them one of few organizations releasing high-quality pretraining data at scale (which comes with non-negligible legal risk).

Pretraining Data

Collection — CC-v2, CC-v2.1, CC-Code-v1, Code-v2, Specialized-v1, CC-Math-v1. Math paper: arXiv:2508.15096.

Post-Training Data

Core post-training dumps (SFT/RL blends):

* Llama Nemotron Post-Training v1.1 (Apr 2025)

* Nemotron Post-Training v1 (Jul 2025)

* Nemotron Post-Training v2 (Aug 2025)

2025 reasoning/code SFT corpora:

* OpenMathReasoning (Apr 2025)

* OpenCodeReasoning (Apr 2025), OpenCodeReasoning-2 (May 2025)

* AceReason-1.1-SFT (Jun 2025)

* Nemotron-Math-HumanReasoning (Jun 2025), Nemotron-PrismMath (Apr 2025)

NeMo Gym RLVR datasets: Collection

Nemotron v3 post-training (Dec 2025): Collection

HelpSteer (human feedback/preference):

* HelpSteer (Nov 2023)

* HelpSteer2 (Jun 2024)

* HelpSteer3 (Mar 2025)

And others, not linked here.

Chapters

* 00:00:00 Intro & Why NVIDIA Releases Open Models

* 00:05:17 Nemotron’s two jobs: systems R&D + ecosystem support

* 00:15:23 Releasing datasets, not just models

* 00:22:25 Organizing 500+ people with “invitation, not control”

* 0:37:29 Scaling Nemotron & The Evolution of Megatron

* 00:48:26 Career Reflections: From SVMs to DLSS

* 00:54:12 Lessons from the Baidu Silicon Valley AI Lab

* 00:57:25 Building an Applied Research Lab with Jensen Huang

* 01:00:44 Advice for Researchers & Predictions for 2026

Transcript

00:00:06 Nathan Lambert: Okay. Hey, Bryan. I’m very excited to talk about Nemotron. I think low-key, one of the biggest evolving stories in twenty-five of open models, outside the obvious things in China that everybody talks about, that gets a ton of attention. So th- thanks for coming on the pod.

00:00:22 Bryan Catanzaro: Oh, yeah, it’s my honor.

00:00:23 Nathan Lambert: So I wanted to start, and some of these questions are honestly fulfilling my curiosity as a fan. As like, why does NVIDIA, at a basic level, release Nemotron as open models?

00:00:39 Bryan Catanzaro: Well, we know that it’s an opportunity for NVIDIA to grow our market whenever AI grows, and we know that having access to open AI models is really important for a lot of developers and researchers that are trying to push AI forward. you know, we were really excited by efforts from some other companies around the industry to push openly developed AI forward. You know, Meta did some amazing work, obviously, with Llama and you know OpenAI released GPT OSS, which was exciting. And the Allen Institute, of course, has been, you know, really leading the charge for research, open research and, you know, also things like the Marin Project and OpenAthena. You know, like there’s, there’s a bunch of things that we’re always excited to see develop.

And, you know, as we think about where AI is gonna go, you know, NVIDIA believes that AI is a form of infrastructure. it’s.. AI is a very useful technology when it’s applied, but on its own you know, it’s kind of a foundation and infrastructure. We think that technology generally works better when there’s openness to the infrastructure so that people can build things in different ways. You know, you think about the way that the internet transformed every aspect of the world economy is pretty profound, and we’re not done yet.

But the way that, for example, retail uses the internet is different from the way that healthcare uses the internet. And the fact that you know, different sectors of the economy were able to figure out how to incorporate the internet into the beating heart of their businesses in different ways was possible because the internet was built on open technologies that, you know, allowed people to try different things. And we think AI is gonna evolve in a similar way, that organizations across every sector of the world economy are gonna find new and surprising and fun, and important things to do with AI, and they’ll be able to do that better if they have the ability to customize AI and incorporate it directly into the work that they do. and so -- and by the way, this is not to detract from any of the you know, more closed approaches to AI, you know, the APIs that we see from a number of leading labs that, you know, are just extraordinary and have amazing capabilities. We’re excited about those, too.

You know, NVIDIA loves to support AI in all of its manifestations, but we feel like right now the sort of closed approaches to deploying AI are doing pretty well but we, you know, could use some more energy in the openly developed AI ecosystem, and so that’s why we’ve been putting more effort into it this past year.

00:03:42 Nathan Lambert: Yeah. So I’m definitely gonna dig into this a lot ‘cause I have seen this. We’re sitting here recording in January twenty-six, which is in the midst of the rollout of these Nemotron three models. There’s the-- I think the Nano has released in the fall, which was probably one of the biggest splashes the org has made, and everybody’s eagerly awaiting these super and ultra-larger variants.

And it’s like how far are you, how far are you willing to push this Nemotron platform? Like, is it just depending on the users and the uptake and the ecosystem? Like, like, what is the-- is there a North Star in this? Or you hear a lot of.. if you listen to a lot of other open labs, they’re like: “We want to build open AGI,” which is like, I don’t necessarily think grounded, but there’s like a very unifying vision.

Is there something that you try to set the tone for it that goes through the organization? I mean, AI too, it’s like-

00:04:31 Bryan Catanzaro: You know, my North-

00:04:32 Nathan Lambert: .. academics is so-

00:04:34 Bryan Catanzaro: For Nemotron.

00:04:36 Nathan Lambert: Okay, go ahead.

00:04:37 Bryan Catanzaro: Oh, sorry. Go ahead.

00:04:39 Nathan Lambert: I was just, like, gonna compare to, like, AI too, where we can have such a-- like, we have a very specific vision, being so open that it’s like, I think, like, research is so needed, and there’s so little recipes to build on, like, with really credible research. So there’s, like, a research infrastructure, and then when you have something like Llama, it was, like, built on Zuckerberg’s vision, and he changed his mind, which I actually thought his vision was ex- was excellent, the way he articulated the need for open models, and it kind of faded. So it’s like, is there a way to set a vision for an org that, like, permeates every- everyone and is really compelling and exciting?

00:05:17 Bryan Catanzaro: Right. Well, we built Nemotron for two main reasons. The first is because we need to for our main product line. So what I mean by that?

Well, accelerated computing, what NVIDIA does, we build fast computers, right? But the point of building fast computers is to help people do new things. and actually every fast computer is also a slow computer. you know, the observation that it would be nice if computers were faster and could do more things isn’t new. that’s been around since the beginning of computing. So what makes accelerated computing different from standard computing is that we’re prioritizing, you know, we’re focusing, we’re deciding we’re gonna accelerate this workload. This other workload, which is like ninety-nine percent of all of the workloads, we’re gonna let somebody else do that, right?

So, like, you do not buy NVIDIA systems to do any general purpose computation. You buy them for a purpose, right? Which is these days, all about AI. But when you think about the workload, the compute workloads involved in AI there’s a, there’s a lot of diversity and there’s a lot of really important -.. parameters, hyperparameters, or algorithmic approaches that all have enormous imp- impacts on the systems that we need to build for AI.

So things like numeric precision MoE architecture, which of course, influence net-- it influences network design. you know, we’re dreaming about sparsity. We, you know, we’ve had, we’ve had sparse neural network acceleration in the GPU since Ampere. I don’t think that it’s being used enough. you know, so how do we, how do we figure out how to use that? These, these sorts of things have an enormous impact on the future of NVIDIA’s main product line, and we have to understand the answers to those questions deeply ourselves in order to know what we’re going to build.

We can’t just go to our customers and do a survey and say, “Hey “ you know, Meta, for example, since we were just talking about them, “what would you like to see in a future product line from NVIDIA?” Of course, Meta’s always trying to help us as much as they can, but there’s limits to what they can tell us because, you know a lot of the information that influences the design of these systems, it’s very expensive to derive, and so therefore, it’s, it’s very closely held. And so we need to be able to understand these questions very deeply in order to understand what kind of systems to build, in order to understand what we’re accelerating in AI and what we’re not gonna worry about. and so that’s kind of the first job for Nemotron models, is to make it possible for NVIDIA to continue to exist as a company. And I think it’s important that the community knows that because that’s the reason why NVIDIA is making the investments in Nemotron, is because we believe it’s essential for the future of our company. and so this isn’t-- and although as much, as much as it feels good to say, you know, NVIDIA believes in open openly developed AI because you know, we’re so charitable, but actually, that’s not the case. This is actually a business decision-

00:08:34 Nathan Lambert: It’s smart

00:08:34 Bryan Catanzaro: .. like, for NVIDIA, our business needs us to know about AI very deeply. And and so, you know, the amount of investment that is justified to carry on NVIDIA’s ongoing business, I think, is large. and so that’s that’s job number one for Nemotron. Now job number two for Nemotron is to support the ecosystem more broadly outside of NVIDIA. and, you know, NVIDIA has a special position in the AI landscape. of all of the big AI companies I think we’re the one that works with the most other companies. We support every company small and large, AI native company to old established enterprise.

We work with hyperscalers, we work with tiny little startups, we work with countries around the world. so we have this unique position and I think also a uni- unique responsibility and al- maybe also a unique opportunity, that whenever AI is able to grow in any sort of direction, in any capability, then you know, that’s an opportunity for us to grow our business. Obviously, it’s not automatic, right? you know, the AI market is diverse, and it’s getting more diverse, and it should be, ‘cause it’s the most important market in the history of humanity. So so we acknowledge that, and at the same time, we know that it’s in our interest to develop the AI ecosystem. The more people that are building, inventing, and deploying AI, the more opportunity that we have as a company.

So that’s job number two for Nemotron.

00:10:17 Nathan Lambert: Yeah. I really appreciate you saying it so directly ‘cause it’s like we’ve worked.. We- I launched this thing, the Adam Project, last summer, which is trying to get more investment in the US open models, and it’s like the only company that has an obvious business model for open models is something like NVIDIA, where you need to make sure that the open models and the research ecosystem plays nicely on CUDA, because then you’re gonna be able to be one-- You’re so many steps closer to research that’s happening. If not, like, if it like- There’s such an advantage to have research happen mostly on GPUs relative to AMD or anything like this, so.

00:10:49 Bryan Catanzaro: Well, you know, we are-- we’re, we’re not thinking about how to prevent competition. You know, we welcome competition. There’s lots of competition. There should be more competition in this space, but we are very self-interested in staying engaged with the community.

You know, it’s very important. You know, CUDA not many people remember this because it happened so long ago, but you know, CUDA started out with a lot of outreach from NVIDIA to the academic and industrial community saying, “Hey, we have this new way of doing computing. we’d love to see what you can do with it.” In fact, you know, I started using CUDA in 2006 when I was a grad student at Berkeley because David Kirk, who was the chief scientist of NVIDIA at the time, came over to Berkeley and said, “Hey we just released this new GPU, and it has this new programming model called CUDA. You should give it a try.” And I was-- at the time, I was working on machine learning on FPGAs, and I had been working on this one particular piece of support vector machine training on the FPGA, and I decided to take that little piece and write it in CUDA, and it took me like fifteen minutes, and then I ran it, and it was like two hundred times faster than my single-threaded CPU code, and I was like: “Whoa, that was way easier than what I was doing before. I’m just gonna go do that,” right?

So, like, my own personal involvement with CUDA and NVIDIA came about because of this outreach that NVIDIA conducted right from the beginning of CUDA. you know, of course, that led to a lot of great things for NVIDIA, including AlexNet, which was another academic project, you know, where Alex Krizhevsky and Ilya Sutskever were thinking about: “How do we train larger neural networks on more data? we’re gonna go write a bunch of GPU code that uses the GPU in a, in a kinda new and clever way, so that we can train a better image classification model.” And, you know, that had such astonishing results, it kicked off the deep learning era for the whole community. and again, not something that-.. could have been done top-down. That was a, that was a very much a result of NVIDIA supporting open development and re- research in parallel computing and artificial intelligence. And so we remember that, and we’re thinking about in twenty-six, what does it look like to help, you know, the Alex Krizhevsky of the future, who’s, who’s a grad student in a lab somewhere, invent the next technology that changes the world? It seems really difficult to do that without something like Nemotron or, or the other openly developed AI projects out there. yeah, I also wanna say in regards to this Nemotron is not trying to be the only project out there.

We’re part of the community. We love other people doing great work in openly developed AI. We learn from things that other people do and you know, so we’re, we’re trying to support the community because it’s in our interest, but we you know, we’re very happy to see other people contributing as well.

00:13:57 Nathan Lambert: Yeah, I mean, I can transition into something I wanted to ask about is like, I see multiple ways, twenty-five Nemotron mat-- in, I don’t wanna use the word maturing ‘cause I wanna ask you about how it feels in the org, but just like the output reached levels that were more noticed by the community and people building with models. And there’s a lot of ways that can happen, but one of them is like, in my niche community, I’ve been using Nemotron datasets a lot. Like we-- when we redo our post-training recipe, one of the only people we look at is like, okay, NVIDIA, Nemotron has released a lot of high-quality, openly licensed post-training data. this year, you also started releasing some pre-training data, which among AI2 got a lot of notice. Like, what is that? is that like a distinct shift within Nemotron?

Is that something that you’ve wanted to do for a while and finally just did? But it’s ‘cause it’s like-- it is just like a zero to one moment where releasing pre-training data comes with legal risk for any company, but so few people do it, where on my side of the world, it’s like pretty easy to normally say what the best pre-training dataset is, and it had, for a long time, oscillated between like Hugging Face, AI2, DCLM, and there was like literally only two or three options. So in terms of fundamental research, like I think that’s a big step from an org to support the community and take on some risk. So if you have any story you can tell and or just say like, I appreciate it, that’s, that’s all.. that’s all I got.

00:15:23 Bryan Catanzaro: Well, yeah. I mean, so I think it’d be great if more people could understand that Nemotron is not just a model, right? Like, what we’re trying to do with Nemotron is to support openly developed AI, because, again, that’s our big opportunity, right? Now, there’s a lot of organizations that are incentivized to build a model, and the model is maybe the thing that runs their business, right?

But at NVIDIA, the model is not the thing that runs our business, it’s the systems. So when we’re thinking about how do we support the ecosystem, it’s clear to us that the ecosystem needs more than just a model. There’s a lot of models out there already, you know? And of course, we want Nemotron to be awesome, but you know, if Nemotron can convince other people to work on AI because of a dataset or a technique, you know, we’re, we’re trying to be very open with all of the things we learn, you know, including..

I mean, we do a lot of expensive experiments in order to figure out how to do blending for our datasets or to figure out, you know, optimize our settings and, you know, these sorts of things. we’re very happy for other people to pick that up and run with it if it’s useful to them, you know. And so that makes Nemotron a different kind of AI effort. Of course, there is a model component, and that’s a tangible thing, and it’s, it’s easy to focus on that, but we see Nemotron as you know, an effort that includes models, but also includes datasets, techniques, all of all of the research that goes into Nemotron. And again we’re a unique kind of AI organization because of the way that we work with AI companies around the industry and because of the way that our business works, we can afford to be more open with some of these things than maybe some other organizations could be.

Now to your question about, like, does it take some courage in order to be open? Yeah, absolutely it does. and you know, I think there’s been-- one of the things that’s happened in twenty-five is that there’s been an evolving understanding within NVIDIA about the benefits of openness, and that has really enabled the company to make some investments that perhaps it was a little gun-shy to make in the past. And so that’s really encouraging for me. it’s something that I’ve you know, advocated for a while, and so it’s, it’s great to see the company kind of lining up behind it. I also, you know, to your point about like twenty-five being a, a year where Nemotron really made some strides, I want to say thank you for noticing that, and then maybe tell you a little bit about how that happened, because I think it’s instructive for me about how I think the work is gonna go forward in the future.

So you know, NVIDIA is a very decentralized company with a lot of volunteers. You know, everybody that works at NVIDIA is a volunteer. And what do I mean by that? Well, I mean, look, the industry is moving quick.

You know, people can always move from one job to the next. So the way that we think about the work that we do is like, it’s very decentralized, it’s very much let smart people figure out what they should be doing and then kind of self-organize. Now one of the challenges of self-organization in a field that’s moving quickly is that sometimes a whole bunch of people decide to-.. do similar kind of overlapping things but aren’t really coordinated. and that’s okay at the beginning because, you know in a place like NVIDIA, it’s just great to have some energy. It, it took us a while, I think, as a company to figure out that Nemotron was better together.

That rather than having, like, this group has a, has a model and that group has a dataset, and like, you know, then we end up publishing papers that kind of you know don’t really acknowledge each other and aren’t really coordinated. And then, of course along with that, we need to have k times the GPUs, where k is the number of independent efforts. we realized that, you know building AI, you really do need to figure out how to collaborate. the AI efforts that are built from teams of people focused on the overall effort succeeding rather than their own particular piece of the project succeeding, those are the ones that, you know, really change the world. And, you know, of course, NVIDIA works that way for the systems that we build, right? So, like, the people working on the memory controller on the GPU know that they also have to work with the people working on the SM that does the math, right?

Like, you can’t, you can’t make a GPU where it’s just like, “Well, we’ve got an awesome memory controller,” if the math doesn’t work, right? It all has to, has to kinda work together. And so that coordination, I think in the field of AI, it took us a little bit longer to do maybe than you could imagine that it could have. and I think that slowed the progress for Nemotron. so I give a lot of credit to the Nemotron team for realizing over the past, I don’t know, year and a half or so, that it was really time to join up and build one thing and make it awesome, and deeply understand that the success of the Nemotron project was more important than the success of any individual piece of that project. And the reason why I’m telling you all of this is because I think that’s actually true more broadly than just inside NVIDIA, and I think it’s, it’s difficult. you know, researchers like those of us with PhDs, for example, we are taught how to be independent, you know, and how to, how to build up our Google Scholar profile, and there’s, like, an incentive to go ahead and focus on that.

And a lot of successful academics and people researchers you know, they manage to push that pretty far and get some pretty amazing results. But, you know, I do believe that in 2020- in the 2020s you know, that the best research is done as part of a larger team. so how do we figure out how to work together? You know, how do we figure out how to put the success of the team first? That is a thing that is challenging to do but if we can achieve it, I think yield significant results.

And, you know, to the extent that we made progress in that part of the organization, I think we also saw progress in the technology. and that’s.. That gives me great hope for 2026 for Nemotron because the way the team is working together, I think is you know, pretty extraordinary. There’s just an enormous number of brilliant people that have decided that they’re gonna volunteer to make Nemotron awesome, and we’re, we’re starting to see some pretty great things come together.

00:22:25 Nathan Lambert: I agree with everything you said. Do you have any advice for making the orgs come together? I think we’ve seen big-- Wait, I’ve seen two class-- there’s two classes of AI companies right now. One is startup, does everything, and you have a model in six months, but you’re building from zero, and you have-- you p-- everybody agrees when they start that they do this. And then you have Google’s famous long-winded reorgs, which they actually eventually got right. Like, they got it very right with what’s going on with Gemini and Google DeepMind-.. right now. And it’s like, do you have any advice on doing this? I think, like, I’m, AI too, also advocating for this, but it’s very hard. I think personally-

00:22:58 Bryan Catanzaro: It’s-

00:22:58 Nathan Lambert: .. it’s like, I mean, I’m, I’m a special case ‘cause I’m also visible, where it’s e-- very easy for me to turn internet activity into, like, reputation points because of algorithms and size. But it’s very hard to do bottom-up technical work and get all of this and get all the culture alignment. So do you have any advice on actually, like, what works in this domain?

00:23:20 Bryan Catanzaro: You know what’s worked for us is invitation and not control. so you know, one way that, like, for a while I kinda wanted to try to implement was, like, nobody gets to publish any papers in AI unless they’re clearly part of Nemotron. So this is kind of a top-down, like, we’re gonna make you do it, right? I came to the realization that which we never implemented this, by the way, but I came to realization that this was a bad idea because it would just breed resentment, and, you know, NVIDIA is a company of volunteers. Everybody here is a volunteer.

So what we need to do is create the conditions by which it makes sense for people to volunteer to be part of Nemotron. And so the way that we went about doing that first of all it involved like, some top-level agreements between me and some of the other leaders of Nemotron, for example, John Cohen and Kerry Briski. I work very closely with the two of them. And you know, that hadn’t always been the case.

Like, we kind of had all come to this place independently. but we realized, like, Nemotron, better together, all three of us, and then we started telling our teams that: “You know, we really think Nemotron is gonna be better together.” so that top-down alignment, I think was really helpful. We-- again, we weren’t telling people exactly what to do, but we were just sending a con constant message like, you know, “Nemotron’s better together.” And then we built some structures that facilitated collaboration. So in the past decisions in the Nemotron project tended to be made in kind of a an opaque way. and the reason for that is just, you know-.. it’s hard to tell everybody about the middle of the sausage-making process. You know, it’s, like, messy and dif- difficult, and so, like, you know, it’s natural.

Like, researchers, we’re used to doing this, right? It’s a fait accompli. Like, “Here’s my ICML paper,” and like, you know, the fact that you spent, like, two years failing at that task before you finally succeeded, and then you tied a bow around it and gave it to the ICML committee, you don’t really talk about that, right? And so it’s difficult for researchers to, to be open about the middle of the process of research.

There’s a lot of failure, and it’s hard for people to feel like they’re, they’re not looking amazing. But what we, what we decided to do is we structured the project with.. There’s about twenty different areas for the project. Each of them has a clear leader, what we call a pilot in command.

Their job is to-- the job of the pilot in command is to land the airplane. You know, you just want the airplane to land, okay? So somebody, if you’re landing an airplane, there might be multiple pilots on board, but only one of them is gonna land the airplane at any time, right? Because it would be chaos if two of them tried to land at the same time, people would die.

So so this is not a committee structure; it is a delineated responsibility structure. And then the purpose of that pilot in command for each of these sections is to gather together all the best ideas, help the group of people that are interested in working on that space to come up with data-driven answers to what we should do, what technical decisions we should make, and then document that, you know, in a, in a way that other people can review. and you know, the thing that’s been really great about that is that it is inviting to people because when they see, like, okay, here’s the group of volunteers that are working on this area of Nemotron and then they want to contribute, it’s much clearer about how they could go about doing that, and it’s also clearer what the group needs because you know, these meetings are being held in the open. and we have-- we actually have a website where all of the ideas are submitted. they each get, like, a unique identifier, and then they get engaged with, you know, the PIC is trying to understand what the implications are, what kinds of experiments need to be run in order to prove or disprove the idea? how do we do what I call integration studies? You know, I, integration studies are so key for bringing researchers together, and they’re so opposite of what we are taught when we’re learning how to do ablations as a graduate student. You know, rather than, like, isolating the particular contribution of one idea, integration studies are about putting a hundred ideas together and seeing if they’re better than what we had before. so this kind of thing, doing that in a structured way and in a, in an open way internally has then made it possible for more people to volunteer, and that has then generally raised the rigor of the experiments and also the I think the outcome of the work.

00:28:15 Nathan Lambert: Yeah, this is great. I think that over the last few years, there’s been more consensus on things that work for research. And I think the- we also do integration tests very regularly of like, is this feature gonna land for the model? And that’s kind of a..

It’s a good- it’s a nice mirror to ablations, where we know research is changing so much. There’s a lot of turmoil in the academic research community, and it’s nice to have things that are tangible as ways that are a little bit different when you’re doing these large-scale projects. So people that underst- like, you still need to do ablations. But then it needs to survive, like, an additional test in order to land into the model.

So it’s like an additional type of work that needs to be done, and I just like to have words to describe what is actually happening. I think on the Nemotron-3 Nano front, I do a lot of analysis on just looking at basic adoption metrics and Nemotron we created this, what we called like a relative adoption metric, which is essentially looking at downloads over time for models, because it’s easy to know which models have a ton of downloads that are released a while ago. But to, like, look at the trajectory of downloads changing over time, this is a lot-- this is a mouthful. It’s kind of an aside, but, like, Nemotron Nano 3 was in the thirty B size range, like, on track to be one of the top ten models downloaded of all time.

The point that I bring this up, other than to just flatter you, is like, do you think last mile adoption takes a substantial amount of work other than making, like, a very functional model? Or does adoption-- like, do you need to, like, change the recipe that you’re making and put a lot of focus and evaluation and, like, change this over time so that you actually get people to really use the model, rather than, like, “Oh, the benchmarks are good,” look at NVIDIA flying high?

00:30:03 Bryan Catanzaro: Right. Yeah, I mean, wow, it has taken the whole company coming together in order to make Nano V3 have more of an impact than the models that we released before. and there’s so many different aspects to that. obviously, there’s a lot of technical aspects which frankly, I think we have more work to do. So, like you know, making sure that on day zero, when we release something, that the quantizations, all the quantizations, the best quantizations are out there, that the speed on all of the important inference frameworks is out there, that it runs on all of the edge devices that we care about fla- flawlessly, that the install experience is great. You know, this kind of work is extraordinarily important because you know, it’s a crowded world.

There’s so many different things that people could choose to work with, and any amount of friction that gets in the way of people even evaluating something that you do is gonna blunt the results, no matter how good that technology is.. I don’t think that we’re amazing at this yet, so this is something that I anticipate we’re gonna see a lot more investment in as the, you know more people at NVIDIA from all over the company, from marketing, from developer relations, from software engineering, you know as they-- as we all come together in support of this effort. so yeah, so it does, it does take an enormous amount of work. and then, you know, something that I’m particularly interested in is you know, how do we work engage-- i-in a new way, sort of engage with the community to make future Nemotron models even stronger? You know if the only things that we were to optimize for with a Nemotron model would be kind of academic benchmarks that are, you know, highly cited it’s likely the case that the model wouldn’t be general enough to really be useful. And so what we’re trying to build is a technology that other people can extend and deploy, and that means we need to have, like, other ways of understanding the strength of a model besides you know, a handful of academic benchmarks.

I think we have a lot of room to grow here. I’m hoping over time that we develop the muscle of being able to engage with the community and learn from them. Like, you know, okay, this particular thing that I tried to do with Nemotron, it didn’t work. It did this other thing that, you know, I wasn’t expecting, it was wrong. well, that can become feedback that then is used to make the next version better.

I think we’ve got a lot of work to do in that regard.

00:33:10 Nathan Lambert: Do you think there’s any magic to it? I’ve-- I’m blown away by how successful OpenAI’s two open-source models are. Like, yes, they’re obviously the number one name brand in AI, but on the same metric that I see you guys, like, overperforming, like, what I would expect. I’m like, “Wow, great job, NVIDIA.” They’re, like, totally off the charts, like, on track to like, beat Llama’s, like, most downloaded numbers ever with these two GPT OSS models.

And I feel like what they-- like, even on release, they had hiccups where people were pretty negative on it. But for whatever reason, it has just like.. People figured it out, and it just clicked, and then just, like, for a company to say so little about it. Like, we-- Meta put so much effort into Llama being adopted, and you obviously are putting a lot of effort into this.

Like, I’m just like, did OpenAI just crack the code, or is there sometimes a bit of luck?

00:33:59 Bryan Catanzaro: Well, I don’t think I, I don’t think about OpenAI as a, as a lucky company. I think of them as a visionary company that works incredibly hard and you know, I think their success is well deserved. I love the GPT OSS models. You know definitely they’re an inspiration for us here at Nemotron. and yeah, so I think OpenAI also has, like, some other ways of engaging with the community just because of the large number of people that use their services, and that helps them learn things about what are people trying to do with AI, that then they can address when they’re building models, and you know, obviously, you know, people talk about that as a flywheel. you know, I think that’s really interesting and really important.

NVIDIA is never going to have the same kind of flywheel as OpenAI does. We’re not trying to build a service like ChatGPT. What we’re trying to do is help the ecosystem, you know, be strong and enduring. we think that it’s important for there to be this openly developed AI ecosystem, and also we’re, we’re trying to build our next generation of systems, and so we have our own reasons for doing this. But we’re not ever going to have the same exact user base or flywheel that OpenAI does.

On the other hand, you know, we are able to work with institutions around the world in our own way, that I think offers us different opportunities and hopefully, that helps us make things that are, that are useful, too.

00:35:38 Nathan Lambert: Yeah, this makes me realize, I’m having a lot of conversations on.. There are many open model efforts, especially even among people that are fully open, and it’s like, how do we better coordinate? So especially at the smaller scale, it’s like AI2 and Hugging Face. So they’re not big teams.

Like, how do we make sure we’re not doing the same data project at the same-- the same exact thing at the same time? And it’s like, I wonder if there’s opportunities for open companies, like LM Arena has historically released a lot of user data to, like, better help us close this kind of what are people using models for flywheel. And but it’s just-- it’s very hard to build cross-organizational model improvement pipelines, is something that I think. I think models become pretty vertical in terms of somebody at NVIDIA getting the feedback and the model making better.

So that’s what would be something I would like to see this year, but I don’t have ideas for doing it well.

00:36:28 Bryan Catanzaro: Yeah. You know at NVIDIA, we have a tradition of working really closely with, you know, organizations that use our technology. and, you know, we really-- we have, we have teams of engineers that their job is to enable success for our customers. in fact, there’s more people at NVIDIA that care about the success of people outside of NVIDIA than I feel like sometimes there are people that care about the success of things inside NVIDIA. So, like, sometimes I’m like, I’m like: “Hey, could we use a little bit of that e-energy to support Nemotron?” And, and the answer is yes, and NVIDIA is doing that. But I think as Nemotron matures, we’re gonna find that you know, the organizations that work with NVIDIA to make Nemotron awesome for their business, for their use case are gonna have a say in how Nemotron evolves and hopefully, that helps Nemotron address their needs.

00:37:29 Nathan Lambert: .. Yeah, a basic question: how many people, like, how many employees does it take to build all the different versions of Nemotron? I haven’t brought this up because you also have other great types of models. I think our, like, open model analyst, Florian, is obsessed with the Parakeet model, ‘cause- Much faster at typing and is much faster at speaking than typing.

So there’s a lot of other-- I don’t know-- I don’t have the full list of other NVIDIA models off the top of my head, but you are releasing a lot of varieties of models. So I think it’s a bit of a there’s more context to my original question, which is I think about language models ‘cause I’m a n-- like, I just think of AI’s progress is gonna continue to go very fast, so I focus as that as the engine. So but it’s like, how many people is putting this kind of movement into place?

00:38:16 Bryan Catanzaro: Yeah. Well, it’s, it’s, it’s hard to know exactly, and as I said, NVIDIA is a company of volunteers. But and also these days, things are changing, right? Like, so the Parakeet team, which is an excellent team, by the way they I would say a year ago wouldn’t have really considered themselves so much part of the core Nemotron effort, but these days they absolutely are. for the obvious reason that, you know, LLMs these days need to be able to consume all sorts of data, right?

Including audio data. And so you know, as the pro-- as the characteristics, the capabilities of Nemotron models expand obviously, the number of people contributing is gonna expand. I’d say right now there’s about five hundred people that are working pretty much full-time on Nemotron technologies in different ways. This is everything from numerics quantization recipes to speech recognition or image understanding or, you know, pre-training, post-training, RL systems inference software. you know, there’s, there’s a, there’s a whole bunch of different dimensions, right?

So I’d say it’s about five hundred people. but also we’re having our Nemotron all-hands meeting this week, and so I took a look to see how many people were invited to that all-hands meeting, and it was about two thousand. so those are people around the company that are interested in working with Nemotron and either expanding its capabilities or helping its adoption. and so I think you know, the number is somewhere in between and it’s hopefully gonna keep growing as, as Nemotron matures.

00:40:07 Nathan Lambert: Yeah, I mean, that’s one of the greatest attestations to what you’re saying is like, if the interest outside the company-- inside the company is four times as big as the people doing it, you’re gonna, you’re gonna keep scaling up, it seems. People are gonna-.. find ways to help. - One of the other things I’m interested in, I don’t know, like, on the point of five hundred, it’s like, it sounds like a lot of people, but with how many things you have going on, it seems also very few. ‘Cause I’m transitioning to thinking about the long-standing, like, open-source software that you’ve had for NeMo, and I think Megatron, and it’s like they’ve been around for a long time. I think Megatron has gone through many eras. I have a note here.

It’s like these softwares have been going around since, like, twenty nineteen in some form. And it’s, it-

00:40:51 Bryan Catanzaro: Publicly. We had our first public release in twenty nineteen, but we started earlier.

00:40:56 Nathan Lambert: And it’s something that I’ve found is that when I started doing lang- language models, so I was a late bloomer, and we’ll transition to some career talk in a few minutes at Hugging Face. Like Megatron had, like, a bad rap of being very hard to use. But now, like three years later, I hear from anyone that’s founding a new language modeling startup, they’re like, “Just use Megatron.” like, do you pick up on things like this? Is it just, like, random-

00:41:22 Bryan Catanzaro: Well, we-

00:41:22 Nathan Lambert: .. but it’s like-

00:41:22 Bryan Catanzaro: We hard on it. You know, we’re trying really hard to make Megatron easier to use. It’s difficult. Megatron is a complicated piece of technology, and, you know, when we originally started Megatron, the point was to show the community that you could make state-of-the-art large transformer language models with NVIDIA.

I don’t know if you recall, but it-- there was some assertions by some other companies back in twenty seventeen when the transformer was invented, that they could only be made without NVIDIA. in fact, there were statements to that effect on bl-- on official blog posts, which I think got redacted later on. But it was important for NVIDIA to show up and say, “We love language models. We love transformers. Let’s see what we could do, you know, if we partitioned the work properly on lots of GPUs with an amazing interconnect, what kinds of models could we train?” And so that’s where the Megatron project started.

You know, I actually came up with the name Megatron. one of my proudest moments, I suppose. I was thinking about it, I was like: This is a really big transformer. What’s the biggest and baddest transformer? Oh, it’s Megatron.

So that’s, you know, where the name came from. but you’ll think about that had nothing to do with usability, right? Like, I wasn’t, I wasn’t thinking about, like, how do we make a platform that’s really easy for other people to use? I was just trying to show the world that, like, NVIDIA systems could be awesome for transformers. You know, that was, that was my goal.

Over the years, you know, it has evolved. We have a lot more people trying to use Megatron. We got a lot of complaints about how hard it was to use, and then we did a lot of work to try to improve the software engineering around Megatron. You know, these days Megatron software engineering is actually shared between about four different teams at NVIDIA. and we have to coordinate that work very closely.

That has also not been easy. There has been times when you know, people wanted to fork Megatron, and then there were times when we, like, had to bring it back together, and it’s like: Look, I know forking things is always tempting, but look, better together. It’s better for all of us to keep working together.. and so I feel like Megatron the-- and especially Megatron Core, which is like a subset of Megatron that’s, like, especially protected, and we try to put more software engineering into that that has gotten dramatically better since we started paying more attention to it as a company. are we done yet? No, there’s a lot, a lot, a lot more work.

00:43:52 Nathan Lambert: a ba-- a basic question: Is is Megatron or Megatron Core, like, this is what Nemotron is trained on? And also-- And it’s also something that many of the hottest, like, AI startups are training their models on. I would guess that there’s nothing else that does that. So, like, could you summarize why it’s so hard?

00:44:11 Bryan Catanzaro: Well, you know, there’s a, there’s a lot of other great frameworks out there. Megatron’s not the only one. and you know, we’re happy about that. NVIDIA doesn’t need to control the space. What we, what we do wanna do is make sure that we’re putting our products forward in the best light, you know, and it’s a challenging problem.

We’ve got so many things going on with precision and you know, the networking. Like, those questions, like, the software is so complicated. these days, you know, we’re pre-training our Nemotron-3 Super and Ultra models using FP4 which is a thing that, you know, hasn’t been done publicly anyway and something that, you know, we’re pretty excited about because our GPUs have really awesome FP4 throughput. But obviously, the numerical challenges of, like, trying to train a state-of-the-art language model using four bits is non-trivial. So, like, you know, all of that work has to go into Megatron, into Transformer Engine which is a, another open-source project that Megatron relies on and, you know coordinating all of that making sure that, you know, we can actually deliver the benefits of NVIDIA systems to people that are trying to make state-of-the-art models, that’s really important to us.

And, you know, of the five hundred or so people working on Megatron, like, a pretty good fraction.. or on Nemotron, a pretty good fraction of them are working on these kinds of systems issues, right? Because NVIDIA at its core, is a systems company. and Megatron, you know, Nemotron’s first job really is about systems, you know, and so we, we care, we care deeply about that.

00:45:51 Nathan Lambert: Yeah. I mean, from my perspective, I was at Hugging Face before AI2, and Hugging Face is, like, the best company at doing public work. But also, and switching to AI2 and focusing on, like, we’re focused on the output artifact the most. Seeing the different type-- Like, it’s such a different type of work, going from you’re trying to build a tool that’s good for training models, to build a tool that’s good for everybody else and whatever heck use case they are.

00:46:13 Bryan Catanzaro: It’s different.

00:46:13 Nathan Lambert: So I think-

00:46:13 Bryan Catanzaro: Yeah. Different work.

00:46:14 Nathan Lambert: To do both is like.. I’m, I’m happy that AI2’s repos aren’t that popular in terms-

00:46:21 Bryan Catanzaro: Oh,

00:46:21 Nathan Lambert: .. of open-source adoption because, like, we can’t handle it. We just can’t. It’s, like, so hard because it’s people-- it’s, like, it ends up being researchers that are supporting it, and we don’t have the ability to scale the organization structure. So I just think, like, that’s a, that’s a very fun turnaround for me to think of all these things happening at once.

00:46:39 Bryan Catanzaro: Yeah. Well, thanks for noticing we’re putting effort in. I would say Megatron is still not nearly as user-friendly as Hugging Face libraries. Like-.. Hugging Face libraries are legendary, and I admire the work they’ve done to make the community so productive. people, you know, are able to get so much research done thanks to the work that, you know, Hugging Face has put into to their library. So you know, my hat’s off to them as well.

00:47:06 Nathan Lambert: Yeah. One of my hot takes, you don’t have to reply, is that Hugging Face and NVIDIA have been very good partners.

00:47:10 Bryan Catanzaro: Oh, absolutely.

00:47:10 Nathan Lambert: And it’s like bringing that Hugging Face culture to the NVIDIA stuff would be so good. It’s just so hard, so I don’t know how that would work, but-

00:47:17 Bryan Catanzaro: We’re trying, you know, and you know, it is, it is challenging. NVIDIA is always a company that is gonna prioritize speed like hardware speed, above really anything else, ‘cause that’s, like, who we are. I am always trying to make the case that developer speed is important, too, right? It’s like there’s different ways of thinking about speed. and it is definitely the case that a lot of NVIDIA’s software is so cumbersome to use that you know people can’t get the actual hardware speed as fast as it should be because they just give up.

You know, they just don’t, don’t even figure out how to use that. So I think NVIDIA’s making strides there. I think the, the company is understanding more deeply how important developer experience is, and I hope we continue to push that, so that the benefits of all of the systems technology that NVIDIA works so hard on can be more widely used. but at the same time, you know, there is gonna be a tension between those things. It’s, it’s not gonna go away, and you know, to a certain extent, I think that’s just life on planet Earth.

00:48:26 Nathan Lambert: It is. I think you’re do- you’re doing a good job, and I’m gonna kind of shift gears in this interview. So I’ve.. In becoming more back in language- in becoming a person that works in language models, I’ve seen your name more and more times.

I was like, “Bryan Catanzaro, like, where have I seen this?” And then I went and did the research of the Berkeley PhD in, like.. It says April of 2021, you gave a Berkeley EECS Colloquium titled “Applications of Deep Learning and Graphics, Conversational AI, and Systems Design.” I’m not even gonna posit that I actually went, but that’s definitely where I remembered the name from in grad school. And we both have backgrounds that aren’t traditionally in AI and end up working in language models. I just wanted to, like-- what have you learned from your path th- through NVIDIA into what, like, people should be thinking about with AI or open models today?

This could be career reflections, like technical reflections. I just think that there’s-- there are actually a lot of people that come from all over the, like, STEM field to work in AI, so giving it-

00:49:29 Bryan Catanzaro: Sure

00:49:29 Nathan Lambert: .. space to think about is-

00:49:31 Bryan Catanzaro: .. useful, even if it’s just like, it was the big problem, and I wanted to go solve it. Well, I think, you know I’ve, I’ve had a lot of opportunity and a lot of luck in my career. I think in hindsight, it seems like an extraordinarily lucky thing that, you know, I did my first internship at NVIDIA in 2008, and I was, like, building machine learning models on the GPU, and I went to NVIDIA, and nobody else was really doing that. And I was like, “Hey, like, we should have more people doing machine learning on the GPU.

I think this could be an opportunity.” And you know, it took a few years for me to make any headway. NVIDIA didn’t really wanna listen to me. I was a brand-new PhD. I was in the research organization, which is very independent, but, you know, sometimes struggles to change the way that the, you know, the bigger company thinks about things.

And and yet, I just had this conviction, you know, I just was following my heart about what I think is gonna be important, what do I think could really change the world? And that has been, I think, the thread that has taken me through my whole career, is that I’m constantly trying to refine my beliefs about what matters and then hold to them. And that.. I don’t know how helpful it is to say that, but I feel like sometimes people you know, tend to follow the, whatever the thing is that people are talking about on Twitter.

And like I’ve- I’ve done a lot of unpopular things during my career because I believed in them, you know? I remember I published my first paper in 2008 on, at ICML, on training support vector machines on the GPU, and I actually had somebody at the conference, it was in Helsinki at dinner, you know, we were all telling each other what we’re doing, and, and I was like: Yeah, I wanna help people train bigger models on bigger data sets with GPUs. And, and I had you know, a couple of people just say, “Well, why are you here at ICML? That just doesn’t really feel like a good thing for us.” And in 2008, ICML was momly- mainly about new mathematical frameworks for thinking about data, and you know, maybe if you trained a model at all, you would train one on your laptop.

You know, that was the state of machine learning in 2008. So for somebody to come in and say, “I think I want to focus on, like, parallel computing, new kinds of hardware for machine learning, programming frameworks for machine learning, so that, you know, we- more people can try inventing new models on complicated machines with a lot more compute throughput on bigger data sets,” that was like a, an unpopular thing. At least it felt very unpopular. I felt very marginalized at the time by the community.

But I believed in it, you know? I just felt like, look, technology.. Like I have this sense of, like, where do I think technology is going? I knew that traditional computing was running out of steam.

You know, I had, I had done a few internships at Intel, and I was trying to help Intel make processors that ran at, like, ten gigahertz back in 2001, and, you know, it was, like, clear that th- they were running into a wall. And I was thinking: Okay, so if the compute hardware is gonna have to be different, it’s gonna be more restricted. It’s not gonna be able to be so general-purpose in order to get speed. What kinds of applications are gonna have, like, an infinite need for more computing?

And I thought, well, machine learning and AI, that could really change the world if it ever actually worked. But, you know, but, you know, back then it, back then, it kinda worked inside of Google. outside of Google, it kind of didn’t work. and so I had kinda these signals, like it was possible, but it was hard. It was a little weird. It was a little niche.

I was a little bit caught in between different fields, like the systems people didn’t think I was systems enough, and the machine learning people didn’t think I was machine learning enough. But, but I believed in what I was doing, and I found a way to keep following that belief. And, you know, ultimately it was very rewarding when all of a sudden NVIDIA decided, “Hey deep learning is changing the world. What do we know about deep learning?” And then it was like: Oh, well, Bryan’s been doing that for several years, and he’s written some libraries that we could turn into a product.

Let’s go do that. And, you know, so that all happened really quickly after many years of nothing happening, you know? And that was really obviously an amazing opportunity for me. you know, an- another thing that was important to me, I left NVIDIA in 2014 to go work at the Silicon Valley AI Lab at Baidu with a group of really talented people, including Andrew Ng and Dario Amodei and Awni Hannun and Adam Coates, and you know, this was a, a really once-in-a-lifetime opportunity, I think for me, to learn some things that would have been hard for me to learn on my own. you know, I felt at the time at NVIDIA that although I had this great opportunity to help NVIDIA become an AI company, and I was doing that, and I was succeeding at that back in 2013 2014, I also felt like I really wanted to learn from a broader community of people applying machine learning and AI to solve really important business problems. And so going to work at Baidu really gave me that chance. and I was there for a couple of years, learned a ton. very grateful to the team there especially to Andrew Ng, who, who encouraged me to, to join with him on that. and then, you know, I ran into limits of what I could do in California, working for a Chinese company.

I was thinking about, you know, what should I do next? And Jensen asked me to come back and build an applied research lab at NVIDIA in 2016. and -.. I wasn’t sure, like, if that was a good idea. I thought NVIDIA’s already grown so much, you know.

The, the years from twenty fourteen to twenty sixteen, NVIDIA actually grew a lot. these days you look back at it, and you’re like: It was still really tiny. But, but back then, I was like: I don’t know, maybe NVIDIA’s already tapped out. I don’t know if you recall, in twenty sixteen, there was already, like, ten different companies making GPU competitors, right? The TPU had already been out for a while and you know, it, it wasn’t clear that NVIDIA was gonna become as large as it, as it has.

But I believed in the opportunity. I believed in the people. you know, one of the things I loved about NVIDIA was that it’s a very stable organization. So Jensen, he’s been running it since he founded it in nineteen ninety-three. my boss, Jonah Alben, who’s an absolutely extraordinary person has been here for you know quite a, quite a long time, almost since the very beginning of NVIDIA. And these people a lot of the leadership at NVIDIA they love the work.

Their heart is in the work. Jensen and Jonah and many other leaders at NVIDIA, they don’t need to be doing this, right? They, they have earned the right to go sit on a beach and drink mai tais all day, but their heart is in the work, and they work incredibly hard. you know, the.. I feel like if there was an Olympics for email, you know Jensen would get the gold medal.

You know, like it’s, it’s unfathomable to me, like, how much information he’s able to process. and it’s a skill that he’s built up over a long time running this company, but it’s also a reflection of his commitment to the work. And I felt like working at a place where we’ve got this very stable organization that loves the work, that really wants to change the world. You know, why does, why does Jensen get up in the morning? Well, it’s-- this is his chance to do something meaningful.

I thought, associating with these people, you know, I could do worse. I could-- I think I could learn from this as well. And so I came to NVIDIA, and back then it was really hard to explain to people why I was trying to build an AI lab inside of NVIDIA. At, at the time, NVIDIA wasn’t doing very much AI, and so I had to kind of develop a vision for that and then explain it to people. that’s ended up being a really good idea for me as well.

You know, the lab, I think, has really helped NVIDIA. you know, Megatron, I think, has really shown the industry, like, how valuable NVIDIA systems can be for language modeling, which is, which is awesome. DLSS, you know I’m continuing to, to push DLSS forward. Very excited about making graphics, you know more efficient with AI. These days, you know, fifteen out of every sixteen pixels a gamer sees are rendered by AI models that, you know, my team developed, and that then makes the GPU ten times more power efficient.

This is a really exciting you know, thing for me to be involved with, something that I’ve, you know, dreamed about for years. So, so that’s the kind of thing that continues to push me forward, is that I have strong beliefs about what I think is possible, where I think technology’s going, and I’m willing to do things that are we- weird and unpopular but, you know, basically following my convictions. I’m very much always thinking about the people I’m working with, the tribe. You know, I think tribes matter enormously. like you know if I..

So, so back when I was a grad student, I was working on programming models for machine learning. I joined the Python tribe. There are other people that were in the Scala tribe, and the people that did their work in the Scala tribe, trying to make programming models for machine learning in, like, two thousand and ten you know, that work, although a lot of it was technically excellent, didn’t matter to the community as much as the people who were in the Python tribe. It ended up.. and, you know, it kind of sucks sometimes that the world is tribal like this, but it’s just the case.

You know, that like the people that you work with, the community that you work with has a big impact on the problems you think about and then the impact that your work has. So I think a lot about the people and the tribes that I’m collaborating with or that I’m part of. and you know, that’s, that’s kind of been the thread that has carried me through my career.

00:59:56 Nathan Lambert: Yeah. Than- thanks for sharing this full arc. I think you’ve said things that I tell people but in different languages, and the first one, the early days, it seems like there can be space in between fields, where people-- two fields will have their way of describing things, but both of them are probably incomplete, and there can be space there, which is a lot of what I was doing transitioning from novel robots to model-based RL, where I, like, didn’t sit and bear in the actual AI lab, but I started doing AI with my, like, total electrical engineering friends. And then the second thing is, like, I’d wholeheartedly recommend this to people, is, like, choose your work based on the people and people that sincerely are in it for-.. the, what they want to do, and a lot of-

01:00:41 Bryan Catanzaro: And follow your beliefs. You know, think about it. What do you believe in? And it’s okay to change your mind, you know, but, like, figure out what is it that you believe in.

Ask yourself every day: Do I still believe in that? If I do, what next? You know. If I don’t, well, what do I believe in?

You know, that’s been really important to me. I think too many people end up kind of just following trends. That’s not usually helpful because the trends are too late. So if you wanna, if you wanna change the world, you need to be ahead of the trends, and you need to know, you know, it-- trends-- I don’t think trends in computing are just fashion.

I think there’s truth that drives those trends. Not always, but often. You know, it’s just-- this is, it’s there’s kind of an inevitable force of gravity. It just can be really hard to par- parse out the noise and figure out what is the truth that is gonna push the industry forward, and how can you push that with it.

You know, if you can join with that, you can accomplish great things.

01:01:36 Nathan Lambert: Yeah, I agree. I think in building language models, it’s like you want to build a model that the community wants in six months. I think if you’re building a model to compete-.. with the models that are already out, you’re not gonna keep up. And I think that it’s like, what is the right thing is building open language models in six months, and like, where do you need to try to steer things is one of the hardest problems that I think about. So I don’t-- if you want to close with any predictions where you see, like, open models, like, if we’re-- if you’re gonna be here at the end of twenty-six, if there’s anything you think will be far more obvious than it is today, or any bets that you want to make, I think it’s kind of a good place to wrap.

01:02:18 Bryan Catanzaro: Well predictions are always hard, and I don’t feel like I’m very good at making predictions. But I am-- I feel like I am good at identifying what I believe in, and what I believe in right now is that compute remains one of the fundamental challenges behind AI. It has been that way for a very long time and I think it continues to be. I think as we find new ways to apply compute to AI, we discover new forms of scaling laws that help AI become more useful and therefore, it becomes more widespread.

So I’m gonna keep thinking about compute. I continue to believe that the fastest-- that, you know, the way to think about AI is not just in terms of absolute intelligence, but rather intelligence per second. You know, there’s some sort of normalization in there that relates to how fast a model can think, how fast a model can be trained or post-trained. You know, that models that kind of incorporate this compute acceleration characteristic, where they’re thinking about intelligence per unit time, those are gonna end up winning because they end up getting trained on more data, they end up getting post-trained with more cycles, they end up with more iterations during thinking when they’re deployed. and you know, of course, if they happen to fit the hardware really well whatever hardware that is then, you know, that can have a pretty non-trivial effect on the intelligence as well.

So that’s something that I really believe in. I really believe in AI as an infrastructure. You know, there’s, there’s different ways of thinking about AI. I think some people believe AI is more like the singularity, like once AGI has been declared, then the whole world is different forever, and all humans have lost their jobs and, you know, there’s a lot of like-- there’s a lot of things about AI that people believe that I personally don’t believe.

You know, I believe, first of all, that intelligence is very multifaceted that it is not easy to pin down, that as soon as we try to pin down intelligence, we find that there’s very many more forms of intelligence that aren’t covered by that. So, for example, a model that achieves gold medal status on the International Math Olympiad, that’s an extraordinary achievement, but it doesn’t make me have no job, right? Like, I’m actually not solving math problems all day, even though, like, having the ability to solve math problems is clearly very useful. And you know, it’s also the case that intelligence is, you know, is kind of like a potential energy it’s not a kinetic energy, right?

In order to transform intelligence into kinetic energy, it needs to have a platform. It needs to be applied in the proper way. and you know, that is why I believe in open models and open- openly developed and deployed intelligence. I believe every company, every organization, has secrets that only they know. They have special data, they have special ways of thinking about their problems, their customers, their solutions, and they’re gonna know how to apply AI better than anyone else.

And so AI as infrastructure that transforms companies, turbocharges them, allows them to take the things they know and multiply their impact, that’s something that I believe in more than AI as an event, that one day, when it happens, makes everyone obsolete. I don’t.. I just don’t believe in that. you know, I often joke that, like if, for example, the CEO were to retire at some point, and we needed to find a replacement you know, handing out an IQ test or asking, you know, who has the highest SAT score that would not be a very good way of finding a replacement, you know? intelligence is just far too complex for that. And so you know, so this, these beliefs, you know, you can disagree with me about anything that I just said, and I’m not offended by that.

I have a lot of friends that do. but you know, I’m asking myself, well, if I believe that intelligence has these characteristics and that AI is gonna change the world by turbocharging institutions that exist a-and also creating new applications that we haven’t even dreamed of yet rather than replacing all humans, then, you know, how do I go about building that, you know? And so that’s, that’s kind of the direction that I’m on right now.

01:07:00 Nathan Lambert: Yeah, I love it. I agree, I agree that we’re entering an interesting area where the open models are taking so many different shapes and sizes and have so many different strengths and trade-offs, that there can start to be interesting interplay as an ecosystem, where there’s just so many different things going on. And I think I like your idea of potential energy, and you have to build things that are kind of unclear of what-- It’s like you have to build the energy in a way, and you don’t really know what the goal is, but you have to do.. try to build these good models. So I appreciate it, and-

01:07:30 Bryan Catanzaro: Yeah, and then let people apply it. Let it-- let them make the kinetic energy happen.

01:07:35 Nathan Lambert: I agree. Thanks for coming on.

01:07:37 Bryan Catanzaro: Thanks so much for inviting me. It’s been a great conversation.

There’s a pervasive, mutual challenge in the job market today for people working in (or wanting to work in) the cutting edge of AI. On the hiring side, it often feels impossible to close, or even get interest from, the candidates you want. On the individual side, it quite often feels like the opportunity cost of your current job is extremely high — even if on paper the actual work and life you’re living is extremely good — due to the crazy compensation figures.

For established tech workers, the hiring process in AI can feel like a bit of a constant fog. For junior employees, it can feel like a bit of a wall.

In my role as a bit of a hybrid research lead, individual contributor, and mentor, I spend a lot of time thinking about how to get the right people for me to work with and the right jobs for my mentees.

The advice here is shaped by the urgency of the current moment in LLMs. These are hiring practices optimized for a timeline of relevance that may need revisiting every 1-2 years as the core technology changes — which may not be best for long-term investment in people, the industry, or yourself. I’ve written separately about the costs of this pace, and don’t intend to carry this on indefinitely.

The most defining feature of hiring in this era is the complexity and pace of progress in language models. This creates two categories. For one, senior employees are much more covetable because they have more context of how to work in and steer complex systems over time. It takes a lot of perspective to understand the right direction for a library when your team can make vastly more progress on incremental features given AI agents. Without vision, the repositories can get locked with too many small additions. With powerful AI tools I expect the impact of senior employees to grow faster than adding junior members to the team could.

This view on the importance of key senior talent has been a recent swing, given my experiences and expectations for current and future AI agents, respectively:

Every engineer needs to learn how to design systems. Every researcher needs to learn how to run a lab. Agents push the humans up the org chart.

On the other side, junior employees have to prove themselves in a different way. The number one defining trait I look for in a junior engineering employee is an almost fanatical obsession with making progress, both in personal understanding and in modeling performance. The only way to learn how the sausage gets made is to do it, and to catch up it takes a lot of hard work in a narrow area to cultivate ownership. With sufficient motivation, a junior employee can scale to impact quickly, but without it, it’s almost replaceable with coding agents (or will be soon). This is very hard work and hard to recruit for. The best advice I have on finding these people is “vibes,” so I am looking for advice on how to find them too!

For one, when I brought Florian Brand on to help follow open models for Interconnects, when I first chatted with him he literally said “since ChatGPT came out I’ve been fully obsessed with LLMs.” You don’t need to reinvent the wheel here — if it’s honest, people notice.

For junior researchers, there’s much more grace, but that’s due to them working in an education institution first and foremost, instead of the understatedly brutal tech economy. A defining feature that creates success here is an obsession with backing up claims. So a new idea improves models, why? So our evaluation scores are higher, what does this look like in our harness? Speed of iteration follows from executing on this practice. Too many early career researchers try to build breadth of impact (e.g. collecting contributions on many projects) before clearly demonstrating, to themselves and their advisors, depth. The best researchers then bring both clarity of results and velocity in trying new ideas.

Working in academia today is therefore likely to be a more nurturing environment for junior talent, but it comes with even greater opportunity costs financially. I’m regularly asked if one should leave a Ph.D. to get an actual job, and my decision criteria is fairly simple. If you’re not looking to become a professor and have an offer to do modeling research at a frontier lab (Gemini, Anthropic, OpenAI is my list) then there’s little reason to stick around and finish your Ph.D.

The little reason that keeps people often ends up being personal pride in doing something hard, which I respect. It’s difficult to square these rather direct pieces of career advice with my other recommendations of choosing jobs based on the people, as you’ll spend a ton of your life with them, more than the content of what you’ll be doing. Choosing jobs based on people is one of the best ways to choose your job based on the so-called “vibes.”

Working in a frontier lab in product as an alternative to doing a Ph.D. is a path to get absorbed in the corporate machine and not stand out, reducing yourself to the standard tech career ladder. Part of what I feel like works so well for me, and other people at Ai2, is having the winning combination of responsibility, public visibility, and execution in your work. There is something special for career progression that comes from working publicly, especially when the industry is so closed, where people often overestimate your technical abilities and output. Maybe this is just the goodwill that comes from open-source contributions paying you back.

If you go to a closed lab, visibility is almost always not possible, so you rely on responsibility and execution. It doesn’t matter if you execute if you’re doing great work on a product or model that no one ever touches. Being in the core group matters.

This then all comes back to finding the people hiring pipeline.

There are many imperfect signals out there, both positive and negative. For individuals building their portfolio, it’s imperative to avoid negative signals because the competition for hiring is so high. A small but clear negative signal is a junior researcher being a middle author on too many papers. Just say no, it helps you.

The positive signals are messier, but still doable. It’s been said that you can tell someone is a genius by reading one Tweet from them, and I agree with this. The written word is still an incredibly effective and underutilized communication form. One excellent blog post can signify real, rare understanding. The opposite holds true for AI slop. One AI slop blog post will kill your application.

The other paths I often advise people who reach out asking how to establish a career in AI are open-source code contributions or open research groups (e.g. EluetherAI). I’ve seen many more success cases on the former, in open-source code. Still, it’s remarkably rare, because A) most people don’t have the hardware to add meaningful code to these popular LLM repositories and B) most people don’t stick with it long enough. Getting to the point of making meaningful contributions historically has been very hard.

Doing open-source AI contributions could be a bit easier in the age of coding agents, as a lot of the limiting factors today are just bandwidth in implementing long todo lists of features, but standing out amid the sea of AI slop PRs and Issues will be hard. That’ll take class, creativity, humanity, and patience. So, to be able to run some tiny models on a $4000 DGX Spark is an investment, but it’s at least somewhat doable to iterate on meaningful code contributions to things like HuggingFace’s ML libraries (I’ve been writing and sharing a lot about how I’m using the DGX Spark to iterate on our codebases at Ai2).

Back to the arc of hiring, the above focused on traits, but the final piece of the puzzle is alignment. The first question to ask is “is this person good?” The second question is, “will this person thrive here?” Every organization has different constraints, but especially in small teams, the second question defines your culture. In a startup, if you grow too fast you definitely lose control of your culture. This isn’t to say that the company won’t have a strong or useful culture, it’s to say you can’t steer it. The culture of an organization is the byproduct of how all the individuals interact. You do not want to roll the dice here.

Interconnects AI is a reader-supported publication. Consider becoming a subscriber.

Personally, I’m working on building out a few more spots in a core post-training methods team at Ai2. Post-training recipes have gotten very complicated, and we’re working on making them easier to run while doing research on fundamentals such as post-training data mixing and scaling laws. To be a little vague, getting the post-training recipes done for both Olmo 3 and Olmo 2 was... very hard on the team. At the same time, post-training hasn’t gotten much more open, so hiring through it and doing the hard work is the only way.

Ideally I would hire one engineer and one researcher, both fairly senior, meaning at least having a Ph.D. or a similar number of years working in technology. Junior engineers with some experience and the aforementioned obsession would definitely work.

This callout serves as a good lesson for hiring. It is intentional that people should self-filter for this, no one likes when you way overreach on selling yourself for a job. I also intentionally make people find my email for this as an exercise. The art of cold emailing and approaching people in the correct pipelines is essential to getting hired. Many people you look up to in AI read their emails, the reason you don’t get a response is because you didn’t format your email correctly. The best cold emails show the recipient that they learned from it or obviously benefitted from getting it. Platitudes and compliments are of course nice to receive, but the best cold emails inspire action.

Two of the most recent people I helped hire at Ai2 I learned of through these side-door job applications (i.e. not found through the pile of careers page applications). I learned of Finbarr through his blogs and online reputation. Tyler sent me an excellent cold email with high-quality blog posts relating to my obvious, current areas of interest and had meaningful open-source LLM contributions. Both have been excellent teammates (and friends), so I’m always happy to say the system works, it’s just intimidating.

All together, I’m very torn on the AI job market. It’s obviously brutal for junior members of our industry, it obviously feels short sighted, it obviously comes with tons of opportunity costs, and so on. At the same time, it’s such a privilege to be able to contribute to such a meaningful, and exciting technology. My grounding for hiring is still going to be a reliance on my instincts and humanity, and not to get too tied down with all the noise. Like most things, it just takes time and effort.

Other posts in my “life thoughts” series include the following. I send these to people when they ask me for career advice generally, as I don’t have time to give great individual responses:

* Apr 05, 2023: Behind the curtain: what it feels like to work in AI right now

* Oct 11, 2023: The AI research job market s**t show (and my experience)

* Oct 30, 2024: Why I build open language models

* May 14, 2025: My path into AI

* Jun 06, 2025: How I Write

* Oct 25, 2025: Burning out

Arcee AI is a the startup I’ve found to be taking the most real approach to monetizing their open models. With a bunch of experience (and revenue) in the past in post-training open models for specific customer domains, they realized they needed to both prove themselves and fill a niche by pretraining larger, higher performance open models built in the U.S.A. They’re a group of people that are most eagerly answering my call to action for The ATOM Project, and I’ve quickly become friends with them.

Today, they’re releasing their flagship model — Trinity Large — as the culmination of this pivot. In anticipation of this release, I sat down with their CEO Mark McQuade, CTO Lucas Atkins, and pretraining lead, Varun Singh, to have a wide ranging conversation on:

* The state (and future) of open vs. closed models,

* The business of selling open models for on-prem deployments,

* The story of Arcee AI & going “all-in” on this training run,

* The ATOM project,

* Building frontier model training teams in 6 months,

* and other great topics. I really loved this one, and think you well too.

The blog post linked above and technical report have many great details on training the model that I’m still digging into. One of the great things Arcee has been doing is releasing “true base models,” which don’t contain any SFT data or learning rate annealing. The Trinity Large model, an MoE with 400B total and 13B active tokens trained to 17 trillion tokens is the first publicly shared training run at this scale on B300 Nvidia Blackwell machines.

As a preview, they shared the scores for the underway reasoning model relative to the who’s-who of today’s open models. It’s a big step for open models built in the U.S. to scale up like this.

I won’t spoil all the details, so you still listen to the podcast, but their section of the blogpost on cost sets the tone well for the podcast, which is a very frank discussion on how and why to build open models:

When we started this run, we had never pretrained anything remotely like this before.

There was no guarantee this would work. Not the modeling, not the data, not the training itself, not the operational part where you wake up, and a job that costs real money is in a bad state, and you have to decide whether to restart or try to rescue it.

All in—compute, salaries, data, storage, ops—we pulled off this entire effort for $20 million. 4 Models got us here in 6 months.

That number is big for us. It’s also small compared to what frontier labs spend just to keep the lights on. We don’t have infinite retries.

Once I post this, I’m going to dive right into trying the model, and I’m curious what you find too.

Listen on Apple Podcasts, Spotify, YouTube, and where ever you get your podcasts. For other Interconnects interviews, go here.

Guests

Lucas Atkins —X,LinkedIn — CTO; leads pretraining/architecture, wrote the Trinity Manifesto.

Mark McQuade — X, LinkedIn — Founder/CEO; previously at Hugging Face (monetization), Roboflow. Focused on shipping enterprise-grade open-weight models + tooling.

Varun Singh — LinkedIn — pretraining lead.

Most of this interview is conducted with Lucas, but Mark and Varun make great additions at the right times.

Links

Core:

* Trinity Large (400B total, 13B active) collection, blog post. Instruct model today, reasoning models soon.

* Trinity Mini, 26B total 3B active (base, including releasing pre-anneal checkpoint)

* Trinity Nano Preview, 6B total 1B active (base)

* Open Source Catalog: https://www.arcee.ai/open-source-catalog

* API Docs and Playground (demo)

* Socials: GitHub, Hugging Face, X, LinkedIn, YouTube

Trinity Models:

* Trinity models page: https://www.arcee.ai/trinity

* The Trinity Manifesto (I recommend you read it): https://www.arcee.ai/blog/the-trinity-manifesto

* Trinity HF collection — (Trinity Mini & Trinity Nano Preview)

Older models:

* AFM-4.5B (and base model) — their first open, pretrained in-house model (blog post).

* Five open-weights models (blog): three production models previously exclusive to their SaaS platform plus two research models, released as they shifted focus to AFM — Arcee-SuperNova-v1, Virtuoso-Large, Caller, GLM-4-32B-Base-32K, Homunculus

Open source tools:

* MergeKit — model merging toolkit (LGPL license return)

* DistillKit — knowledge distillation library

* EvolKit — synthetic data generation via evolutionary methods

Related:

* Datology case study w/ Arcee

Chapters

* 00:00:00 Intro: Arcee AI, Trinity Models & Trinity Large

* 00:08:26 Transitioning a Company to Pre-training

* 00:13:00 Technical Decisions: Muon and MoE

* 00:18:41 Scaling and MoE Training Pain

* 00:23:14 Post-training and RL Strategies

* 00:28:09 Team Structure and Data Scaling

* 00:31:31 The Trinity Manifesto: US Open Weights

* 00:42:31 Specialized Models and Distillation

* 00:47:12 Infrastructure and Hosting 400B

* 00:50:53 Open Source as a Business Moat

* 00:56:31 Predictions: Best Model in 2026

* 01:02:29 Lightning Round & Conclusions

Transcript

Transcript generated with ElevenLabs Scribe v2 and cleaned with Claude Code with Opus 4.5.

00:00:06 Nathan Lambert: I’m here with the Arcee AI team. I personally have become a bit of a fan of Arcee, ‘cause I think what they’re doing in trying to build a company around building open models is a valiant and very reasonable way to do this, ‘cause nobody really has a good business plan for open models, and you just gotta try to figure it out, and you gotta build better models over time. And like open-source software, building in public, I think, is the best way to do this. So this kind of gives you the wheels to get the, um... You get to hit the ground running on whatever you’re doing. And this week, they’re launching their biggest model to date, which I’m very excited to see more kind of large-scale MoE open models. I think we’ve seen, I don’t know, at least ten of these from different providers from China last year, and it’s obviously a thing that’s gonna be international, and a lot of people building models, and the US kind of, for whatever reason, has fewer people building, um, open models here. And I think that wherever people are building models, they can stand on the quality of the work. But whatever. I’ll stop rambling. I’ve got Lucas, Mark, um, Varun on the, on the phone here. I’ve known some of them, and I consider us friends. We’re gonna kind of talk through this model, talk through building open models in the US, so thanks for hopping on the pod.

00:01:16 Mark McQuade: Thanks for having us.

00:01:18 Lucas Atkins: Yeah, yeah. Thanks for having us. Excited.

00:01:20 Varun Singh: Nice to be here.

00:01:20 Nathan Lambert: What- what should people know about this Trinity Large? What’s the actual name of this model? Like, how stoked are you?

00:01:29 Lucas Atkins: So to- yeah.

00:01:29 Nathan Lambert: Like, are you, like, finally made it?

00:01:32 Lucas Atkins: Uh, you know, we’re recording this a little bit before release, so it’s still like, you know, getting everything buttoned up, and inference going at that size is always a challenge, but we’re-- This has been, like, a six-month sprint since we released our first dense model, which is 4.5B, uh, in, in July of last year, 2025. So, um, it’s always been in service of releasing large. I- it’s a 400B, um, thirteen billion active sparse MoE, and, uh, yeah, we’re, we’re super excited. This has just been the entire thing the company’s focused on the last six months, so really nice to have kind of the fruits of that, uh, start to, start to be used by the people that you’re building it for.

00:02:16 Nathan Lambert: Yeah, I would say, like, the realistic question: do you think this is landing in the ballpark of the models in the last six months? Like, that has to be what you shop for, is there’s a high bar- ... of open models out there and, like, on what you’re targeting. Do you feel like these hit these, and somebody that’s familiar, or like MiniMax is, like, two thirty total, something less. I, I don’t know what it is. It’s like ten to twenty B active, probably. Um, you have DeepSeeks in the six hundred range, and then you have Kimi at the one trillion range. So this is still, like, actually on the smaller side of some of the big MoEs- ... that people know, which is, like, freaking crazy, especially you said 13B active. It’s, like- ... very high on the sparsity side. So I don’t actually know how you think about comparing it among those. I was realizing that MiniMax is smaller, doing some data analysis. So I think that it’s like, actually, the comparison might be a little bit too forced, where you just have to make something that is good and figure out if people use it.

00:03:06 Lucas Atkins: Yeah, I mean, if, if from raw compute, we’re, we’re roughly in the middle of MiniMax and then GLM 4.5, as far as, like, size. Right, GLM’s, like, three eighty, I believe, and, and thirty-four active. Um, so it-- you know, we go a little bit higher on the total, but we, we cut the, uh, the active in half. Um, it was definitely tricky when we decided we wanted to do this. Again, it was July when... It, it was July when we released, uh, the dense model, and then we immediately knew we wanted to kind of go, go for a really big one, and the, the tricky thing with that is knowing that it’s gonna take six months. You, you can’t really be tr-- you can’t be building the model to be competitive when you started designing it, because, you know, that, obviously, a lot happens in this industry in six months. So, um, when we threw out pre-training and, and a lot of our targets were the GLM 4.5 base model, um, because 4.6 and 4.7 have been, you know, post-training on top of that. Um, and, like, in performance-wise, it’s well within where we want it to be. Um, it’s gonna be... Technically, we’re calling it Trinity Large Preview because we just have a whole month of extra RL that we want to do. Um- But-

00:04:29 Nathan Lambert: I’ve been, I’ve been there.

00:04:31 Lucas Atkins: Yeah, yeah. But i- you know, we’re, we’re in the, um, you know, mid-eighties on AIME 2025, uh, GPQA Diamonds, uh, seventy-five, um, at least with the checkpoint we’re working with right now. We’re still doing more RL on it, but, um, you know, MMLU Pro, uh, eighty-two. So we’re, we’re, we’re happy. We’re really-- Like, for it being our first big run, like, just getting it trained was, was an extreme accomplishment, but then for it to actually be, like, a, a genuinely useful model is a, a cherry on top.

00:05:03 Nathan Lambert: Yeah, let’s go big picture. Uh, like, let’s recap. We have all of the... We have this full trinity of models. I think that there’s a fun note. Uh, did I put it in this doc? Yeah, on Nano Preview, which was the smallest- ... you’re, like, charming and unstable. The model card’s really funny. Um, ChatGPT, doing deep research on this, I was like, ChatGPT Pro just tagged next to it, “charming and unstable.” And I was like: Is this a hallucination? And then in the model card, you have, like: “This is a chat-tuned model with a delightful personality and charm we think users will love. Uh, we think- ... it’s pushing the boundaries, eight hundred million, um, active parameter, and as such, may be unstable in certain use cases.” This is at the smallest scale- ... which is like, I appreciate saying it as it is, and that’ll come up multiple times in the conversation. And then you have Mini, which is like, um, I think it was, like, 1B active, 6B total type thing. In my-- I, I don’t have it, the numbers right in front of me. I have it somewhere else. Um-

00:05:52 Lucas Atkins: Yeah, Nano was, Nano was the 6B, uh, 1 active.

00:05:55 Nathan Lambert: Oh, yeah, yeah.

00:05:55 Lucas Atkins: And then, and the Mini was twenty-six, 3B active.

00:05:58 Nathan Lambert: Yeah. So, like-

00:06:00 Lucas Atkins: Um, yeah.

00:06:00 Nathan Lambert: -are these based on more of, like, you need to build out your training chops, or are you trying to fill needs that you’ve-... heard from community, and like, I think for context, previously, your first open model was a base and post-trained model, which was Arcee 4.5B, which was a dense model- -which people like. And prior to that, you had, like, a long list of, like, post-training fine tunes that you had released. So before that, it was like a post-training shop, and I think that kind of history is i- important to fill in, ‘cause I think most people-- a lot of people are gonna meet you for the first time listening to this.

00:06:34 Lucas Atkins: Yeah, it, it, um, we chose those sizes for Mini and Nano, uh, specifically Mini, um, the 26B, 3B Active, because we wanted to de-risk, uh, large. Like, th- this has all been in service of getting to a model of, of, you know, the 400B class. So, um, we, you know, learned from doing the original 4.5B, that you might have everything on paper that you need to train a model, but i- inevitably, there’s tremendous, you know, difficulties that come up, and, um, it, it’s-- we, we definitely knew we wanted to make sure that we, you know, solved some of... E- especially when it came to just doing an MoE model performance, uh, you know, like a, like an efficient, fast train of an MoE. So, um, we thought that that was a good ground where we could, you know, it wasn’t crazy expensive, uh, but gave us a lot of data, uh, going into large. And then Nano just came about because we had some extra compute time, and we really want to do more research on, like, smaller models that are very deep. Um, and we hadn’t really seen that in an MoE before, so that one was very much we started training it, and then it, you know, early benchmarks were good, so we said, “Well, we’ll just do the whole dataset.” Um, and, uh, but most of the love for those releases went into, to Mini. So I, I definitely think that long term, uh, from an ROI perspective, the smaller models are going to be where we shine, just because there’s a tremendous amount of, of cost savings a company can get from, from optimizing on a, on a smaller model. Um, but, but we, uh, w- we’re definitely gonna be trying to push the, the large frontier, too.

00:08:26 Nathan Lambert: Yeah. Um, I’d like to kind of double-click on training before going back to the small model that’s useful for companies, ‘cause we’re gonna have-- we’re gonna end up talking for, like, twenty minutes plus about open ecosystem. So I kind of am curious, like, philosophically, how your company feels about, like, sharing scientific details. So if I ask you, like, what are the things you’re technically most excited about in the model, or, like, what are the pain points? Like, uh, like, are you willing to talk about these things? Like, I- Do you feel like it’s kind of orthogonal to the company? Like, I feel like a lot of it is just, like, things that happen. I think your framing of all of this is in service of getting the big model going. And particularly, of, like, you have to be thinking about your model as landing in six months, is probably... Like, for people not training models, it’s hard to think about, ‘cause even I- ... like, I’m thinking about trying to refresh our post-training stack for OLMo 3, and I’m like, the thinking model, the, um, we are pretty SFT heavy right now, and it makes it not very dynamic in terms of the thinking time. But it’s just like, I can’t see people deploying this model, or probably will have a hard time fine-tuning it. And it’s like to think about where tool use models are going in six months, like, seems pretty hard. Um, it’s a very hard task to do, so it takes a lot of gumption to actually set out and do it. So I, I would just appreciate the framing, kind of self-reflecting on what I go through. So if you have anything that you think was, like, particularly hard to actually land the six-month outlook, because you use Muon as an optimizer, or is it Muon? And some of these things. I think the data, it’s well known that Datology is cranking a lot of this, and you probably provide-- I think of it as like you’re kind of driving and working with these partners, and I’m sure you provide a lot of feedback on what’s working and what’s not. So- ... anything you’re willing to share, I think it’s useful.

00:10:08 Lucas Atkins: Uh, I, I think, um, I mean, on the data side, like Datology, I-- at least for these models, that, that partnership has very much been almost an extension of our own research team. Like, we’ve worked very closely with them, and, um, obviously, our model’s doing well, you know, i- is, is, is good for them. So, um, but it, it-- there was definitely, you know, and you know this better than most, like, small-scale ablations, when you throw them at scale, sometimes, you know, uh, the-- i- it doesn’t always turn out how you want. So there was quite a lot of iterating there to at least get the dataset we used for Large. Um, I, I would say that as far as looking out six months and then figuring out how we wanted to... Obviously, the big one was compute. We don’t, um, you know, we, we never raised as, like, a foundation model company, so we’ve ne- we haven’t signed massive commits for, you know, thousands of GPUs before. Um, we didn’t have a, a, a massive cluster that was always active, uh, for a lot of our post-training. So if they came before, um, you know, we had sixty-four, uh, H100s, that was pretty sufficient for that kind of work, but obviously, this necessitated quite a bit more. Um, but the first thing was-

00:11:29 Nathan Lambert: That’s still less than people would guess. Like, you’re releasing models- ... that weren’t like, your models weren’t catching national news, but people in the community knew about them. And, like, uh, i- I think of, like, Moondream when I think about that. Like, vik has- ... such little compute, and he puts it to so use. Like, you, like, see how successful he is? And he tells you that he has, I don’t know, thirty... Like, l- it might be, like, sixty-four GPUs. Like, uh- ... there’s, uh, uh, that’s a whole separate conversation on building- ... actual good ML output on little compute. I, I should ta- I should chat with vik about this, but aside

00:12:03 Lucas Atkins: No, it’s, it is-- I think it was... Yeah, it, it, it was very much a gift going into the pre-training side because-... we were kind of already thinking, All right, how do we do the mu- you know, the most with the, the least amount of compute? But, um, you know, we-- it took us quite a while to get the cluster that we have been training large on, which is twenty-two thousand forty-eight B300s. Um, and once we figured out when we were going to get that, get access to that cluster, everything else kind of became clear as far as, like, timelines for Mini and Nano and, and when we wanted to do that. Uh, obviously, you know, five hundred and twelve H100s was easier to come across, um, for Mini and Nano. So once we figured that out, um, it really became, uh, this game of, okay, how can we find, like, the best research on the topic of, of pre-training, and what is kind of... What are the, the, the papers and publications that are coming out, um, that have enough potential and enough precedence, either because, uh, another lab used them, it comes from a reputable team, uh, the ablations and the, the evaluation setup, like in the paper, was sufficient enough to give us confidence. Uh, and then we basically spent, I don’t know, it was probably about two months just figuring out what we wanted our architecture to be for the MoE, then figuring out, okay, now that that’s what we want to do, how do we implement all of that in the actual training pipeline? Uh, how can we-- you know, at that time, there had been many people who’d done Muon, but, um, for post-training, and, and then other-- some Chinese labs had used it, but there wasn’t, like, a widely available distributed Muon, um, to do it that scale.

00:13:54 Nathan Lambert: What do you think that, like, looks like in decision-making? ‘Cause that seems like a risky decision, if you ask me. I think for one, the ti-

00:14:00 Lucas Atkins: Muon?

00:14:00 Nathan Lambert: ... the timing, the, the, like, timing sharing that you’re saying is good. Like, you said this for two months, and then, like... But, like, even Muon is like, that’s a bet that would even take-- like, somewhere like AI2, that would take some serious evidence to go with it. We would want to ablate it. So like- ... on a single track, it’s like y- you had probably had a process for becoming fairly confident in it then.

00:14:24 Lucas Atkins: It- yes, but it, it was also, like, Kimi had, had just come out, and we knew that that one used Muon, and so we knew that it, at least, if implemented correctly, could deliver a good model. There weren’t outstanding ablations done around like... You know, there wasn’t a Kimi scale model done with Adam, and then compared to Muon and see the difference. But, um, that at least gave us enough confidence that if-

00:14:50 Nathan Lambert: What does Muon give you? Does it give you, like, memory saving, uh, in-

00:14:55 Lucas Atkins: No, it’s actually a little bit more memory. It’s, it’s, it’s mostly-

00:14:58 Varun Singh: It’s, uh-

00:14:58 Lucas Atkins: ... like the loss converges a bit quicker.

00:15:00 Varun Singh: It’s, it’s less memory, actually. It’s, uh, uh, only one momentum buffer instead of Adam’s two, uh, beta buffers, and then it’s also better convergence.

00:15:10 Nathan Lambert: Okay. So it’s, like, mostly designed around convergence, and then I know the math is different, which is where this momentum term changes.

00:15:15 Lucas Atkins: Well, it, it kind of came out... I mean, it had its, its, its big, you know, uh, explosion of popularity in the kind of nanoGPT speedrunning community. So it was kind of all built around converging to a certain, you know, validation loss faster, and, uh, that, that, that was, um... As for why we chose it as opposed to Adam, we’d used Adam for 4.5b, uh, but we also knew that if we wanted to move this fast, that we were going to have to make some pretty big bets, educated. Um, but, but still, we would have to make some, some, some risky decisions, um, beyond just, you know, training in general. So, um, there were a few that Muon we went with, uh, I think was, was one of our bigger bets. Uh, we ended up not doing, like, multi-token prediction or, or, or FP8 because we were throwing so many new things into the run at once, um, that-

00:16:12 Nathan Lambert: Do these apply for-

00:16:12 Lucas Atkins: ... if something were to go wrong-

00:16:13 Nathan Lambert: um, Mini and Nano? Are those also Muon, or are those- ... Adam as well? Okay, so then you- ... you get some de-risk from that. Do you know off the top of your head how many days it take to train each of those? Like, a, a good-

00:16:25 Lucas Atkins: Uh-

00:16:25 Nathan Lambert: ... ballpark for people, before-

00:16:27 Lucas Atkins: Yeah, so-

00:16:28 Nathan Lambert: going into the bigger run.

00:16:29 Lucas Atkins: So, so Mini, uh, so Nano on it was five hundred and twelve H200s, uh, took a little over thirty days. Um, and then Mini was about forty-five days.

00:16:45 Nathan Lambert: Okay. I think another thing- ... off the top of my head is I know that, like, a OLMo 1B dense would take us, like, eleven days on a hundred and twenty-eight H100s for a dense model. So, like, sixteen. So, like, the numbers- ... just go up from there. ‘Cause then it’s like the question is like, I’m guessing i- if those are forty-five days, and then you have-- you up the number of GPUs, it’s gonna be like a similar amount of time, or forty days for the big model, but much more stressful.

00:17:16 Lucas Atkins: Yeah, the big model was... But again, that was- we knew that we, we wanted- we felt confident that we could deliver a competitive and exciting model in January 2026. Like, we knew that it would-- we could... Who knows kind of where the research and what, what class and, and, and, and skill and performance of model is gonna come out in the next three months? Um, so we also knew that we really wanted to land sometime in January, and that’s also why we also took- we went with B300s, even though definitely the largest public train of that size on B300s and, and the, um, you know, a lot of the software was not-- did not have, like, out-of-the-box B300 support. It was the only way we were gonna be able to train a model of this size in-

00:18:06 Nathan Lambert: Did you have to do this? Did you have to implement the... like, help solve version issues or other issues on B300s? ‘Cause I’ve heard that-

00:18:13 Lucas Atkins: W-

00:18:14 Nathan Lambert: ... the rollout has been rough.

00:18:16 Lucas Atkins: We had to add-... a, a bit. There, there were a couple days where the, the data center had to take it offline to implement some bug fixes. It was, it was definitely, like, a very cool experience being on the bleeding edge, but, um, also, like, a little frightening ‘cause you just know, like, “Oh, we’re not getting the most out of these that we possibly could.” So, um, a little bit of both.

00:18:40 Nathan Lambert: Uh, was your final training run stable, or did you have to do interventions through it?

00:18:46 Lucas Atkins: Uh, it was very stable, actually. Uh, it took-- the beginning of it was not. The, the, the first ten days were absolute, um... It, it would start very well and, and looked, you know, uh, the dynamics and the logs, and the graphs looked very similar to Mini and Nano, and then after, uh, around a trillion tokens, it- the- we- you know, you’d get collapsing, experts would start to go crazy. Uh, part of this is just, again, we are very sparse compared to what you, you, you have. So, um, you know, four hundred billion total, um, thirteen billion active, two hundred and fifty six experts. Like, it was, it was-

00:19:26 Nathan Lambert: Did you do a, uh, expert routing loss or some sort of balancing loss?

00:19:30 Lucas Atkins: Yeah. Yeah, yeah. Yeah.

00:19:32 Varun Singh: We did, um, we used DeepSeek’s, uh... We, we modified DeepSeek’s Auxiliary-loss-free, um, uh, loss balancing with our own, like, uh, with some tweaks, and then we also added a sequence loss like they, uh, did as well.

00:19:47 Nathan Lambert: Uh, was there Auxiliary-loss-free one from DeepSeek V3, or was that a later model?

00:19:51 Varun Singh: That was V3.

00:19:52 Lucas Atkins: It was V3.

00:19:52 Varun Singh: They did a separate paper on it as well. Yeah.

00:19:55 Nathan Lambert: Yeah. Yeah, that makes sense. I think a lot of people have derived from there. Um, have you- ... had issues on post-training as well? So I have a theory that the new algorithms we’re getting from the Chinese labs, like GSPO and SysPO, are primarily for problems that you solve when you have big MoEs and you have expert problems when trying to do the RL. And that’s the whole reason that, like, I think our very serious AI two RL setup, like, we’re doing it on dense models, and we’re just like, “It’s fine. We don’t have this big clipping problem, and as much like we don’t have as much of a need to get the batch size as big to ac- activate all the experts.” So you’re saying you have so many experts and so much sparsity, that potentially sounds like you’re making RL harder.

00:20:36 Lucas Atkins: Um, yes. I will also... I will say that from just, like, a purely post-training side, we added as much as we po- we used- we... So our code base started from TorchTitan. We’ve had to make a ton of modifications to it to get it where we need it to be, but that was an excellent base. And from one of the bigger learnings from Mini and Nano was treating, uh, at least the SFT side of it, as a s- as a separate phase. Um, ‘cause with, with Mini and Nano, we finished the pre-training, we did context extension, then we took those and then ran those on, like, the sixty-four H100s we usually would do post-training on. Um, that presented a lot of challenges, uh, with the MoEs. They, they really... And that’s kind of been a thing in the open space, is post-training MoEs, like, really, um, can be frustrating, even for SFT. So for Large, we added, uh, like, fine-tuning directly to TorchTitan, um, and did it all on the same cluster. So, um, from a performance standpoint, like, SFT was very, um... actually ended up being totally different.

00:21:42 Nathan Lambert: What is the actual difference between the q- the, the implementations then? Is it just kinda like you end up with different batch sizes and parallelism and stuff? Like why-

00:21:50 Lucas Atkins: Uh, I mean, we ended up, we... Yeah, we ended up needing to get it to do really, like, to get context parallelism really well, really good, ‘cause we’re obviously going at a higher sequence length, and then, um, just adding the proper loss masking. Um, it, it, it, it ended up being a relatively easy implementation, especially ‘cause we did all the pre-processing, uh, outside of TorchTitan.

00:22:13 Nathan Lambert: Interesting.

00:22:14 Lucas Atkins: Uh, and then on the RL side, yes, I would say it’s not, um, it didn’t present itself as, as, as significantly harder than, than, um, Mini and Nano. However, that many GPUs does, so we didn’t end up using, uh, two thousand of the B300s for that. That ended up being, uh, a thousand. So two, we just split the nodes in half.

00:22:39 Nathan Lambert: Yeah. That makes sense.

00:22:40 Varun Singh: On the dense model side of things, uh, you mentioned that you didn’t need to use all the tricks and stuff. I, I think it is, uh... I think the, the, it- MoEs are just, in general, harder to RL, but I think it’s also, like, uh, b- because of, like, the KL mismatch between trainer and inference engine, right? Um, where you have, like, uh, sometimes the inference engine can pick different experts compared to, like, the trainer, uh, when you, like, do a forward pass on the same tokens. So I think there is definitely some, like, inherent instability with, with RL on MoEs.

00:23:13 Nathan Lambert: Yeah, that makes sense. Are, are... Okay, um, another question of, like, how much do you want to say? How do you feel about the state of public post-training recipes? Like, do you... Like, I, I feel like there’s so little out there, and there’s an opportunity to be seen as technical leaders by sharing just, like, more of what you’re doing. ‘Cause I feel like we’ve seen for years how complicated things can be, but also at, kind of at the same time... Like, we see this from the likes of Llama, has these really complicated recipes. But at the same time, I feel like just executing on a simpler recipe can get pretty close. But it’s just, like, very uns- I feel, uh, currently unsatisfied with how much I know about what are the actual core trade-offs of doing post-training well. And I think you could do a lot with SFT, but there’s definitely, in this RL regime, more trepidation of kind of narrowing your model to either downstream use or, like, being able to do this multi-week RL run where you get the most performance.

00:24:06 Lucas Atkins: Yeah, I mean, I, I, from-- since RL has become such a pivotal part of the process beyond what, you know, DPO and, and, uh, and kind of your, your typical RLHF was in the past, like, we used to get quite, uh-... sophisticated with, with how we would do SFT and, and even our, our RL. We, we obviously, we make MergeKit, so we, we utilized merging, and we used to do a lot of distillation, um, to eke out as much performance as we could. Now that RL is such a massive part of the entire post-training stack, I, I have almost reverted us to just really solid but simple SFT. Um, like in, in large, I mean, we’ve-- our post-training data set for, uh, Trinity Large is, uh, two hundred and thirty billion tokens. Like, like, it just like a really, really, really large-

00:25:09 Nathan Lambert: That’s ten X what we did. At least in SFT.

00:25:10 Lucas Atkins: And even that-- and even, even your tenant, like that was bef- before this kind of w- going at this scale and even kinda thinking and, and reasoning models. Like our largest SFT before that was five billion to-- we’d do, like, three epochs, but it was like five billion, you know, tokens, so- Um-

00:25:28 Nathan Lambert: Our non-reasoning model is, like, te- another ten X. So, like, our most latest instruct model is, like, two billion.

00:25:34 Lucas Atkins: Yeah, which is, uh, already a lot, you know. So, um, I, I’ve definitely... We-- you know, simplicity’s key because it also makes debugging anything easier, and then, um, devoting a lot of that sophistication to the RL. Our RL part is, like, really important. I do think that, I mean, the next, uh, phase of reinforcement learning for models of this scale is, is just scale. Is, is... Okay, we went from, you know, twenty billion SFT to two hundred and thirty, now we’re going from, you know, ten environments to a hundred. I think that that really is where you’re gonna get the biggest benefit. I also think that’s why, you know, MiniMax and, and, and other players like GLM are so performant and just, like, have that extra bit of, of usefulness that goes beyond just what you see in the benchmarks, is they’ve, they’ve really embraced, like, long-form, uh, RL. And, and so, um, yeah, I mean, to be quite frank, our, our RL pipeline’s rather... immature might be the wrong word. Like, it’s, it’s, uh, there’s definitely a lot more work we could do and a lot more work we need to do, but, um-

00:26:43 Nathan Lambert: Have you started the tool use side of RL?

00:26:46 Lucas Atkins: That-

00:26:46 Nathan Lambert: Or are you mostly... Well, um, beyond like, if you’re training on code, just verifying the code answer, I don’t count yet as tool use. I would say, like, search and code integrated reasoning is what I think is gonna be like minimum table stakes, but do it- to do it well is really hard. Like, we have to, like- ... like, you, you really, like, uh... That’s what I want to do. I want all of our models to have that this year. Search is prob- you have to have, like, a partner to do search or just, like, illegally scrape Google if you’re gonna- ... you’re gonna serve this model onto a customer, and it’s gonna- ... what? Go, go to Google, like, what?

00:27:16 Lucas Atkins: Yeah. Yeah, no, I mean, I, I... Beyond, like, like, really kind of like long-form, like deep research or, um, you know, even like GPT-OSS style or, or G- GPT 5 style, where, you know, it’s doing a hundred tool calls before it gives you a response. Not there yet, um, but that is kind of... Once we get past the, the final kind of RL of Trinity Large, and, and we kinda look at where we go next, like, that is the next major hurdle, um, for sure, and it’s intimidating.

00:27:56 Nathan Lambert: How big is your, your team of- of... Like, how many people are spending the majority of their time on the model? And then I think we c- start to wrap up technical talk and zoom out a bit to ecosystem and company strategy.

00:28:09 Lucas Atkins: Uh, there’s thirteen at Arcee- ... that are just, like, every, every single day is working on it. Yeah.

00:28:16 Nathan Lambert: And I guess that’s a good number because these people are talking about data, but there’s also, like, the whole data thing that’s coming somewhere else. But also somebody else that wanted to pre-train a model, like they could just download the best fully open data set. And I don’t think it’s gonna be quite as good, particularly in the fact that, um, like, if you look at OLMo’s models, we don’t have a lot of tokens, so we need to, like, acquire- ... more tokens in the open still. But to, like, get a number of thirteen, where some are spending a bit of time on data, but there’s the whole data abstraction, is actually kind of nice for somebody that’s like... To do a serious modeling effort, you need to have this many people, I think.

00:28:50 Lucas Atkins: It, it was-

00:28:51 Nathan Lambert: It’s reasonable to me.

00:28:52 Lucas Atkins: It was, it was a good number. I mean, I would say that, um, it, it was helpful to be able to, you know... This was like, how do we alleviate as many concerns as possible? Or how do we check off as many boxes, right? And it’s like, if we’re trying to do this in the shortest possible amount of time, like, we need to focus on what we’re good at, which is we- pretty good at post-training, and how do we get to the point where we’re able to do that? Well, we have to have a pretty strong base model. How do we get a strong base model? We’ll-- we have to, you know, figure out how to do it, perform, you know, efficiently across many, many GPUs, and then data’s, you know, extremely important, so getting a partner that could, you know, help us with that, and we could offload some of that. It, it- there ended up being, obviously, as you, you know, alluded to earlier, like, a lot of, uh, working with Datology and, and, and others to make sure that the data accomplished what we needed it to. Um, I think that that is gonna be an interesting... You know, as we, as we- now that we have Large and we’re looking at, you know, kind of going further, it’s like, okay, you know, the, the pre-training data really has to be in service of what you wanna do in the post-training, uh, work.

00:30:10 Nathan Lambert: How did you identify this?

00:30:11 Lucas Atkins: Like, like-

00:30:11 Nathan Lambert: Like, like- ... did, did you identify this through Mini and Nano, or, like, how’d you come to think that this was so important?

00:30:19 Lucas Atkins: Data in general or, or just-

00:30:20 Nathan Lambert: Or like this in form of post-training

00:30:21 Lucas Atkins: ... of optimizing it for the post-training? Um, I- really ob- observing other, other players, I think. I mean, it’s, it’s... You know, the, the true base model has kinda stopped really being a thing.... around Qwen2, but definitely around Qwen 2.5, um, where you started to see how much post-training data was making its way into the, the, the base models themselves. Um, and then you start to see the models that have done that, how malleable they are with RL, Qwen 2.5, Qwen3 being a good example. And you start to see like, oh, yeah, like they are, uh, doing as much in the last probably thirty percent of training to make it so that when they go to do RL or post-training, they’re gonna have a really good time. Um, you know, they’re just complete-- they’re way easier, way more malleable, way more performant than what you had in Llama 2 or Mistral 7B. So, um, I knew that i-in-intuitively, kind of going into this, but it wasn’t until after Mini and Nano, yeah, where, where we kind of... Well, definitely 4.5B, where we were like, “Yeah, we definitely need to juice our mid-training quite a bit.”

00:31:31 Nathan Lambert: Yeah, I agree. Okay, this was fun. We could- we’ll probably revisit themes from this. I think that, um, I can definitely go over time and keep chatting because I’m enjoying this. And for context, Mark and I had coffee at some point when I was at some conference in SF, and I was like: Damn straight, this is a fun bet that you’re making. So I’m trying to recapture as much of this as you can. Um, for context, it’s like in July, which is similar to when you decided to start this model, which is when, like, Qwen Coder came out, Kimi came out, um- ... GLM 4.5 came out, and I was just, like, looking- and Llama had kind of been, like, become a meme of going away. And that’s why I launched the Adam Project, where I was like: Come on, we need to have some people doing this. And I think that it’s, like, hard in the US because I think there’s so much money to be made on AI. Like, the company- the big tech companies are like: “We see it, and we’re gonna take it, so I don’t need to bother with, like, caring about open models ‘cause we don’t need it.” But from, like, an ecosystem co- perspective and a long-term tech perspective, I don’t think that works very well for the country. So it’s kind of this weird middle ground of like, how do you convince people to actually build open models? I was on... Like, I have calls with people in government asking me, like, what would I actually do? So it’s, like, very hard to think about this. And I have this- and then it’s just, like, to hear that you guys are just making this bet on this is very fun to me, but it’s also, like, based on actual learning from trying to do this. So you’ve been trying to train open models. I think Mark and I have both been at Hugging Face in our past, and you’re, you were trying to sell people on using open models, and there is a market for this, but it wasn’t enough to not have the base models. So I think, like, talking about your experience in selling on-prem open models and why you needed to train your own end-to-end, and why you needed to train bigger, is great because I hope there are more stories like this, and it kind of fills a void and inspires people to work in it. So how- however you want to take this prompt.

00:33:24 Mark McQuade: Yeah, I can jump in. Um, I mean, yeah, I mean, wh- when I started Arcee in 2023, right, uh, it was... All we did was post-training. Uh, and we worked with, uh, a lot of large organizations and did model customization, you know, for their use case on their data. Um, and we were using Llama-based models, Mistral-based models, and then, you know, some Qwen. I don’t even know if we actually did much Qwen, right, Lukas, at that time, but-

00:33:54 Lucas Atkins: No, we did. Yeah, we, we- Later on, but and then-

00:33:56 Mark McQuade: Later on, right? Uh-

00:33:57 Lucas Atkins: We did, and then we ended up not, because after a lot of Chinese models started to come out, then the companies didn’t wanna use Chinese models, so then we kind of went... Yeah, it was kind of just tricky.

00:34:08 Mark McQuade: Yeah, and people don’t realize that that’s real.

00:34:10 Nathan Lambert: People don’t realize that that actually happened.

00:34:13 Mark McQuade: Yeah, no, that’s, that’s a real thing. That’s why we, we started going down to pre-training was because, well, you know, Meta did their thing and kind of got out of it, right? So there was the, the main US player got out of it, and, and we were working with a lot of US-based enterprises that were not comfortable using Chinese-based architectures. And if you wanted to use the best open models of the day, it started to really trend towards, you know, the Chinese labs. Um, and to the point where we are now, where it’s like, you know, ninety-plus percent of the top mo- open models are coming out of China, um-

00:34:47 Nathan Lambert: Yeah, like, Cursor’s building on it and stuff. Like, people are building on these things.

00:34:52 Mark McQuade: Yeah. So, um, we said, “Okay, let’s...” Instead of we were so reliant on the Metas of the world, the Mistrals of the world, and Mistral largely stopped open sourcing, uh, you know, fully. So we said: You know what? We’ll just go down the stack, and we feel we’re capable enough to, to, to train our own models from scratch, and then we control the, you know, the stack. We can, you know, we, we control the core of, of... as opposed to relying on others to release great models. And, um, and then during this time, you know, it just happened to be that, um, you know, there wasn’t a tremendous amount of US companies doing it. So, um, from our perspective, it was kind of a, a win-win, in that we were able to own more of the stack by going down to pre-training and creating our own models, as well as we were entering into a, like, a space that there wasn’t a tremendous amount of competition, to be honest. Um, and, you know, I-- Lukas and I had said this yesterday, I, you know, I think as a startup, every startup doesn’t want to directly compete with, you know, X or OpenAI, or Anthropic, or Google because they have more money than God, and they can do whatever they want. Um, but when you’re doing open weights, you don’t-- it’s, it’s a different kind of compe- they, they don’t sit in there, right? You’re kind of going into your own path, where there isn’t a tremendous amount of players, and you can kind of find your, your way and, and build your niche and, and kind of go from there and, and become something big. So, um, it kind of happened to all coincide for us back in, in July, and, and we went all in.

00:36:23 Nathan Lambert: Yeah, yeah, like, uh, the, the all-in thing is real because this is expensive. I think that- ... I could dig up in my research the cost of daily, um, twenty-four T8 B300. So I think I’ve seen this type of cost at AI too, where we have long rentals, and we’re like: I know exactly how much this costs, and it’s like, it’s not cheap. Are you... A, a way to transition this is like-... do you see the demand? Like, you were selling open models, like, does this kind of be continuous, where people are like: “You helped us deploy this model, but it’s not good enough.” Like, is, is that something that’s happening, and you’re like: “Well, we have this, and we can help you do it coming in this time?” Or is it like you need to build it... It’s like, is it a we will build it, and they will come type of situation? Like, how much- ... continuity is there in this?

00:37:17 Mark McQuade: Yeah, I think it’s largely-

00:37:19 Nathan Lambert: I-

00:37:19 Mark McQuade: I, uh, from my perspective, I think it’s largely if you build it, they will come. Because we stopped, you know, focusing on that whole revenue generation side of the house when we started to go all in on being this, you know, frontier lab in the open source side. So, um, there’s a couple pieces to that, that, that I think we should all be very proud of inside of Arcee, is that we not only went all in by committing a significant amount of capital. Like, we, we committed, you know, sixty-five, seventy percent of our capital to these models, which is a large amount for a startup. I mean, we didn’t... So that’s not like a dip your toe in, that’s like, we’re all the way in.

00:37:55 Nathan Lambert: Yep.

00:37:55 Mark McQuade: Um, but we did that at the same time as abandoning essentially the whole revenue angle to go all in on it, because we couldn’t focus on both. So we said, “We know how to make revenue on open models. We’ve been doing it for two years. Now, let’s take a step back, because it wasn’t, uh, in a repeatable or sustainable way that we h- the way we had that business set up. Let’s take a step back, let’s build these models from scratch, let’s come up with the, the Trinity family, then let’s go back to generating the revenue side of the house and the monetization piece,” which I think we are in a good position to capitalize on even more now, but we, we took a... We, we, we kind of walked away from it to do what we’re doing here.

00:38:36 Nathan Lambert: Yeah, I love this.

00:38:36 Lucas Atkins: Yeah, I mean, when you have... When there’s only, like, thirteen, you know, uh, researchers who would... Well, we’re, we’re doing obviously our own products and own models, but when you’re working with customers, like, inevitably, those are the same people that need to help train those models for customers, and we got to a point where we were really beginning to, like, do mini and nano. We were getting down to, like, the start date of the cluster, where, um, having myself or Mark, or even, you know, Varun and others, like, pulled into customer or, or, or, uh, conversations or contracts, like, it was not-- we would not be where we are if we had continued, you have know, working with, you know, ten customers at once. So-

00:39:19 Nathan Lambert: But-

00:39:19 Lucas Atkins: ... we, we scaled that down pretty drastically. I do think that when... You know, Mark and I put a lot of thought into, “Okay, well, we’re gonna spend all this money to train these models, like, you know, w- how do we not...” I think, uh, one of the things that makes the idea of, of going all in on training open weight models hard, is that you’ve seen other people try it. And, and like M-

00:39:42 Nathan Lambert: Um, like, like do you think Meta or do you think Meta or Mistral went all in?

00:39:46 Lucas Atkins: I, I think, well-

00:39:48 Nathan Lambert: Meta obviously did.

00:39:48 Lucas Atkins: I think they, they both... Yeah. I think, I think that when I say all in, I mean more like Mistral was, was one of the core ones I’m thinking of, where- ... they were a venture-backed company that, like, had a, a, a fiduciary responsibility to bring in money, but were also trying to release open weight models, uh, for, you know, the West, and for their communities, and for the world. And, um, they tried doing closed versions, and then monetizing off of that. They, they also kind of have more recently, luckily, for all of us, gotten back to their kind of Apache 2.0 roots, and-

00:40:30 Nathan Lambert: Oh, my God.

00:40:30 Lucas Atkins: And-

00:40:30 Nathan Lambert: Have you seen the download numbers on Mistral 3 Large?

00:40:33 Lucas Atkins: I haven’t. No, what is it?

00:40:35 Nathan Lambert: Oh, s- no bueno, sir.

00:40:38 Lucas Atkins: Hey.

00:40:39 Nathan Lambert: Carrying on. Sorry.

00:40:41 Lucas Atkins: But, I mean, yeah, you know-

00:40:42 Nathan Lambert: Um, Mist- the, the Large Instruct model has downloads in the last month. I honestly don’t know what’s going on. Maybe there’s some, like, quantized version out there. I, I was confused.

00:40:50 Lucas Atkins: Maybe. Well, I mean, yeah. But I think that we-

00:40:52 Nathan Lambert: It’s, it’s hard to get adoption. The competition is insane.

00:40:55 Lucas Atkins: Hmm. Well, that’s, that’s- ... yeah, I mean, and that could be a whole conversation also, is, like, how do you actually get people to use it?

00:41:00 Nathan Lambert: I was gonna ask you, like, how do you get people... How do you get people to- - really sell into this? You said you’re good at it.

00:41:06 Lucas Atkins: Yeah, I think that the-

00:41:08 Nathan Lambert: Continue your point, we can come back to it.

00:41:11 Lucas Atkins: No, no, but they... I think they all kind of tie into it, is, is... We knew that the, the market was there for, for custom models. It was two years ago, frankly, and it’s even more so now, because RL has drastically, uh, increased the areas that you can hill climb and become really powerful with a tiny model. Um, and but, but also, people are beginning to see how powerful, you know, uh, te- uh, cust- or, or training in a, a, a product is. Like, you see Claude Code, you see Codex, you see, um... I think Deep Research was kind of one of the first ones that really kind of opened my eyes to what was possible, when you kind of are kind of training in the same environment that you’re serving your users. So we knew that, that people wanted it. We’d, we’d had good success with, with customers in the past using other people’s open models. So, um, it was less of a question of, like, could we monetize it, or will we? And it was just a matter of, um, could we get a model, you know, that pe- that, that we would feel that, you know, given a, a wide suite of basically being able to pick any model in the world, would, would our researchers and, and would our teams re- reach towards our own? And, uh, luckily, I think we’re there. Um, on, on the-

00:42:31 Nathan Lambert: Uh

00:42:31 Lucas Atkins: ... on the topic of, like, how do you get people to use it? How do you get adoption? You know, I’ve never wanted Trinity, uh, or our biggest advertising thing to be, like, US. You know-

00:42:45 Nathan Lambert: Yeah, I know

00:42:45 Lucas Atkins: ... like, if, if your entire-

00:42:47 Nathan Lambert: I know, man, it hurts me.

00:42:48 Lucas Atkins: Yeah, if your-

00:42:48 Nathan Lambert: I spent months reckoning with this.

00:42:50 Lucas Atkins: Yeah. If, if your entire, uh, you know, value prop is that you’re an American company-... great, but ultimately people are gonna use the best. Um, and so I think that we’re gonna be able to serve and, and the people like that need a US-based model because their compliance or legal teams won’t let them use something out of China, it’s gonna be a fantastic option. But I think, you know, kind of the next phase of what we’re doing as a company is, all right, now we’ve, we’ve proved to ourselves and maybe the, the wider industry that like we deserve to be in the conversation, and we can train models of this scale. Um, then it’s like, okay, how do we train the best one? Uh, ‘cause really, I mean, people’s loyalties are very fickle, and, and, yeah, you, you go to what’s the best. I guess it’s like, how much do you think

00:43:41 Nathan Lambert: you’ve learned about being able to tune a model narrowly by going and building the whole stack? Um, something we talk about is like ability- ... to specialize models, and I kind of, of opinion that you just make a better general model right now ‘cause the pace of progress is so high. And but the question is like, can we tune a OLMO that’s very good at science or something? And I- ... w-would guess that training the entire model, you’re going to be able to actually do a better job at what you were doing, but I don’t know how to articulate why or what that looks like.

00:44:18 Lucas Atkins: Um, I mean, the, the, the simplest answer to that being yes is just that... or the simplest reason why that’s the answer to the question is yes, is because we know what went into the model. Like, we know what it actually saw at the later stages of training during the decay. Um, and so that all- that helps influence, A, what are we tr- what kind of data and what topics and, and what format are we giving these models, uh, in post-training? But it also allows you to know like, okay, where, where do I absolutely wanna crank, you know, how, how many- how much of this, say, 230 billion dataset, do we want it to be math or, or, or, or coding? And a lot of that’s influenced by what you’re able to put in-

00:45:06 Nathan Lambert: How, how much of your post-training-

00:45:07 Lucas Atkins: ... post-training

00:45:07 Nathan Lambert: -do you expect to redo? Like, uh, how much can you say about when you’re serving something on-prem? Um, you- you’re not gonna redo the pre-training. You might, for a very big customer, redo mid-training or do continued pre-training- ... in which, in that case, you do need the pre-training data to keep, keep it being stable. Which is a use case where like I’m- I would love to see a paper that’s like, “Because of OLMO being open, we continued to pre-train on biology, and we mixed half of their exact mid-training dataset in with our dataset, and it, and it worked,” yadi, yadi. Like, you could obviously- ... do that, but how much do you think is gonna be like the standard, you fine-tune the last instruct model, or do- are you gonna have to retouch the post-training for a customer? Because that, like, I, I really feel like-

00:45:48 Lucas Atkins: Um

00:45:48 Nathan Lambert: ... it’s just at the end.

00:45:50 Lucas Atkins: It, I think, I think-

00:45:50 Nathan Lambert: But it would be fun if you had to change it.

00:45:52 Lucas Atkins: For the most part, um, I think a lot of tasks will be fine just starting from our, our, our, po- uh, like the released, you know, official post-trained version. Um, now, that’s for maybe simpler tasks, is the wrong way to frame it, but if it’s like, “Oh, hey, we’re doing a deep search agent. We want it to do 30 calls and, before...” That would be a good use for just starting with the finished model that we released that’s already post-trained. Now, if we’re going into something along the lines of, um, a very low-resource programming language or, um, something that it didn’t see a lot of in, in, in pre-training, um, or it’s kind of like a, you know, we’re wanting to train this thing to be really good at humanities last exam, but tools. Um, once we get into the world where we’re having to, especially... Actually, I have a much better answer to this question as I was thinking through it, but most of that holds the same. I think that the, the, the world where we’re gonna be doing a lot of extra instruct and, and SFT and, and post-training is gonna be when we’re trying to distill capabilities from large, like into mini or nano. So say like, oh, you know, this large is, is, is really great at invoice processing, but it’s also 400b, and the, you know, the company doesn’t wanna be hosting that on-prem, you know-

00:47:24 Nathan Lambert: Ah

00:47:24 Lucas Atkins: ... let’s go out generate a new one.

00:47:25 Nathan Lambert: Do you have costs off the top of your head for, like, what the hosting costs are for each of the model? Like, do people... Are people all gonna host these models in the same way, or is there actually-

00:47:32 Lucas Atkins: Uh

00:47:32 Nathan Lambert: ... a wide variance? And if you have, like, the same three models- ... do almost all of your customers end up hosting the same way, or do you end up doing a lot of, like, how do you configure the model to fit in the right hosting for them? Like, is that part of-

00:47:44 Lucas Atkins: It depends

00:47:44 Nathan Lambert: ... the business model?

00:47:45 Lucas Atkins: It, it, it, it kind of... And we tried to move a, a, a little bit further away from that because you get into the risk of being like, like a consultancy, and it’s- that becomes tricky, where there’s not a very clear separation of concern. But, um, for the mo- it would change depending on, were they using AWS? Did they have a commit with Azure? Um, if not, okay, then we, we can go to, you know, someone like Prime Intellect or Parasail and, and get a, you know, maybe a, a cheaper rack of eight. Uh, it just really depended. Uh, there’s quite a bit, um, of, of people that were also serving them, just using, like, Llama CPP. So, like, on CPU-

00:48:25 Nathan Lambert: Uh, is the 400b designed to be, to fit onto one rack of eight 80 big gigabytes in FP8? Is that how you designed it? ‘Cause Llama- ... Llama four, whatever, Llama 405b was the same. It was like one rack in FP8 works pretty well.

00:48:41 Lucas Atkins: It’ll do- we... well, you’ll be able to get really good throughput, a little bit lower concurrency on a, a rack of eight H100s at FP8, and then for, like, our, you know, what we’re serving, we’re serving them on, uh, a series of H200s, but we’re not doing, like, multi-node inference. Uh, but that’s just to add more, you know, replicas and- ... other kinds of things.

00:49:03 Nathan Lambert: Hopefully, eventually. I think that the-... Do you have anything else to say about selling open models? I think that generally, like, how do you think about the market for AI? ‘Cause I see the market as being so big, but the- with specifically with open models, it’s so hard to measure. I think I’ve started talking to some of the Chinese labs at all- as well, and I like to ask them, like, this is very US-centric and like Fortune 500 or whatever, and it’s just like, who the heck uses these models? I think- I guess another question is, like, what license or do you know the licenses you’re gonna use for the biggest models? And I think they’re, like, you’re, you’re playing with fire ‘cause people can use it for free, obviously, but potentially- ... you’ll get to hear like, “Oh, s**t, somebody actually used our model for this.” And I think any successful business, you’re gonna want... You, you, you know that this model is not gonna be very relevant in a year with the pace of progress. So like- ... how do you think about your license decisions?

00:49:55 Lucas Atkins: Uh, we- you know, with the 4.5B, we tried to do like a, like a, a reve- one of those revenue-gated licensing. So it’s like, oh, it’s completely free for you to use for commercial and whatnot, but if you or your company made over, I think it was like $1.7 million last year, then you need to come to us and get a license. And what we ultimately found was like, it, it didn’t... Maybe for some people who are just only trying to train the model, release it on Hugging Face, and then just call it a day, maybe that is a huge requirement. But when so much of our, our, our company is built around, you know, training custom versions of the models, and, and not even just ours, but in general, even before we did pre-training. Like, at the end of the day, i- as long as we were using it, a- and we knew that we were in full control of, of whether- if we really succeed, it’s because we trained the models, we did them well, and we executed on it well. If we fail, it’s because we, uh, didn’t execute, instead of, oh, some company just stopped releasing good open models. Um, so we eventually switched to just Apache 2.0, and Trinity Large is also gonna be Apache 2.0. Um, you know, I’m- I think it is-

00:51:23 Nathan Lambert: I think this is the right approach. I have a big investor-

00:51:25 Lucas Atkins: Yeah, I think it-

00:51:25 Nathan Lambert: Without, without naming other companies, it’s easy- like, raising a lot of money, whe- or being Meta and releasing open models, and do it- and you could release it with non-commercial, and you could get all these, like... You could talk to, I don’t know, f*****g Adobe, whoever. Oh, Adobe’s too big. They’ll have good AI. Some... I don’t know, a bank. Bank of America. You could run Llama on Bank of America and make good money on this. But I just feel like the cultural home of open source AI, and I don’t think- it’s impossible to know who wins it, and I don’t think that you’re in the prime position, and I don’t think that it’s easy to win, but you’re doing a thing that aligns with it. It’s the person that just, like, commits to building the models and learning how the ecosystem works, and to rebuild the models based on the feedback th- that you get from people, and to just kind of commit to an evolving process. And if the whole thing works out, there will be a lot of value, and the person who understands it best should be able to learn how to extract said value. And I think that I’m personally, like, sometimes frustrated with Hugging Face, ‘cause I feel like they have sat on that s- a sort of position like this, and they- ... haven’t figured it out. Not that it is easy to figure it out, but I think that has to be the ideal of open source AI, of like, if it’s really gonna work, that’s, that’s what I hope it looks like. And it’s like, I, I don’皮 know, maybe you guys could do some of that. Like, I have a question of like, could you figure out how to make models that are more fine-tunable- ... after all this post-training? Because you need to sell it to a- you need- ... you, you know the customer’s not gonna want it off the shelf. And I don’t know how to train to post-training to make sure that you don’t, you don’t cook it. Maybe you just learn that you need to warm up the model in a l- in the right way, and you just learn the technique of training downstream. But when you talk to people doing research, the different base models have such different characteristics. I think one of them is character training. I did this paper, and the guy was like: “Qwen and OLMo love their character,” and I’m like, “I have no idea why.” And but it’s like Llama and Gemma, you can change them so much. And I’m like, “Dog, like, please figure out why this is the case.” And for one thing, it’s really cool, but also, like, in your case, that would unlock a lot of value to be like, we know exactly what the model’s gonna do, and we know exactly how to change it. So.

00:53:35 Lucas Atkins: Yeah-

00:53:36 Nathan Lambert: Uh

00:53:36 Lucas Atkins: ... it, it, that’s- no, you’re, you’re, you’re right on the money. I think that even, uh, going into the post-training at large, we, uh, one of our researchers came out with, like, a pretty cool, um, experiment and ablation run that they did on drastically reducing catastrophic forgetting. And I almo- I mean, this was, like, three days before we were gonna start doing SFT, and then we ultimately just... I, I ended up pausing on it because it was just throwing something in that wasn’t tested. But, um, yeah, I think-

00:54:08 Nathan Lambert: A good research lead. You did the right thing.

00:54:10 Lucas Atkins: Yeah, I think, I think one of the most important things long term, you know, as we look at kind of what our research priorities are for this year is, is there’s obviously just how to scale RL and, and make these- the end result of the model as good in as many situations as possible. Um, but I think the other half of that is, you know, how do we make the, the, the speed and efficiency and, and performance of customizing them as, as fast as possible, and as easy as possible.

00:54:42 Nathan Lambert: Yeah. Do you learn in making open models from your experience just kind of running these open software things in MergeKit and DistillKit? I know there was a whole license journey on one of those as well.

00:54:52 Lucas Atkins: Yeah, DistillKit.

00:54:52 Nathan Lambert: Do you feel like they’re kind of isolated?

00:54:54 Lucas Atkins: Or MergeKit. Um, yeah, I mean, I think so. I think that, that, um, you kind of have to play the tape out. With MergeKit-... it was by far our most popular piece of software we’d ever released, but it was so popular because it took something that isn’t fundamentally very complicated, but we ma- but it’s time-consuming, and standardization is great for things like that, and we made it, uh, you know, streamlined and easy to do and fast, and you could experiment and ablate really quickly for, you know. And, and so I, I think that when we switched that to, like, a, you know, a, a similar, uh, revenue-based licensing, like, it, it didn’t end up having the value prop that was important because are you gonna pay Arcee, you know, thousands of dollars, or are you just gonna have one of your researchers-

00:55:52 Nathan Lambert: You’re gonna have clone code in a week, right?

00:55:52 Lucas Atkins: recreate it in a week, right? Yeah, so it’s-

00:55:55 Nathan Lambert: In a day.

00:55:55 Lucas Atkins: It’s, it’s kind of... It, it’s remi- it’s remembering like, okay, what is- what problem is this solving, and is this even a prob... Like, is the solution to this monetizable? Um, and so MergeGit, we brought it back to the original license, but I think with even viewing the models in the same way, it’s like it’s... Open source is an unbelievable marketing tactic. Like, there’s no one would care about Arcee if we weren’t open sourcing stuff, ‘cause as soon as you do something closed source, if you’re not the best or the cheapest for your price point, I mean, your performance point, no one’s gonna use it. Because-

00:56:30 Nathan Lambert: Um, another question on this. Um, do you think that open models are kind of at a disadvantage when progress is so high? Because it’s potentially easier to swap APIs than open model configurations, especially if, like, model weights are changing sizes or something like this. Where it’s like, “Oh, I can just upgrade to the new Opus, and I do this.” Like, does that, like, uh, decentivize people from using it? Or do you think most of the people are like: “I can only use open models, therefore, I’m gonna use open models?”

00:56:56 Lucas Atkins: Uh, I think for the people who are using, like, s- either self-hosted or, you know, um, uh, bespoke, uh, you know, engines to, to run it, where they have complete... You know, in a VPC or they have complete control over, like, data in and out, egress, ingress. I don’t think that’s really gonna be so much of a problem because they’re obviously doing it for a reason. Um, like, they’re either for privacy or security or, or HIPAA or SOC 2. For whatever reason they’re doing it, um, I, I don’t think that that’ll be, um, so much of a blocker, but I definitely do think that, um, you know, by far, e- even, even with some of the, the larger open... You know, like inference players, like Together and Fireworks, that, that host a lot of open models. Like, being feature- being on feature parity with a lot of these, these larger labs’ APIs is gonna be extremely important, um, o- of being able to serve, you know, um, with features that they’re used to, like prompt caching, that kind of stuff.

00:58:03 Nathan Lambert: Yeah, are- like, I, I think I saw that you guys are setting up an API as well. Is that kind of what the vision there is, is being able to o- offer parity at least, or, like, make it easy for people to consider it?

00:58:13 Lucas Atkins: I think so. I, I- we’re- we very... Yeah, we are doing our own API. We are hosting it. Um, we haven’t- we, we push a lot of that through Open Router just because it’s such a great place to get, like, discovered. Um, as... If we see, like, tremendous growth there, that would obviously be where we’ll, we’ll invest very heavily. Um, whereas the right move might be to let other people host it, and we invest super hard on the infra for, like, make- taking advantage of the models, um, and, and customizing them. There’s, there’s, there’s a few avenues we have ahead of us then, and we have, you know, projects going kind of toward to poke at each one. Um, and we’re just kinda getting as much data as we can before we... I mean, we’re gonna have to go all in on another direction soon. Not, not like pivoting away from pre-training, but now that we’ve done that, now w- what’s the next big bet we’re gonna make, and how do we go fully into that? So we’re trying to figure out what that is.

00:59:12 Nathan Lambert: Yeah. My two last kind of, like, real questions are, like, one is... I guess I can start with, like, where do you see the open model ecosystem? Do you think- where would you see it changing substantially in the next six or twelve months? I, like... Or, or do you? Or you just kinda think we’re marching along for a while?

00:59:31 Lucas Atkins: No, I think we’ll, I think we’ll, we’ll be... I, I, I don’t think it’s an unrealistic prediction to make that by the end of 2026, like, the best model in the world is, is some degree of open. Uh, I think that’s very, very possible, especially with, like, what I’ve seen GLM and, and MiniMax do recently. Um, they have started to find that secret sauce that takes you out of just being good on benchmarks and, like, genuinely useful in people’s day-to-day workflows. And, um, I wouldn’t- like, if, if I, you know, came back, and I... Someone came from the future and told me that the best model in the world was, uh, an open-weight model, I wouldn’t be surprised. I actually think we’re on a, a, a super good trajectory, and, and, and fostering and, and promoting that kind of work and adoption here in the United States is gonna be extremely important.

01:00:24 Nathan Lambert: And where do you see the company going? ‘Cause like, like, I have my guess. Like, you kind of hopefully-

01:00:31 Mark McQuade: What’s, what’s your guess? I wanna hear your guess.

01:00:31 Nathan Lambert: Um, you can hopefully do a mix and kind of oscillate into trading when you get... Like, you need to start having the feedback of the real world. I think that’s obvious. Like, it’s o- like, it’s... Well, obviously, you need to make money to survive as a company, but then you need to start using that as the feedback to guide training. And then it’s like, you need to figure out how to balance and do some of them at each time, and you can plan your cluster at different times, and then you kind of... Hopefully, they become a, a loop across each other, and they kind of make it so obvious of why you each need them, ‘cause it, it seems somewhat natural.

01:01:03 Mark McQuade: Yeah, I mean, exactly. You know, you kinda hit, hit it right on the head. Um, you know, getting feedback and then kinda steering the ship from there, um, is, is probably-

01:01:15 Lucas Atkins: ... exactly what we’ll do, but we have a good idea already. I mean, first and foremost, you know, we talked about it earlier, w- we’ve spent a tremendous amount of money. So, uh, we need to go raise some money after we - after we get, you know... We need people to back the, the, the mission and the vision of US open source and, and, you know, so, um, because, uh, you know, we, i- i- Lucas had mentioned about, like, MergeKit and how we flopped the license and, you know. I mean, we’re a smaller-sized start-up. We have-- we’re-- we gotta think of kinda unique ways to try and generate revenue because we don’t have the money of the large labs. So, uh-

01:01:52 Nathan Lambert: Well, I think it’s a benefit to the employee. I think a lot of these labs have over-raised.

01:01:56 Lucas Atkins: Yeah, I like, uh- uh, I-

01:01:57 Nathan Lambert: OpenAI, Anthropic, and all of them are fine. Like, with the OpenAI, Anthropic, Cursor scale, like, let it rip. They should, they should really rip the raising. But all the other companies that are stuck at the, like, the one to two billion range without, like, obvious traction, like, the risk goes to the... I mean, you could-- a lot of them do secondary, so a lot of the founders get out. But it’s like, the risk is the employees get nothing.

01:02:21 Lucas Atkins: Yeah. Yeah.

01:02:22 Nathan Lambert: There is a lot of money, but that’s also why I like the approach, ‘cause it’s like, “Oh, you’re doing the actual start-up thing.”

01:02:28 Lucas Atkins: Yeah, yeah. Yeah, I mean, I think... W- what I was gonna add to what Mark... is just like, what- whatever we do from, uh, uh, uh, scaling and, and speeding things up and growing, um, my goal is to keep our research and engineering teams pretty small. I think, I think that one of the reasons we’ve been able to, to move as quickly as we have is it’s been, like, a small group of, like, highly intelligent, smart, and opinionated people sitting in a room, debating in good faith on decisions. And I think that that’s, uh, uh, under the constraints of, “Hey, we don’t have five hundred million dollars to go and, you know, to rip on, on, you know, X, Y, and Z.” So and I think that’s kind of where creativity comes from, and I think that fostering a culture like that over time is how you can kind of make it so that excellence is less of like a, um, an accident, and it’s actually, like, a by-product of the way that you work. So, so we’re gonna stay small, we’re gonna stay lean, but, um, I, I do think that, like, the, the major, um, kind of challenge for us over the next probably six months, beyond any other models we might have, kind of, uh, think or we’re thinking about, is, is getting up to, like, post-training parity with the likes of DeepSeek, and GLM, Qwen, and others.

01:03:47 Nathan Lambert: Yeah. I, I hear lots of horror stories about this, where it’s usually and-- it’s-- you end up having people that are going after different important abilities, but, uh, like, doing each of the abilities alone is pretty easy to hill climb, but then you just end up with such a mess. It’s like you’re- ... building a custom puzzle, and you’re building all these custom pieces, and they’re magnificent, and then you’d have to, like, pick up these pieces and assemble this unknown thing at the end. And it’s like-

01:04:12 Lucas Atkins: Like they didn’t have the same designer, right? Yeah.

01:04:15 Nathan Lambert: As AI2 is barely scratching the surface of this. Like, you talk to the people at the frontier labs, and it’s like, holy cow, like, post-training is really the Wild West. But a lot of it works. I think, like, we find-- like, even like model merging gives a ton of performance across the whole- ... training pipeline. It’s like- ... you merge at pre-- you merge after each pre-training stage, you merge in post-training. It’s like-

01:04:35 Lucas Atkins: Roon can tell you.

01:04:36 Nathan Lambert: But merging post-training becomes a lot more complicated because you- ... can have all these domains and things, uh.

01:04:41 Lucas Atkins: Well, in, in merging, you know, it, it actually, it used to be very YOLO, um, the way we used to do it, and, and Charles, who, who created MergeKit, I call him, like, chief alchemist, and, like, you’d kinda just send him ten promising checkpoints, and he’d come back a day later with, like, some insane, you know, model that was really good at all of them. And, and you can’t do that as much in post-training anymore because of, uh, of just the, the formatting and the way that RL is done. Like, you do have to be a little bit more surgical about it, but yeah, everyone can tell you, like, any time we start to see anything worrisome at all in training or, or, or even something going really good, you know, “Lucas, what do we do?” I’m like: Merge it. I’m like, just-

01:05:21 Nathan Lambert: Merge.

01:05:21 Lucas Atkins: ... I’m like: “Just take it, just merge it. Let’s see.” And more often than not, it fixes it, so...

01:05:27 Nathan Lambert: Um, do you merge during RL? Like, you could just, like, merge the last few checkpoints and resume or something?

01:05:32 Lucas Atkins: We’ve ex-- we’ve, we’ve dabbled in that, not, not for what we’ve done. You know, again, a, a lot of the, the mini, nano, and large story for Trinity is, like, getting to a level of... what was my level of complexity I was comfortable with us undertaking, and then, uh, not introducing anything more. So, um, not yet. But we, I mean, we, we, uh, regularly merged. We didn’t do it for LARP, but we used to merge a lot, um, during just, like, your standard, uh, um... When we’d do, like, RLHF, we used to do a bunch of merging. We’d do it, like, every five checkpoints. We would-

01:06:11 Nathan Lambert: Online RLHF or D-DPO?

01:06:13 Lucas Atkins: There’s DPO.

01:06:15 Nathan Lambert: Yeah. It’s so much easier to get started. One of my goals is to have somebody figure out how to do actual online RLHF, pure LM feedback, obviously, for scaling. But it’s just like- ... it’s, it’s unsavory to it’s just, like, doesn’t look like DPO-

01:06:28 Lucas Atkins: Yeah, I mean, if, if, you know, if GRPO and kind of op-- in, in the, the present day RL regime, like, if that hadn’t materialized when it did, I think that would’ve been a big topic in 2025. But I do think that, you know, GRPO and just the overall, um, DeepSeek and o1 style reasoning and thinking and RL kind of... Any, a- any person who is thinking of doing that for, like, performance reasons, realize that there was something that had fifty thousand papers released every day on how to do it. Um- ... that was kind of probably right where you’d get the same amount of performance.

01:07:07 Nathan Lambert: Um, do you force dog feeding? Do you make yourself-- do you guys use your own models to understand them? Like, do you, like, make that a thing?

01:07:14 Lucas Atkins: Uh, Mini was the first one we could actually start doing that with, um, a- at least for, uh, a more general day-to-day tasks. So a lot of our, like, internal Slack, we have stuff that, like, monitors Twitter and LinkedIn for feedback on Trinity and, and, and that kind of stuff. That all runs on Trinity Mini now. Um, and then, uh-... you know, we, we put a good amount of work into, into large being, um, you know, good in, in a bunch of your, like, OpenCode and, and Cline, uh, and, and Kilo Code. So, um-

01:07:45 Nathan Lambert: Uh, what does that, what does that work look like?

01:07:49 Lucas Atkins: Uh, working with those guys to get data. And then, um-

01:07:53 Nathan Lambert: That’s, I mean- Good for me to know.

01:07:55 Lucas Atkins: I mean-

01:07:55 Nathan Lambert: I should do that, I guess.

01:07:58 Lucas Atkins: Yeah. Yeah, working with, uh... Or, or I mean, it- the way it started was us, like, using open models and then, like, passing those through as the base URL, and then, like, getting the logs from that. Um, and then realizing that, like, that translated pretty well. Um, and then over time, obviously turning this-

01:08:16 Nathan Lambert: Um, can you expand on this? So I was gonna ask you-

01:08:19 Lucas Atkins: So-

01:08:19 Nathan Lambert: -if you’re, like, using these open models regularly, ‘cause I, I’m just, like, Claude Code psychosis, man. I’m like, “Can’t take that away from me.”

01:08:26 Lucas Atkins: Yeah, I, I use, I use four... I’ve used 4.7 a lot. I think 4.7 from GLM was one of the first ones that could replace a lot of my day-to-day. Uh, I’ll still reach for Claude Code or even 5.2 Pro if it’s, if it’s, like, something that’s, like, really... I- if I do not know how to measure what success looks like for something, I’ll usually use those. Um, but, uh, yeah, I mean, it, it- even using DeepSeek before, um, kind of their May update was hit or miss. But, um, yeah, w- the reason I decided to, like, start talking to these people and working on, like, how can we get data and, and start making our models good in these systems was I would use them. I had a, um, you know, something that would grab the logs, like, it, you know, inter- as a proxy, so it’d like grab the logs and then format them in the messages format. And then I saw that and went, “Yeah, that’s... You can make a pretty good filter for just, like, standard stuff that you don’t want, and kind of hit a scale.”

01:09:30 Nathan Lambert: Yeah, it makes sense. So, so you’re like, uh, open code will let you look at the data, and then you’re probably gonna get a sense for... Like, I don’t even actually know how the, on the back end, the code agents in open code format data, which I think is actually something I should just go look at, ‘cause then you can design around.

01:09:44 Lucas Atkins: Uh, they’re all different. Yeah. Yeah, but you just have to- you just- basically, it all starts from like, what do you want your format to be? And then how can you take what, what those look like to, you know, to... How do you force it into that? The hard thing, though, is, is with newer models like MiniMax and 4.7, the way they do interleaved thinking is, is like... You know, I’m a big believer in post-training. Like, if you’re gonna do interleaved thinking, like, every sample in your data set should be that. Um, it, you know, it should follow that same format and that same behavior. So, um, that gets tricky if you’re trying to, like, take a bunch of Nemo tr... Or, or, or, well, like, uh, DeepSeek data and Qwen data, and then, oh, we’re also trying to mix in MiniMax, and at that point, you’re- it, it gets really difficult ‘cause they all handle thinking slightly differently.

01:10:34 Nathan Lambert: Yeah, I can buy this. Um, okay, this was fun. Any last predictions or things you want people to know about the model? I will say that, um, when you debuted the Trinity models, you had a great blog post that was very to the point, that covered a lot of this. So I’ll definitely link to the, um, what is it? The Trinity manifesto. I enjoyed reading it. So I’ll link to that in the show notes, and, oh, hopefully you have a new one for me to read when you’re done with the model.

01:10:58 Lucas Atkins: Yeah, we’ll do- we will have a tech report. We’ll have a tech report for you, too. So we, we never, we never did a tech report for 4.5B Mini or Nano because we were so focused on just getting to large, but we also thought it’d be very interesting to write it under the, the... How do you go from 4.5B to a 400B MoE in six months, and, like, what did we learn-

01:11:19 Nathan Lambert: That’s right

01:11:19 Lucas Atkins: ... when you’re viewing it as a whole, so.

01:11:21 Nathan Lambert: That’s about the timeframe that, um, Ant Ling took, too, as well. Ant Ling, uh, the anchor, we talked about, they’re like... It took us about six months to do, um, Ring-1T and their 1T models, which, like, it sounds like a lot more, but I think that’s about the same. It, it depends on compute and configs and stuff to go from, like- ... basic modeling to big MoE, which is pretty interesting to see a lot of people speedrun this sort of thing.

01:11:46 Lucas Atkins: Yeah, it’s, it’s a really, uh... It is a logistical nightmare, but, like, I think everyone on the team has had a tremendous amount of fun over the last, uh, six months. So now the fun begins.

01:11:58 Nathan Lambert: Yeah. Congrats on the milestone. Congrats on the model existing. That has gotta be an almighty relief, and I’ll look forward- ... to see what you all are up to soon. I’ll stop by at some point next time I’m in the Bay.

01:12:10 Lucas Atkins: Yeah. Yeah, come by. Yeah, come by.

01:12:12 Nathan Lambert: Thanks for-

01:12:12 Lucas Atkins: Thanks for having us.

01:12:14 Nathan Lambert: Yeah. Thanks, guys.

Two weeks ago, I wrote a review of how Claude Code is taking the AI world by storm, saying that “software engineering is going to look very different by the end of 2026." That article captured the power of Claude as a tool and a product, and I still stand by it, but it undersold the changes that are coming in how we use these products in careers that interface with software.

The more personal angle was how “I’d rather do my work if it fits the Claude form factor, and soon I’ll modify my approaches so that Claude will be able to help.” Since writing that, I’m stuck with a growing sense that taking my approach to work from the last few years and applying it to working with agents is fundamentally wrong. Today’s habits in the era of agents would limit the uplift I get by micromanaging them too much, tiring myself out, and setting the agents on too small of tasks. What would be better is more open ended, more ambitious, more asynchronous.

I don’t yet know what to prescribe myself, but I know the direction to go, and I know that searching is my job. It seems like the direction will involve working less, spending more time cultivating peace, so the brain can do its best directing — let the agents do most of the hard work.

Since trying Claude Code with Opus 4.5, my work life has shifted closer to trying to adapt to a new way of working with agents. This new style of work feels like a larger shift than the era of learning to work with chat-based AI assistants. ChatGPT let me instantly get relevant information or a potential solution to the problems I was already working on. Claude Code has me considering what should I work on now that I know I can have AI independently solve or implement many sub-components.

Every engineer needs to learn how to design systems. Every researcher needs to learn how to run a lab. Agents push the humans up the org chart.

I feel like I have an advantage by being early to this wave, but no longer feel like just working hard will be an lasting edge. When I can have multiple agents working productively in parallel on my projects, my role is shifting more to pointing the army rather than using the power-tool. Pointing the agents more effectively is far more useful than me spending a few more hours grinding on a problem.

My default workflow now is GPT 5 Pro for planning, Claude Code with Opus 4.5 for implementation. I often have Claude Code pass information back to GPT 5 Pro for a deep search when stuck with a very detailed prompt. Codex with GPT 5.2 on xhigh thinking effort alone feels very capable, more meticulous than Claude even, but I haven’t yet figured out how to get the best out of it. GPT Pro feels itself to be a strong agent trapped in the wrong UX — it needs to be able to think longer and have a place to work on research tasks.

It seems like all of my friends (including the nominally “non-technical” ones) have accepted that Claude can rapidly build incredible, bespoke software for you. Claude updated one of my old research projects to uv so it’s easier to maintain, made a verification bot for my Discord, crafted numerous figures for my RLHF book, feels close to landing a substantial feature in our RL research codebase, and did countless other tasks that would’ve taken me days. It’s the thing de jour — tell your friends and family what trinket you built with Claude. It undersells what’s coming.

I’ve taken to leaving Claude Code instances running on my DGX Spark trying to implement new features in our RL codebase when I’m at dinner or work. They make mistakes, they catch most of their own mistakes, and they’re fairly slow too, but they’re capable. I can’t wait to go home and check on what my Claudes were up to.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

The feeling that I can’t shake is a deep urgency to move my agents from working on toy software to doing meaningful long-term tasks. We know Claude can do hours, days, or weeks, of fun work for us, but how do we stack these bricks into coherent long-term projects? This is the crucial skill for the next era of work.

There are no hints or guides on working with agents at the frontier — the only way is to play with them. Instead of using them for cleanup, give them one of your hardest tasks and see what it gets stuck on, see what you can use it for.

Software is becoming free, good decision making in research, design, and product has never been so valuable.

Being good at using AI today is a better moat than working hard.

Here are a collection of pieces that I feel like suitably grapple with the coming wave or detail real practices for using agents. It’s rare that so many of the thinkers in the AI space that I respect are all fixated on a single new tool, a transition period, and a feeling of immense change:

* Import AI 441: My agents are working. Are yours? This helped me motivate to write this and focus on how important of a moment this is.

* Steve Newman on Hyperproductivity with AI coding agents — importantly written before Claude Opus 4.5, which was a major step change.

* Tim Dettmers on working with agents: Use Agents or Be Left Behind?

* Steve Yegge on Latent Space on vibe coding (and how you’ll be left behind if you don’t understand how to do it).

* Dean W. Ball: Among the Agents — why coding agents aren’t just for programmers.

I’ll start by explaining my current AI stack and how it’s changed in recent months. For chat, I’m using a mix of:

* GPT 5.2 Thinking / Pro: My most frequent AI use is getting information. This is often a detail about a paper I’m remembering, a method I’m verifying for my RLHF Book, or some other niche fact. I know GPT 5.2 can find it if it exists, and I use Thinking for queries that I think are easier and Pro when I want to make sure the answer is right. Particularly GPT Pro has been the indisputable king for research for quite some time — Simon Willison’s coining of it as his “research goblin” still feels right.I never use GPT 5 without thinking or other OpenAI chat models. Maybe I need to invest more in custom instructions, but the non-thinking models always come across a bit sloppy relative to the competition out there and I quickly churn. I’ve heard gossip that the Thinking and non-Thinking GPT models are even developed by different teams, so it would make sense that they can end up being meaningfully different.I also rarely use Deep Research from any provider, opting for GPT 5.2 Pro and more specific instructions. In the first half of 2025 I almost exclusively used ChatGPT’s thinking models — Anthropic and Google have done good work to win back some of my attention.

* Claude 4.5 Opus: Chatting with Claude is where I go for basic code questions, visualizing simple data, and getting richer feedback on my work or decisions. Opus’s tone is particularly refreshing when trying to push the models a bit (in a way that GPT 4.5 used to provide for me, as I was a power user of that model in H1 2025). Claude Opus 4.5 isn’t particularly fast relative to a lot of models out there, but when you’re used to using the GPT Thinking models like me, it feels way faster (even with extended thinking always on, as I do) and sufficient for this type of work.

* Gemini 3 Pro: Gemini is for everything else — explaining concepts I know are well covered in the training data (and minor hallucinations are okay, e.g. my former Google rabbit holes), multimodality, and sometimes very long-context capabilities (but GPT 5.2 Thinking took a big step here, so it’s a bit closer). I still open and use the Gemini app regularly, but it’s a bit less locked-in than the other two.

Relative to ChatGPT, sometimes I feel like the search mode of Gemini is a bit off. It could be a product decision with how the information is presented to the user, but GPT’s thorough, repeated search over multiple sources instills a confidence I don’t get from Gemini for recent or research information.

* Grok 4: I use Grok ~monthly to try and find some piece of AI news or Alpha I recall from browsing X. Grok is likely underrated in terms of its intelligence (particularly Grok 4 was an impressive technical release), but it hasn’t had sticky product or differentiating features for me.

For images I’m using a mix of mostly Nano Banana Pro and sometimes GPT Image 1.5 when Gemini can’t quite get it.

For coding, I’m primarily using Claude Opus 4.5 in Claude Code, but still sometimes find myself needing OpenAI’s Codex or even multi-LLM setups like Amp. Over the holiday break, Claude Opus helped me update all the plots for The ATOM Project, which included substantial processing of our raw data from scraping HuggingFace, perform substantive edits for the RLHF Book (where I felt it was a quite good editor when provided with detailed instructions on what it should do), and other side projects and life organization tasks. I recently published a piece explaining my current obsession with Claude Opus 4.5, I recommend you read it if you haven’t had the chance:

A summary of this is that I pay for the best models and greatly value the marginal intelligence over speed — particularly because, for a lot of the tasks I do, I find that the models are just starting to be able to do them well. As these capabilities diffuse in 2026, speed will become more of a determining factor in model selection.

Peter Wildeford had a post on X with a nice graphic that reflected a very similar usage pattern:

Across all of these categories, it doesn’t feel like I could get away with just using one of these models without taking a substantial haircut in capabilities. This is a very strong endorsement for the notion of AI being jagged — i.e. with very strong capabilities spread out unevenly — while also being a bit of an unusual way to need to use a product. Each model is jagged in its own way. Through 2023, 2024, and the earlier days of modern AI, it quite often felt like there was always just one winning model and keeping up was easier. Today, it takes a lot of work and fiddling to make sure you’re not missing out on capabilities.

The working pattern that I’ve formed that most reinforces this using multiple models era is how often my problem with an AI model is solved by passing the same query to a peer model. Models get stuck, some can’t find bugs, some coding agents keep getting stuck on some weird, suboptimal approach, and so on. In these cases, it feels quite common to boot up a peer model or agent and get it to unblock project.

This multi-model approach or agent-switching happening occasionally would be what I’d expect, but with it happening regularly it means that the models are actually all quite close to being able to solve the tasks I’m throwing at them — they’re just not quite there. The intuition here is that if we view each task as having a probability of success, if said the probability was low for each model, switching would almost always fail. For switching to regularly solve the task, each model must have a fairly high probability of success.

For the time being, it seems like tasks at the frontier of AI capabilities will always keep this model-switching meta, but it’s a moving suite of capabilities. The things I need to switch on now will soon be solved by all the next-generation of models.

I’m very happy with the value I’m getting out of my hundreds of dollars of AI subscriptions, and you should likely consider doing the same if you work in a domain that sounds similar to mine.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

On the opposite side of the frontier models pushing to make current cutting edge tasks 100% reliable are open models pushing to undercut the price of frontier models. The coding plans on open models tend to cost 10X (or more) less than the frontier lab plans. It’s a boring take, but for the next few years I expect this gap to largely remain steady, where a lot of people get an insane value out of the cutting edge of models. It’ll take longer for the open model undercut to hit the frontier labs, even though from basic principles it looks like a precarious position for them to be in, in terms of costs of R&D and deployment. Open models haven’t been remotely close to Claude 4.5 Opus or GPT 5.2 Thinking in my use.

The other factor is that 2025 gave us all of Deep Research agents, code/CLI agents, search (and Pro) tool use models, and there will almost certainly be new form factors we end up using almost every day in released 2026. Historically, closed labs have been better at shipping new products into the world, but with better open models this should be more diffused, as good product capabilities are very diffuse across the tech ecosystem. To capitalize on this, you need to invest time (and money) trying all the cutting-edge AI tools you can get your hands on. Don’t be loyal to one provider.

There is an incredible amount of hype for Claude Code with Opus 4.5 across the web right now, which I for better or worse entirely agree with. Having used coding agents extensively for the past 6-9 months, where it felt like sometimes OpenAI’s Codex was the best and sometimes Claude, there was some meaningful jump over the last few weeks. The jump is well captured by this post, which called it the move of “software creation from an artisanal, craftsman activity to a true industrial process.” Translation: Software is becoming free and human design, specification, and entrepreneurship is the only limiting factor.

What is odd is that this latest Opus model was released on November 24, 2025, and the performance jump in Claude Code seemed to come at least weeks after its integration — I wouldn’t be surprised if a small product change unlocked massive real (or perceived) gains in performance.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

The joy and excitement I feel when using this latest model in Claude Code is so simple that it necessitates writing about it. It feels right in line with trying ChatGPT for the first time or realizing o3 could find any information I was looking for, but in an entirely new direction. This time, it is the commodification of building. I type and outputs are constructed directly. Claude’s perfect mix of light sycophancy, extreme productivity, and an elegantly crafted application has me coming up with things to do with Claude. I’d rather do my work if it fits the Claude form factor, and soon I’ll modify my approaches so that Claude will be able to help. In a near but obvious future I’ll just manage my Claudes from my phone at the coffee shop.

Where Claude is an excellent model, maybe the best, its product is where the magic happens for building with AI that instills confidence. We could see the interfaces the models are used in being so important to performance, such that Anthropic’s approach with Claude feels like Apple’s integration of hardware, software, and everything in between. This sort of magical experience is not one I expect to be only buildable by Anthropic — they’re just the first to get there.

The fact that Claude makes people want to go back to it is going to create new ways of working with these models and software engineering is going to look very different by the end of 2026. Right now Claude (and other models) can replicate the most-used software fairly easily. We’re in a weird spot where I’d guess they can add features to fairly complex applications like Slack, but there are a lot of hoops to jump through in landing the feature (including very understandable code quality standards within production code-bases), so the models are way easier to use when building from scratch than in production code-bases.

This dynamic amplifies the transition and power shift of software, where countless people who have never fully built something with code before can get more value out of it. It will rebalance the software and tech industry to favor small organizations and startups like Interconnects that have flexibility and can build from scratch in new repositories designed for AI agents. It’s an era to be first defined by bespoke software rather than a handful of mega-products used across the world. The list of what’s already commoditized is growing in scope and complexity fast — website frontends, mini applications on any platform, data analysis tools — all without having to know how to write code.

I expect mental barriers people have about Claude’s ability to handle complex codebases to come crashing down throughout the year, as more and more Claude-pilled engineers just tell their friends “skill issue.” With these coding agents all coming out last year, the labs are still learning how to best train models to be well-expressed in the form factor. It’ll be a defining story of 2026 as the commodification of software expands outside of the bubble of people deeply obsessed with AI.

There are things that Claude can’t do well and will take longer to solve, but these are more like corner cases and for most people immense value can be built around these blockers.

The other part that many people will miss is that Claude Code doesn’t need to be restricted to just software development — it can control your entire computer. People are starting to use it for managing their email, calendars, decision making, referencing their notes, and everything in between. The crucial aspect is that Claude is designed around the command line interface (CLI), which is an open door into the digital world.

The DGX Spark on my desk can be a mini AI research and development station managed by Claude.

This complete interface managing my entire internet life is the beginnings of current AI models feeling like they’re continually learning. Whenever Claude makes a mistake or does something that doesn’t match your taste, dump a reminder into CLAUDE.md, it’s as simple as that. To quote Doug OLaughlin, my brother in arms of Claude fandom, Claude with a 100X context window and 100X the speed will be AGI. By the end of 2026 we definitely could get the first 10X of both with the massive buildout of compute starting to become available.

Happy building.

Nathan sits down with Florian, our open model analyst to get spicy into debates of which labs won and lost momentum in open models of 2025. Reflection 70B, Huawei repackaging someone else's model as their own, the fall of Llama — no drama is left unturned. We also dig into the nuances that we didn't get to in our post, predict GPT-OSS 2, the American v. China balance at the end of 2026, and many other fun topics.

Enjoy & let us know if we should do more of this.

For the full year in review post, and to see our tier list, click here:

Watch on YouTube here:

It’s finally here! The public (and most complete) version of my talk covering every stage of the process to build Olmo 3 Think (slides are available). I’ve been giving this, improving it, and getting great feedback at other venues such as The Conference on Language Modeling (COLM) & The PyTorch Conference.This involves changes and new considerations of every angle of the stack, from pretraining, evaluation, and of course post-training.

Most of the talk focuses on reinforcement learning infrastructure and evaluating reasoning models, with quick comments on every training stage. I hope you enjoy it, and let us know what to improve in the future!

Chapters

* 00:00:00 Introduction

* 00:06:30 Pretraining Architecture

* 00:09:25 Midtraining Data

* 00:11:08 Long-context Necessity

* 00:13:04 Building SFT Data

* 00:20:05 Reasoning DPO Surprises

* 00:24:47 Scaling RL

* 00:41:05 Evaluation Overview

* 00:48:50 Evaluation Reflections

* 01:00:25 Conclusions

Here’s the YouTube link:

We present Olmo 3, our next family of fully open, leading language models. This family of 7B and 32B models represents:

* The best 32B base model.

* The best 7B Western-origin thinking & instruct models.

* The first 32B (or larger) fully open reasoning model.

This is a big milestone for Ai2 and the Olmo project. These aren’t huge models (more on that later), but it’s crucial for the viability of fully open-source models that they are competitive on performance – not just replications of models that came out 6 to 12 months ago. As always, all of our models come with full training data, code, intermediate checkpoints, training logs, and a detailed technical report. All are available today, with some more additions coming before the end of the year.

As with OLMo 2 32B at its release, OLMo 3 32B is the best open-source language model ever released. It’s an awesome privilege to get to provide these models to the broader community researching and understanding what is happening in AI today.

Paper: https://allenai.org/papers/olmo3 Artifacts: https://huggingface.co/collections/allenai/olmo-3Demo: https://playground.allenai.org/ Blog: https://allenai.org/blog/olmo3

Base models – a strong foundation

Pretraining’s demise is now regularly overstated. 2025 has marked a year where the entire industry rebuilt their training stack to focus on reasoning and agentic tasks, but some established base model sizes haven’t seen a new leading model since Qwen 2.5 in 2024. The Olmo 3 32B base model could be our most impactful artifact here, as Qwen3 did not release their 32B base model (likely for competitive reasons). We show that our 7B recipe competes with Qwen 3, and the 32B size enables a starting point for strong reasoning models or specialized agents. Our base model’s performance is in the same ballpark as Qwen 2.5, surpassing the likes of Stanford’s Marin and Gemma 3, but with pretraining data and code available, it should be more accessible to the community to learn how to finetune it (and be confident in our results).

We’re excited to see the community take Olmo 3 32B Base in many directions. 32B is a loved size for easy deployment on single 80GB+ memory GPUs and even on many laptops, like the MacBook I’m using to write this on.

A model flow – the lifecycle of creating a model

With these strong base models, we’ve created a variety of post-training checkpoints to showcase the many ways post-training can be done to suit different needs. We’re calling this a “Model Flow.” For post-training, we’re releasing Instruct versions – short, snappy, intelligent, and useful especially for synthetic data en masse (e.g. recent work by Datology on OLMo 2 Instruct), Think versions – thoughtful reasoners with the performance you expect from a leading thinking model on math, code, etc. and RL Zero versions – controlled experiments for researchers understanding how to build post-training recipes that start with large-scale RL on the base model.

The first two post-training recipes are distilled from a variety of leading, open and closed, language models. At the 32B and smaller scale, direct distillation with further preference finetuning and reinforcement learning with verifiable rewards (RLVR) is becoming an accessible and highly capable pipeline. Our post-training recipe follows our recent models: 1) create an excellent SFT set, 2) use direct preference optimization (DPO) as a highly iterable, cheap, and stable preference learning method despite its critics, and 3) finish up with scaled up RLVR. All of these stages confer meaningful improvements on the models’ final performance.

Instruct models – low latency workhorses

Instruct models today are often somewhat forgotten, but the likes of Llama 3.1 Instruct and smaller, concise models are some of the most adopted open models of all time. The instruct models we’re building are a major polishing and evolution of the Tülu 3 pipeline – you’ll see many similar datasets and methods, but with pretty much every datapoint or training code being refreshed. Olmo 3 Instruct should be a clear upgrade on Llama 3.1 8B, representing the best 7B scale model from a Western or American company. As scientists we don’t like to condition the quality of our work based on its geographic origins, but this is a very real consideration to many enterprises looking to open models as a solution for trusted AI deployments with sensitive data.

Building a thinking model

What people have most likely been waiting for are our thinking or reasoning models, both because every company needs to have a reasoning model in 2025, but also to clearly open the black box for the most recent evolution of language models. Olmo 3 Think, particularly the 32B, are flagship models of this release, where we considered what would be best for a reasoning model at every stage of training.

Extensive effort (ask me IRL about more war stories) went into every stage of the post-training of the Think models. We’re impressed by the magnitude of gains that can be achieved in each stage – neither SFT nor RL is all you need at these intermediate model scales.

First we built an extensive reasoning dataset for supervised finetuning (SFT), called Dolci-Think-SFT, building on very impactful open projects like OpenThoughts3, Nvidia’s Nemotron Post-training, Prime Intellect’s SYNETHIC-2, and many more open prompt sources we pulled forward from Tülu 3 / OLMo 2. Datasets like this are often some of our most impactful contributions (see the Tülu 3 dataset as an example in Thinking Machine’s Tinker :D – please add Dolci-Think-SFT too, and Olmo 3 while you’re at it, the architecture is very similar to Qwen which you have).

For DPO with reasoning, we converged on a very similar method as HuggingFace’s SmolLM 3 with Qwen3 32B as the chosen model and Qwen3 0.6B as the rejected. Our intuition is that the delta between the chosen and rejected samples is what the model learns from, rather than the overall quality of the chosen answer alone. These two models provide a very consistent delta, which provides way stronger gains than expected. Same goes for the Instruct model. It is likely that DPO is helping the model converge on more stable reasoning strategies and softening the post-SFT model, as seen by large gains even on frontier evaluations such as AIME.

Our DPO approach was an expansion of Geng, Scott, et al. “The delta learning hypothesis: Preference tuning on weak data can yield strong gains.” arXiv preprint arXiv:2507.06187 (2025). Many early open thinking models that were also distilled from larger, open-weight thinking models likely left a meaningful amount of performance on the table by not including this training stage.

Finally, we turn to the RL stage. Most of the effort here went into building effective infrastructure to be able to run stable experiments with the long-generations of larger language models. This was an incredible team effort to be a small part of, and reflects work ongoing at many labs right now. Most of the details are in the paper, but our details are a mixture of ideas that have been shown already like ServiceNow’s PipelineRL or algorithmic innovations like DAPO and Dr. GRPO. We have some new tricks too!

Some of the exciting contributions of our RL experiments are 1) what we call “active refilling” which is a way of keeping the generations from the learner nodes constantly flowing until there’s a full batch of completions with nonzero gradients (from equal advantages) – a major advantage of our asynchronous RL approach; and 2) cleaning, documenting, decontaminating, mixing, and proving out the large swaths of work done by the community over the last months in open RLVR research.

The result is an excellent model that we’re very proud of. It has very strong reasoning benchmarks (AIME, GPQA, etc.) while also being stable, quirky, and fun in chat with excellent instruction following. The 32B range is largely devoid of non-Qwen competition. The scores for both of our Thinkers get within 1-2 points overall with their respective Qwen3 8/32B models – we’re proud of this!

A very strong 7B scale, Western thinking model is Nvidia’s NVIDIA-Nemotron-Nano-9B-v2 hybrid model. It came out months ago and is worth a shot if you haven’t tried it. I personally suspect it may be due to the hybrid architecture making subtle implementation bugs in popular libraries, but who knows.

All in, the Olmo 3 Think recipe gives us a lot of excitement for new things to try in 2026.

RL Zero

DeepSeek R1 showed us a way to new post-training recipes for frontier models, starting with RL on the base model rather than a big SFT stage (yes, I know about cold-start SFT and so on, but that’s an implementation detail). We used RL on base models as a core feedback cycle when developing the model, such as during intermediate midtraining data mixing. This is viewed now as a fundamental, largely innate, capability of the base-model.

To facilitate further research on RL Zero, we released 4 datasets and series of checkpoints, showing per-domain RL Zero performance on our 7B model for data mixes that focus on math, code, instruction following, and all of them together.

In particular, we’re excited about the future of RL Zero research on Olmo 3 precisely because everything is open. Researchers can study the interaction between the reasoning traces we include at midtraining and the downstream model behavior (qualitative and quantitative).

This helps answer questions that have plagued RLVR results on Qwen models, hinting at forms of data contamination particularly on math and reasoning benchmarks (see Shao, Rulin, et al. “Spurious rewards: Rethinking training signals in rlvr.” arXiv preprint arXiv:2506.10947 (2025). or Wu, Mingqi, et al. “Reasoning or memorization? unreliable results of reinforcement learning due to data contamination.” arXiv preprint arXiv:2507.10532 (2025).)

What’s next

This is the biggest project we’ve ever taken on at Ai2, with 60+ authors and numerous other support staff.

In building and observing “thinking” and “instruct” models coming today, it is clear to us that there’s a very wide variety of models that fall into both of these buckets. The way we view it is that thinking and instruct characteristics are on a spectrum, as measured by the number of tokens used per evaluation task. In the future we’re excited to view this thinking budget as a trade-off, and build models that serve different use-cases based on latency/throughput needs.

As for a list of next models or things we’ll build, we can give you a list of things you’d expect from a (becoming) frontier lab: MoEs, better character training, Pareto efficient instruct vs think, scale, specialized models we actually use at Ai2 internally, and all the normal things.

This is one small step towards what I see as a success for my ATOM Project.

We thank you for all your support of our work at Ai2. We have a lot of work to do. We’re going to be hunting for top talent at NeurIPS to help us scale up our Olmo team in 2026.

First, on the topic of writing, the polished, and more importantly printed, version of my RLHF Book is available for pre-order. It’s 50% off for a limited time, you can pre-order it here!

Like a lot of writing, I’ve been sitting on this piece for many months thinking it’s not contributing enough, but the topic keeps coming up — most recently via Jasmine Sun — and people seem to like it, so I hope you do too!

It’s no longer a new experience to be struck by just how bad AI models are at writing good prose. They can pull out a great sentence every now and then, particularly models like GPT-5 Pro and other large models, but it’s always a quick comment and never many sustained successive sentences. More importantly, good AI writing feels like a lucky find rather than the result of the right incantation. After spending a long time working training these models, I’m fairly convinced that this writing inhibition is a structural limitation to how we train these models today and the markets they’re designed to serve.

If we're making AIs that are soon to be superhuman at most knowledge work, that are trained primarily to predict text tokens, why is their ability to create high quality text tokens still so low? Why can’t we make the general ChatGPT experience so much more refined and useful for writers while we’re unlocking entirely new ways of working with them every few months — most recently the CLI agents like Claude Code. This gap is one of my favorite discussions of AI because it’s really about the definition of good writing is in itself.

Where language models can generate beautiful images from random noise, they can't reliably generate a good few sentences from a couple bullet points of information. What is different about the art form of writing than what AI can already capture?

I'm coming to believe that we could train a language model to be a great writer, but it goes against so many of the existing training processes. To list a few problems at different stages of the stack of varying severity in terms of their handicapping of writing:

* Style isn’t a leading training objective. Language models all go through preference training where many aspects from helpfulness, clarity, honesty, etc. are balanced against each other. Many rewards make any one reward, such as style, have a harder time standing out. Style and writing quality is also far harder to measure, so it is less likely to be optimized vis-a-vis other signals (such as sycophancy, which was easier to capture).

* Aggregate preferences suppress quirks. Language model providers design models with a few intended personalities, largely due to the benefits of predictability. These providers are optimizing many metrics for "the average user." Many users will disagree on what their preference for “good writing” is.

* Good writing’s inherent friction. Good writing often takes much longer to process, even when you’re interested in it. Most users of ChatGPT just want to parse the information quickly. Doubly, the people creating the training data for these models are often paid per instance, so an answer with more complexity and richness would often be suppressed by subtle financial biases to move on.

* Writing well is orthogonal to training biases. Throughout many stages of the post-training process, modern RLHF training exploits subtle signals for sycophancy and length-bias that aren't underlying goals of it. These implicit biases go against the gradient for better writing. Good writing is pretty much never verbose.

* Forced neutrality of a language model. Language models are trained to be neutral on a variety of sensitive topics and to not express strong opinions in general. The best writing unabashedly shares a clear opinion. Yes, I’d expect wackier models like Grok to potentially produce better writing, even if I don’t agree with it. This leads directly to a conflict directly in something I value in writing — voice.

All of these create models that are appealing to broad audiences. What we need to create a language model that can write wonderfully is to give it a strong personality, and potentially a strong "sense of self" — if that actually impacts a language model's thinking.

The cultivation of voice is one of my biggest recommendations to people trying to get better at writing, only after telling them to find something they want to learn about. Voice is core to how I describe my writing process.

When I think about how I write, the best writing relies on voice. Voice is where you process information into a unique representation — this is often what makes information compelling.

Many people have posited that base models make great writers, such as when I discussed poetry with Andrew Carr on his Interconnects appearance, but this is because base models haven’t been squashed to the narrower style of post-trained responses.

I’ve personally been thinking about this sort of style induced by post-training recently as we prepare for our next Olmo release, and many of us think the models with lower evaluation scores on the likes of AlpacaEval or LMArena actually fit our needs better. The accepted style of chatty models today, whether it’s GPT-5, DeepSeek R1, or a large Qwen model, is a bit cringe for my likes. This style is almost entirely applied during post-training.

Taking a step back, this means base models show us that there can be great writing out of the models, but it’s still far from reliable. Base models aren't robust enough to variations to make great writers — we need some form of the constraints applied in post-training to make models follow Q&A. The next step would be solving the problem of how models aren’t trained with a narrow enough experience. Specific points of view nurture voice. The target should be a model that can output tokens in any area or request that is clear, compelling, and entertaining.

We need to shape these base models with post-training designed for writing, just as the best writers bend facts to create narrative.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

Some models makers care a bit about this. When a new model drops and people rave about its creative writing ability, such as MoonShot AI’s Kimi K2 line of model, I do think the team put careful work into the data or training pipelines. The problem is that no model provider is remotely ready to sacrifice core abilities of the model such as math and coding in pursuit of meaningfully better writing models.

There are no market incentives to create this model — all the money in AI is elsewhere, and writing isn’t a particularly lucrative market to disrupt. An example is GPT 4.5, which was to all reports a rather light fine-tune, but one that produced slightly better prose. It was shut down almost immediately after its launch because it was too slow and economically unviable with its large size.

If we follow the voice direction, the model that is likely to be the best writer relative to its overall intelligence was the original revamped Bing (aka Sydney) model that went crazy in front of many users and was rapidly shut down. That model had THOUGHTS it wanted to share. That’s a starting point, but a scary one to untap again. This sort of training goes far beyond a system prompt or a light finetune, and it will need to be a new post-training process from start to end (more than just a light brush of character training).

We need to be bold enough to create models with personality if we want writing to fall out. We need models that speak their views loudly and confidently. These also will make more interesting intellectual companions, a niche that Claude fills for some people, but I struggle with Claude plenty of times due to its hesitance, hedging, or preferred answer format.

For the near future, the writing handicap of large language models is here to stay. Good writing you have to sit in to appreciate, and ChatGPT and the leading AI products are not optimized for this whatsoever. Especially with agentic applications being the next frontier, most of the text written by the models will never even be read by a human. Good writing is legitimately worse for most of the use cases I use AI for. I don’t like the style per se, but having it jump to be a literary masterpiece would actually be worse.

I don’t really have a solution to AI’s writing problem, but rather expensive experiments people can try. At some point I expect someone to commission a project to push this to its limits, building a model just for writing. This’ll take some time but is not untenable nor unfathomably expensive — it’ll just be a complete refresh of a modern post-training stack.

Even if this project was invested in, I don’t expect the models to be close to the best humans at elegant writing within a few years. Our current batch of models as a starting point are too far from the goal. With longer timelines, it doesn’t feel like writing is a fundamental problem that can’t be solved.

This is the first of a handful of interviews I’m doing with teams building the best open language models of the world. In 2025, the open model ecosystem has changed incredibly. It’s more populated, far more dominated by Chinese companies, and growing. DeepSeek R1 shocked the world and now there are a handful of teams in China training exceptional models. The Ling models, from InclusionAI — Ant Group’s leading AI lab — have been one of the Chinese labs from the second half of the year that are releasing fantastic models at a rapid clip.

This interview is primarily with Richard Bian, who’s official title is Product & Growth Lead, Ant Ling & InclusionAI (on LinkedIn, X), previously leading AntOSS (Ant Group’s open source software division). Richard spent a substantial portion of his career working in the United States, with time at Square, Microsoft, and an MBA from Berkeley Haas, before returning to China and work at Ant.

Also joining are two leads of the Ant Ling technical team, Chen Liang (Algorithm Engineer), and Ziqi Liu (Research Lead).

This interview focuses on many topics of the open language models, such as:

* Why is the Ant Group — known for the popular fintech app AliPay — investing so much in catching up to the frontier of AI?

* What does it take to rapidly gain the ability to train excellent models?

* What decisions does one make when deciding a modeling strategy? Text-only or multimodal? What size of models?…

* How does the Chinese AI ecosystem prioritize different directions than the West?

And many more topics. Listen on Apple Podcasts, Spotify, YouTube, and where ever you get your podcasts. For other Interconnects interviews, go here.

Some more references & links:

* InclusionAI’s homepage, highlighting their mission.

* AntLingAGI on X (models, research, etc.), InclusionAI on X (overall initiative), InclusionAI GitHub, or their Discord community.

* Ling 1T was highlighted in “Our Picks” for our last open model roundup in October.

* Another interview with Richard at State of Open Conference 2025.

* Over the last few months, our coverage of the Chinese ecosystem has taken off, such as our initial ranking of 19 open Chinese AI labs (before a lot of the models we discuss below), model roundups, and tracking the trajectory of China’s ecosystem.

An overview of Ant Ling & Inclusion AI

As important context for the interview, we wanted to present an overview of InclusionAI, Ant’s models, and other efforts that emerged onto the scene just in the last 6-9 months. To start — branding.

Here’s a few screenshots of InclusionAI’s new website. It starts with fairly standard “open-source AI lab messaging.”

Then I was struct by the very distinct messaging which is surprisingly rare in the intense geopolitical era of AI — saying AI is shared for humanity.

I expect a lot of very useful and practical messaging from Chinese open-source labs. They realize that Western companies likely won’t pay for their services, so having open models is their only open door to meaningful adoption and influence.

Main models (Ling, Ring, & Ming)

The main model series is the Ling series, their reasoning models are called Ring, and their Multimodal versions are called Ming. The first public release was Ling Plus, 293B sparse MoE in April. They released the paper for their reasoning model in June and have continued to build on their MoE-first approach.

Since then, the pace has picked up significantly. Ling 1.5 came in July.

Ling (and Ring) 2.0 came in September of this year, with a 16B total, 2B active mini model, an 100B total, 6B active flash model, and a big 1T total parameter 50B active primary model. This 1T model was accompanied by a substantial tech report on the challenges of scaling RL to frontier scale models. The rapid pace that Chinese companies have built this knowledge (and shared it clearly) is impressive and worth considering what it means for the future.

Eval scores obviously aren’t everything, but they’re the first step to building meaningful adoption. Otherwise, you can also check out their linear attention model (paper, similar to Qwen-Next), some intermediate training checkpoints, or multimodal models.

Experiments, software, & other

InclusionAI has a lot of projects going in the open source space. Here are some more highlights:

* Language diffusion models: MoEs, sizes similar to Ling 2.0 mini and flash (so they likely used those as base). Previous versions exist.

* Agent-based models/fine-tunes, Deep Research models, computer-use agentic models.

* GroveMoE, MoE arch experiments.

* RL infra demonstrations (Interestingly, those are dense models)

* AWorld: Training + general framework for agents (RL version, paper)

* AReal: RL training suite

Interconnects is a reader-supported publication. Consider becoming a subscriber.

Chapters

* 00:00:00 A frontier lab contender in 8 months

* 00:07:51 Defining AGI with metaphor

* 00:20:16 How the lab was born

* 00:23:30 Pre-training paradigms

* 00:40:25 Post training at Inclusion

* 00:48:15 The Chinese model landscape

* 00:53:59 Gaps in the open source ecosystem today

* 00:59:47 Why China is winning the open race

* 01:11:12 A metaphor for our moment in LLMs

Transcript

A frontier lab contender in 8 months

Nathan Lambert (00:05)

Hey everybody. I’m excited to start a bit of a new series when I’m talking to a lot more people who are building open models. Historically, I’ve obviously talked to people I work with, but there’s a lot of news that has happened in 2025 and I’m excited to be with one of the teams, a mix of product, which is Richard Bian and some technical members from the Ant Ling team as well, which is Chen Liang and Ziqi Liu. But really this is going to be a podcast where we talk about how you’re all building models, why you do this. It’ll talk about different perspectives between US, China and a lot of us going towards a similar goal. I was connected first with Richard, who’s also talked to other people that helped with Interconnects. So we can start there and go through and just kind of talk about what you do. And we’ll roll through the story of building models and why we do this.

Richard Bian (01:07)

Hi. Again, thanks so much, Nathan. Thanks so much for having us. My name is Richard Bian. I’m currently leading the product and growth team of Ant Ling, which is part of the Inclusion AI lab of Ant Group. So Ant Group is the parent company of Alipay, which might be a product which many, many more people know about. But the group has been there for quite some time. It used to be a part of Alibaba, but now it’s a separate company since 2020. I actually have a pretty mixed background. Before I joined the Ling team, I’ve been doing Ant open source for four years. In fact, I built Ant open source from a technical strategy, which is basically a one-liner from our current CTO all the way into a full-fledged multifunctional team of eight people in four years. So it has been a pretty rewarding journey. And before that, my last life, I’ve been spending 11 years in the States working as a software engineer with Microsoft and with Square. Again, it was a pretty rewarding past. I returned back to China during COVID to be close with my family. It was a conscious decision. So far so good. It has been a pretty rewarding journey. And I really love how Nathan you name your column as Interconnects and you actually echoed when you just began the conversation just now. I found that to be a very noble initiative. So very honored to be here.

Nathan Lambert (02:48)

Hopefully first of many, but I think you all have been doing very interesting stuff in the last few weeks, or last few months, so it’s very warranted. And do you two want to introduce yourselves as well?

Chen Liang (02:58)

Me first. My name is Chen Liang and I’m the algorithm engineer of Ling Team, and I’m mainly responsible for the floating point 8 training during the pre-training. Thank you.

Ziqi Liu (03:16)

My name is Ziqi Liu and I graduated, a PhD from Jiao Tong University in China. And I’ve been working at Ant Group for about eight years. And currently I’m working on the Ling language model. That’s it.

Nathan Lambert (03:45)

Nice. I think the way this will flow is I’m going to probably transition. It’ll start more with Richard’s direction. Then as we go, it’ll get more technical. And please jump in. I think that we don’t want to segment this. I mean, the border between product growth, technical modeling, whatever, that’s why AI is fun is because it’s small. But I would like to know how Inclusion AI started and all these initiatives. I don’t know if there’s a link to Ant OSS. I found that in prep and I thought that was pretty interesting and just kind of like, how does the birth of a new language modeling lab go from idea to releasing one trillion parameter models? So like, what does that feel like on the ground?

Richard Bian (04:18)

There’s actually one additional suffix for that in eight months’ time. In fact, we kind of began all of this initiative in February this year. So just to begin with for the audience who probably didn’t know much about Inclusion AI, Inclusion AI basically envisions AGI as a humanity’s shared milestone, not a privileged asset. So we started this initiative back in the February of 2025, inspired by the DeepSeek Research Lab. So the DeepSeek Research Lab and their publication, in fact, motivated a lot of people. I believe not only in China, but globally. Taking one step more closer to the AGI initiative by showing it’s probably not an exclusive game for only the richest people who can afford the best hardware and the best talent. So the way we’re kind of looking at it is like why we named that Inclusion is because we actually have that gene with the company. So the decision was actually made, of course, the decision was made beyond my pay grade, but it was actually very well informed internally for the mission and vision that we want to be more like DeepSeek, which is a research lab with a dedicated effort of pursuing AGI. In fact, I mean, if you kind of think about Ant Group with our business model, like we’re a Fintech company, to some extent, very similar to a combination of Square, Stripe, and many other companies in the States, we have a very broad range of businesses which focus not only on the financial vertical, but on medical insurances and the technical services as well. So a lot of those businesses. In order for us to actually be able to support those businesses, I would say long-term success in the next five to 10 years is going to be critically important for us to be able to really focus on the fundamentals of AI. And we feel that the language model is a key to that door. We cannot give up on that initiative.

Nathan Lambert (06:52)

There’s a lot here and I agree with this. And I think that it’s like, the Ant Group is a big large tech company. And I think large tech companies being able to train AI as like most of the audience here is going to be like, yes, they definitely should be doing this. It’s a transformative technology. I think the two things to double click on are, we’re going to have to define like what you think of as AGI and why you’re pursuing this. Because it has to go deeper than like a term that we are doing. I know like DeepSeek is very ideological in their pursuit of intelligence. So I think it’s good to do that. And then I will also double click on the question of like, why open models and like, because DeepSeek is doing like open and as strong as they can, they’re text only. We’ll talk about this later. But it’s like, let’s do each of these individually to kind of ground the motivation.

Defining AGI with metaphor

Richard Bian (07:51)

Sure. I guess, I mean, for AGI, the way we are looking at it is like, I don’t think there’s a definitive answer to that. I mean, if we kind of search Google or any other search engines, it will give you a line, which means something. But it doesn’t mean anything, honestly, to me personally, just by looking at the definition. I would probably use a metaphor. People are probably very familiar with the navigation era. It’s a glorious navigation era back in the 1400s. Now, I think it feels more like all the ships are just leaving Lisbon last year, or maybe like two years ago.

Nathan Lambert (08:18)

I like it. I agree with this more than most of the definitions, because a lot of the definitions are grounded in like work or something.

Richard Bian (08:26)

The one I’m kind of looking at is like, all the ships are leaving Lisbon. Some of them are heading west, knowing for a fact that, hey, India is over there. But now we all know the truth that India is on the east side. But it doesn’t matter. It’s the whole American continent. So the way I’m kind of looking into the definition of AGI right now is like, I personally have a very firm belief that human intelligence and machine intelligence, to some extent, have their similarities. Humans are trying to, to some extent, explore the limit of human intelligence with the help from the machines. So when everything was beginning, we were kind of using all of this as a co-pilot mode. But moving forward, there are all of these theories indicating that there might be an intrinsic point that the machine intelligence, it goes all the way back from the tooling time. They believe that machine intelligence might, at one point, exceed human intelligence. So I guess we’re looking to that pivoting point. Before we reach there, honestly, I don’t know where we’re going and how long we can go towards that particular direction. But clearly, there are some common consensus right now, including maybe MoE (Mixture of Experts) as architecture, including the pre-training, even to some extent, we’re seeing a diminishing return. But pre-training is still pretty important. And reinforcement learning, to some extent, is probably another general agreement that this might not be wrong. We don’t know if this is right, but it might not be wrong. So there are all of these exploratory directions that we believe in. So we’re just kind of sailing there and see how that goes.

Nathan Lambert (10:20)

I love this. And I think the crucial question is for Chen or Ziqi is like, the team like, how do you build team alignment around this? Is this something that you feel like you walk into the office or get on a call and everybody’s in agreement? Or is this like a vision that you’re still building or trying to sell? Like, to what extent you could say, because I think there’s a big difference between like, I buy the vision for Inclusion AI, but it’s like, how real is this when you’re across the org?

Richard Bian (10:49)

I can maybe share my feeling and Ziqi and Chen can chime in. Of course, at the very beginning, there’s skepticism. It’s by human nature, right? So the way we’re looking at it is like, I think DeepSeek gives a very clear indication that this might be working. There has been this hazy, chaotic era of 2024, which nobody has the tools to navigate. So people are very cautious about sailing. You see ships going out and came back crippled, and you begin to worry about what’s going on there.

Nathan Lambert (11:34)

I think there’s a big difference between the US because I think in the US everybody was bought in. And I’ve talked to a few more labs in China and it’s like there’s so much emotional energy on the DeepSeek moment in China that I think in the US people forget about it where it’s like, I could see this in the sequence of releases as well because it’s like everybody had a few months after DeepSeek like all these labs in China have started releasing models and I just think that it’s good to have you say this, is a shared sense of people so people can internalize like how much has been mobilized. And that’s kind of a culturally salient point.

Richard Bian (12:04)

It’s motivating. To some extent, there was this very famous navigator called Zheng He back in the Ming dynasty. So I think basically when Zheng He was able to pretty much pull through the trip all the way to India from China, people began realizing that, hey, not only the Portuguese can do this kind of long journey sailing, the Chinese can do that too. And we’re exploring different parts of the map. You know, toward the end of the day, nobody knows the whole picture. So the way I’m kind of looking at it is like, first, I’m very bought into the mission to some extent that it kind of feels like, you know, even though we begin sailing late, but we do have our own kind of taste to this game. So we will be able to contribute. And you did ask about the question, you know, like why we chose to be open, right? To some extent, I cannot really believe that open is a choice, just like how the leaders in this game are not the most open player in the game, right? But if you’re kind of thinking about playing poker, the trick leader has their own strategy, which is all understandable. For us, because we’re joining the game at this stage, I guess the best strategy would kind of feel like, A, really trying to follow suit to the right direction to minimize the mistakes we’re making at this moment because we’re so late. Second, stay open and stay polished. So keep a very open mind about what’s going on in the surroundings. And that’s probably the best we can do. That’s my two cents.

Nathan Lambert (13:51)

To provide some color and I’ll have a whole note in the page that I release with this for people listening. The first Ling model, which is like their text only model, very, you could see iterations from DeepSeek and the architecture was in April and then a big updated Ling 1.5 in July. And then in September or recently was Ling 2.0, which also came with a multimodal Ming and a reasoning Ring model. And I think like by this September release is when like me and a couple of people that work at Interconnects were like, Holy crap, like this is a, this is like very much a real deal model. And to kind of ramp in that period of time is not easy. Like there’s a lot of companies in the US that are trying to do this right now. A few companies in China have shown that they can do this. And it’s like, I guess if you want to explain this kind of Ling, Ring, Ming series of models and like if this is a clear strategy behind this or if this is what works like, how did you evolve through the first models through the summer to today to kind of get to this point?

Richard Bian (14:56)

Sure. So I mean, first and foremost, I think the foundation model is really important. To some extent, I’ve been working with many people on the system side, because Ant Group has a very solid cloud-native infrastructure team. So the team has been, when we talk about this, we’re kind of beginning using the metaphor. The model is really like an operating system. It’s not like the operating system itself, but it’s more like the kernel. Right, so only a few people can actually write kernel code, even nowadays. Just like how there’s the most talented people who can actually work on the model team right now. We feel that it’s not only a key leading to the technical future, but it’s also a key leading to the user experience future. Because we do see the, I personally believe in the trend of technology brings in new interactions which will lead to new product, which will lead to new business models, which will lead to potentially new organization structure, rinse and repeat. So we kind of like really choose to do the fundamental model of the Ling series because of that. And the Ring series is an obvious next, given the relationship between V3 and R1. It definitely indicates about how we can potentially take a very polished, well, actually, a very intelligent individual, unpolished, and put some reinforcement learning on it to make it a much better individual in one clear vertical direction. We’re going to be touching on some of those kind of technical aspects in our conversation next. But that has been a very clear direction.

Nathan Lambert (16:48)

Do you see this evolving with kind of feedback from within Ant Group, which is like, you’ve also released this diffusion language model. A diffusion language model is very interesting. I’m going to just go out on a little bit of a side rant because I’ve heard, I was talking to people about these and it’s like very hit or miss with me, whether or not I think they’re going to be big. Because we see that tool use and reasoning is a big thing. So the whole idea of a diffusion language model is you generate a very long sequence at once and that could save on costs because you don’t have this kind of quadratic memory increase and you do very long sequences. So I saw that I was optimistic. And then you see the idea of tool use, which is like, you have to be able to chop up the reasoning. And I was like, I’m really bearish on diffusion models for language again, because you have to be able to search and execute code. But then I was hearing that in like user facing products, like code diffs, where if you’re generating a website and you did take a prompt and go to a huge diff on a code base really fast, then language diffusion is actually really nice. And the motivation of the question is like, do you have this feedback loop in your modeling where Ant Group is trying to use these things for products and might like have a bit of a feedback of like this latency isn’t fast enough or like this area you need to move it to, or is this kind of like a separate play of just build the best models you can and figure it out later?

Richard Bian (18:12)

That’s a very perfect question. We use this metaphor that we’re probably also doing this reinforcement learning in real life by trial and error. Almost kind of feels like, so I think Nathan, you nailed a very good question. And there are some very clear consensus about coding agents, tool use and people kind of going down a path and pursuing their own business models and begin making revenues. So that’s one type of usage patterns for language models. We do that and we see some very clear, I would say feedback loops in that direction. So that’s one pillar. And the second pillar is about the not so clear aspect. By saying the not so clear aspect, it’s like, I believe everyone in the Silicon Valley and in Seattle is still scratching their heads trying to understand about, hey, when can I break even with all this investment? Are we really generating enough user values kind of back to, I’m a product person. So all of those kinds of words keep coming back into my head. And, you know, at this moment, consciously speaking, it’s very hard to come to the conclusion that, you know, all of this is valuable enough for the end user. But, you know, we’re trying to explore the directions for that. I would say a lot of the, you know, generating the whole website, you know, what Labo did, it’s an interesting form of product. But at this moment, we don’t know if it’s A, sustainable as a business model, B, if this is the best type of product we can offer to the user. So all of those are iterative. Within company, we do have some of those explorative products that use our models, not only the Ring model, but Ming as well, like the multimodal. And you mentioned about the, so that’s the second pillar. And the latter is more like the last pillar, because Ant Group does have a research institution called Ant Research. So the model is a joint collaboration between the research and the Ling Team.

How the lab was born

Nathan Lambert (20:16)

I guess there’s another like org chart question, which is like, where in the structure of the big tech company that is Ant did this Inclusion AI slash Ling and all of this grow? Like, is this within cloud that there’s a new modeling or research org or is it kind of separate? Like, do you feel like this is a part of the bigger company or are you kind of insulated from this?

Richard Bian (20:42)

You can actually search on Google and find information about Ant Research which is a joint research lab focusing more on a lot of these frontier technologies like graph, deep learning, reinforcement learning, before all of this. So that’s the background of Ant Research. And second, when we begin forming the AGI initiative of Inclusion AI, we begin getting very serious. So we begin putting all of these resources together to some extent physically, but more from the organizational ways of saying that all of these teams of financial models and research lab institution and the user experience expert focusing on exploratively looking into the next big application that people will actually use. So all of this, we kind of began forming this internal, I wouldn’t call that organization, but more like this internal initiative directly driven by our CTO. So it’s very serious effort. It’s very serious to the extent that, you know, it feels more like when the team actually formed the original DeepSeek initiative. So all of these people, you do nothing else but only focusing on this and this is the only important thing for this.

Nathan Lambert (22:01)

It’s like so much of this is that the mystique I feel like is that in the West, we don’t get what would normally be gossip of what is happening in the Chinese tech ecosystem, which I don’t think this is hard to see if you have friends that work at Ant Group, because it’s probably you’re moving hundreds of people’s jobs around and people talk. Whereas like in my circles, it’s like, Meta is doing another reorg. And then you hear about it in the news a few days later. So it’s just like, I don’t know. That’s my reflection hearing all of this. And I’m mostly learning that all of these orgs end up similar in size. And then you have to prioritize resources per researcher and all of these normal things. I’m going to start transitioning into this section we had prepped on actual modeling things, which is mostly on pre-training, which is fun. I think that state of affairs on my pre-training knowledge from AI2 is that we’ve scaled, done plenty of dense models and some architecture things from up to like 32B, some experiments at 70B that one didn’t work out. MoE is work in progress. So I’m personally very interested in architectural decisions that enable MoEs and long context.

Pre-training paradigms

I think the kind of basic thing is just like, if you’re pre-training, I mean, this is for Ziqi is like, what does your, how do you feel like your trajectory is as a researcher as you’re going through these months? This could be just like, what does your work feel like when you’re trying to boot up like a DeepSeek style, very ambitious lab building new infrastructure and getting models off the ground. And then we’ll kind of go into some more specific discussions around like Ling 1T later and stuff like this. But it’s like, how is building this?

Ziqi Liu (23:45)

Our architecture indeed refers to OpenAI’s scaling law or DeepSeek’s scaling law. They really do a good job. In our Ling scaling law, the non-embedding training FLOPs play the central role of our scaling law. So we set up our own framework that provides foundation for a standardized experimental pipeline. So there are many questions when we start conducting scaling law under the MoE architecture. So the first question is, can we find simple rules for finding optimal hyperparameters with respect to training FLOPs, which are not sensitive to the structure of MoE. Similar to DeepSeek, we first discovered the optimal critical hyperparameters with respect to training FLOPs and the MoE architecture. We find those optimal hyperparameters are not that sensitive to the structure of MoE, like the activation ratio and something others in a mild condition, but more related to the training FLOPs. So this is our first finding. And then we found activation ratio is critical and can consistently improve if we reduce activation ratio.

Nathan Lambert (25:14)

Can you say more about this? I mean, most of pre-training is a lot of different things, which you’re accumulating FLOP efficiency while getting model performance. And then it’s like Chen, you also were saying you focused on FP8 stability, FP8 and training stability in general. So I’m kind of curious of like any major, like, what is your biggest impressions of focusing on kind of this narrow thing in pre-training, which is getting more memory by using lower precision while maintaining stability. So if you have any like high level takes on pre-training stability at that precision, then I’ll zoom into more specific questions on scaling up from there.

Chen Liang (26:00)

At first we heard about the floating point 8 from DeepSeek. They used floating point 8 training through the training of DeepSeek. And we also tried the recipe of them, the block-wise INT8 in the Megatron. And we find that actually the MFU (Model FLOPs Utilization) is not very high. And sometimes it’s even slower than the BF16 (bfloat16) training. And we find that the main costs are the quantization and dequantization. So actually, the floating point 8 is not as fast as they claimed, actually. And we profile the whole training data and try to minimize the quantization and dequantization process.

Nathan Lambert (26:50)

What is getting quantized and dequantized?

Chen Liang (26:53)

If you want to try the floating point 8 training, it’s actually due to GEMM (General Matrix Multiply) in the linear layers. And you want to quantize the weights and the inputs to FP8 (E4M3) type. But the other structure, they compute in the BF16, BFloat16 type. So when you get into the linear layer, you need to quantize it to the floating point 8, and then do the GEMM. And the GEMM output is the BFloat16. So this is the way you need to quantize and dequantize to adapt the other structure.

Nathan Lambert (27:43)

And then what does your work actually look like in getting this? So you find it to be not as fast. Like, what do you actually do to change this?

Chen Liang (27:50)

In the MoE layer, it’s got the FC1 (Fully Connected 1) and FC2 (Fully Connected 2), right? And in the middle of them, they’ve got the switch gated function. So FC1, switch gated function and FC2. And the output of FC1 is the BFloat16. And we fuse the operation of the switch gated function and the quantization function. So we fuse them, the two operations, into one. And so it saves some time. And the MoE layer is a batched operation. So you need to actually do the activation function on all the experts. So that’s a lot of time.

Nathan Lambert (28:52)

For people listening, FC is fully connected, which is just the standard neural network layer. So I might be being silly, but generally the idea with MoEs is that you have the feed forward layers, take up the most parameters and you get more efficient by adding MoEs. And within the MoE, kind of gated to each expert, is it actually standard that it’s like fully connected, MoE gate, fully connected? And it’s kind of alternating because I know this normally like attention block, MoE block is like the higher level of abstraction. And it’s this fully connected, MoE gating and then fully connected, is that actually industry standard? And I just had like a lapse in my brain.

Chen Liang (29:37)

This structure is conventional actually. Some experiments have explained that the switch gated can make your gradient stable during training. So it’s actually a standard architecture.

Nathan Lambert (29:51)

When you’re actually experimenting on this, is this the sort of thing that when you’re doing it at your like first models were about 300B total and you had smaller models? Like, is this a sort of thing done where you get this performance at every scale? Or do you have to revisit this when you’re doing something like Ling 1T, which is this latest model with way more parameters? Because I think the root of my question is like, are the numerical problems you get from scaling like whack-a-mole, where it’s like an old problem that you fixed becomes a problem again? Or is it an entirely new type of thing that comes up when you’re going to big models?

Chen Liang (30:26)

We do the experiment on the size of 100 billion parameters first. Also the situation can be, we can learn from the situation. That size, not just the 1T.

Nathan Lambert (30:43)

And I remember reading, I saw that you guys did QK norm for this as well. Is this just like, you also found this to be standard and work for you because we’ve had some issues with long context and doing QK norm kind of hurting performance there. We still have some ablations to track down.

Chen Liang (30:47)

We actually do the experiment of the QK norm on BFloat16 and the result comes out. The loss is better than if you didn’t apply the QK norm. And actually the one big thing is that when you do the floating point 8 training, if you do not apply QK norm before the rotary embedding, the gradient of the linear QKV may be underflow. Most of the time, it’s underflow because without the QK norm. So if we want to apply the floating point 8 training, you need to add the QK norm to avoid the quantization error. Since the quantization error is propagated from the last layer to the first, and if the last layer got more quantization error, until the first layer it’s amplified error.

Nathan Lambert (32:07)

Let me try to talk through this because I’m mostly working post-training and I’ve heard all these terms and I want to make sure that we’re presenting a fairly clear picture to people. So in attention, you have queries, keys, and values. And these are big matrices that store many different things. And like generally with pre-training, the magnitude of the variables matters a lot because what you’re saying about like gradient flow. And if you have variables that are like too small, you might have no signal and too big or one thing. And what we’re saying is that, God, I guess what’s the order between, when you have, I guess there’s complicated things, which is like where the rotary embeddings are applied relative to the attention computation. And what we’re saying is that you have to put QK norm ahead of the rotary embeddings in this attention module, because then otherwise your gradients are too small when you’re scaling this or with FP8.

Chen Liang (32:53)

During the forward process, you got the QK norm and the rotary embedding, and then you go forward. But during the backward, but if you do not apply QK norm, the Q times K matrix may have large values. And during the backward, the large value may bring a large gradient. And when you do the quantization, actually divide the data by the max of the per channel, the max of the column. So some small values will be divided nearly to the zero. So when you do the dequantize, it cannot find the real value before the quantization.

Nathan Lambert (33:52)

That makes sense. I see. Like, what are you actually looking at to figure this out? Are you looking at like intermediate activation values when you’re scaling? Because I like training loss will only show you so much, or are you like seeing that the training loss is better or worse and then going to investigate this later?

Chen Liang (34:08)

The first is the loss is not right compared to the BFloat16. And we print the quantization error during the intermediate layers and find that without QK norm in the linear QKV, the gradient is too large.

Nathan Lambert (34:34)

I think that this is very good. It gives people a sense for like what the different things moving around when you’re looking at kind of pre-training research is. And then the other side of things, if you make a change and then you have a loss spike, you’re like, okay, then you have like a numerical stability issue. I guess like a loss spike that you can’t skip. So I’m guessing you have things where if you have a loss spike, you can skip some of them. But there’s some numerical stability you can’t get around. This is fun. I’m going to kind of keep rolling through this. I think that you’re also talking about how you have like different pipeline for training your MoE, which you described as like a heterogeneous fine-grained pipeline. I think that this is like, I would read this as matching your training architecture to your compute architecture in order to get a speed up. Because I think with MoEs and the communication bottleneck. So I think that it’s like, if you want to talk about the parallelism strategies you did to get pre-training to be efficient. I think it was also really interesting because it covers multiple layers of the stack and how you design models.

Chen Liang (35:39)

It’s actually a common way, not just for our model. So actually the modern parallelism is just data parallel, tensor parallel, pipeline parallel, and context parallel. And our optimization is only focused on the pipeline parallel. As you can see from the paper, we do not use TP during our pre-training. So the common way to do the pre-training is they name it one forward and one backward type. Let’s see. We just focused on one machine with eight cards. And every card, actually, we name it as a stage. So we got stage 0 to stage 7. And every stage does the forward and the backward after it does the forward and sends the forward data to the next stage and they get the backward data from the next stage, right?

Nathan Lambert (36:49)

So that’s like an eight step pipeline. That’s like a pipeline parallel that you’re describing.

Chen Liang (36:53)

And every stage, they do communication from the prior stage and do the communication with the next stage. And the 1F1B got a problem that the stage 0 and stage 7 always got the most computation load because stage 0, you have an embedding layer. And it’s an index select operation. So it’s close. And stage 7, you got the LM head layer and the loss function. And you also got a large GEMM. So you need to times the hidden states to transfer the hidden states to the vocab size. And the vocab size is always large.

Nathan Lambert (37:45)

How much fine-grained work are you doing to change which part of the model is on each stage? Because that seems like what it would be then. You either have to change the model or you have to change how you split up the model. It’s like your two options.

Chen Liang (37:58)

The common way is just you split the LM head layer and embedding layer and just divide it by the GPU number. So it’s natural that the stage 0 and the stage 7 got much more computation load, since you just ignored the balance of the system when you split the layers. So it’s the common one. So our optimization’s main concern is just to alleviate the computation load of the stage 0 and stage 7.

Nathan Lambert (38:25)

I see. I guess I don’t fully follow like what has happened. I’m trying to be like very clear of whether or not I understand it. Because I think that’s like in a dense model, I think pipeline parallel really makes sense, but you have like a smaller model. And then as you’re getting bigger, it’s like much less of a model. I don’t know what it means to necessarily like de-load the specifically the embeddings or the loss function and how much of a change you can make. But I think that might be like a me limitation. It might be hard to get to, but you can, I’m curious if you want to try.

Chen Liang (39:14)

Actually, it’s quite the same as the dense model. The only difference is that per GPU, you can imagine that during the pre-training, if we got the 32 experts and we use like four machines to gather the expert data, it’s just you can view this four machine as one machine. So in this view, it’s the same like the dense model. So just imagine the dense model. You split the layers according to your GPU cards. And let’s assume that every machine got two layers of the dense model.

Nathan Lambert (40:11)

So I get that. And then it’s like, it’s just like, then you have to shift things around to make it so the loss is less of a bottleneck in the last layer or the final part of this pipeline parallel being the bottleneck is kind of potentially fundamental.

Chen Liang (40:24)

Yeah.

Post training at Inclusion

Nathan Lambert (40:25)

I see. I mean, the next question that I wanted to ask is going to be very related to this, which is like, what are your, how do you scale this to make RL work at the same scale? So the different problems that you have for doing pre-training versus RL with a large scale model. I don’t have the title of the paper, but you’re like in this Ling 1T paper, there’s a ton of RL details. And it’s like, is this kind of just like the next sequential problem that you got to? And then there’s just a lot of, not necessarily similar solutions, but like you’re doing your problem solving in the same way to make RL work rather than pre-training in terms of throughput.

Chen Liang (41:03)

It’s actually got some common tricks like we mentioned in the paper that the VPP (virtual pipeline parallelism). It actually means that the machine, you got double layers than the original one, than the original 1F1B, same things. But the difference is, let us assume that the stage 0 machine got four layers. But actually, during the time, two layers are doing computing and two layers are doing communication. So that’s what they call VPP.

Nathan Lambert (41:47)

What does two layers computing and communicating mean?

Chen Liang (41:50)

In other words, some layers are doing computing and some layers just prepare the data. They get the data.

Nathan Lambert (42:00)

I see, so it’s like some machines.

Chen Liang (42:03)

So when you train, during the computing, communication bandwidth is idle, right? So they utilize this to just like the exploration is the exploration. And our optimization is just to split the pipeline more precisely.

Nathan Lambert (42:31)

So I think I’m seeing that. So it’s within a node. You have very fast communication between eight GPUs. And then in pre-training, you’re kind of doing all sequentially, but in RL, you need to kind of sync this. You need to communicate more between your like generate, you have to move your weights to be able to generate when you’re doing RL. There’s like this sync step. And then I’m thinking what you’re saying is like, you have this chunk on eight GPUs and then you can split this. So half of them are doing compute and half are doing communication at the same time. So it kind of alleviates the bottlenecks. I see. For context and how like there’s a lot of different ways of doing RL infrastructure, it’s just the abstractions that like what we’re doing is much easier where we’re looking at approaches where we have GPUs that are set for generation and training, and that we are primarily looking at ways to make those both faster and then be able to throw the like training GPUs, we sync the weights to the generators and the generators just keep going where this is like it’s much more deeply embedded in the architecture where you have like one cluster where you’re kind of splitting the GPUs and what work is happening across each of the across like the per node basis when you’re doing this RL training. And I’m going to go look at this in more detail.

Chen Liang (43:48)

Yeah.

Richard Bian (43:56)

Just to add a little bit more flavors to this, the reason why we kind of didn’t really cover a lot of post-training details in this interview is because we have some additional technical papers or technical reports we’re writing at this moment about the system.

Nathan Lambert (44:14)

That makes sense.

Richard Bian (44:15)

So it was to some extent intentionally vague, Nathan. But I mean, first thing first, the current paper of Ling 1T and Ring 1T does have the fundamental intro for our system. It’s called a system. I believe the article has been published on ant-ling.medium.com/ on the medium technical paper as well as on Ling Team. So the paper is also available in English on Ling Team as we publish all the details. So specifically, there are several things which we did for the RL aspect. One is about the system itself. You can imagine that we do have an optimized internal hybrid engine which does all the things you described. And the second part is we’re exploring the reward model system. So this reward model system essentially requires some additional design to reach a certain level of parallelism. And the way we’re kind of looking into that is we’re really trying to set up meaningful rewards by doing a parallel structure for that. Last but not least, we have the term called LPO (Language-level Policy Optimization), right? It’s a linguistic unit. So we decided to choose sentence intentionally. So it’s kind of like a different approach from GRPO (Group Relative Policy Optimization) and the GSPO (Group Sequence Policy Optimization), like the session approaches or the token approaches that some of the other labs are using. We intentionally chose language as a linguistic unit to explore the meaning of this. So far, we’ve been seeing very motivating results from doing that. The training stability and the generalism is actually, we see some pretty clear numbers indicating that the LPO can be a very viable option for RL training. So let’s maybe save some of those interesting dessert for our next conversation. And we would love to really be able to share a lot of those details, given your background in post-training. I will try to maybe invite some of the experts from that domain into our next conversation.

Nathan Lambert (46:10)

I think the LPO thing is interesting, that there’s kind of a natural abstraction in a sentence. So in the language model generates, you just split every generation per sentence or per punctuation mark. It’s very linked to kind of these ideas of process reward models that people have looked at and understand to have natural inductive biases for a long time. And there is still some research doing this. So I’m happy to see that you’re doing it. And it’s kind of, I think of it as like value models and other things in RL that are just out of vogue and are likely to come back in some form in the near future, which is cool. In the ecosystem, where do you see open models going? I think it’s like, I guess the high level question is like, I mean, this weighs heavily on myself personally, it’s like, do you think that it’s like a big cake that you can eat out of and everybody does like, is it like, you see a clear path to having models that are meaningful? Does it worry you that the list of handful in China, it’s like, I mean, we know DeepSeek, we know Qwen, we know Kimi Moonshot, we know GLM 4.5, Meituan is releasing good, very strong models right now. You guys are like, the conviction that this is like a winning thing and you have your niche and there are more models coming soon. Like, is that easy for you to see? I mean, you had your metaphors at the beginning that I thought were great. So I think that’s kind of partially answered, but it’s like, it’s a very competitive space. So is that like easy for you to see through and just keep pushing ahead?

The Chinese model landscape

Richard Bian (48:15)

Thanks again for the invitation for really having this conversation. I did actually have my lines at the very beginning. I kind of call myself as a global citizen. Some of the current, I would say, really pains me in that regard. So when I’m kind of looking at it, so first thing first, I’ve been doing open source for years. You did ask about Ant OSS. You can actually find Ant OSS on Twitter. And there’s also a website for that. It’s opensource.antgroup.com. So Ant Group actually has a very long history of doing, as we call nowadays, the traditional or the classic open source, quote unquote, which I believe will be there forever. And you did ask a very specific question about open source models or open models. Last year, this time, it has been a very heated conversation in the open source ecosystem. So people in the open source domain are saying that, hey, this is open-weights. It’s not open source at all, which makes perfect sense. Because if you think about the nature of open source, it has at least three entities which are critically important. One is code itself, and the other one is community, aka the developers and people around it. And the last one is license, which pretty much provides a common consensus of the, I would say, the most common denominator as people agree upon, which is legally viable. But coming to that license requires years of effort. So like last year, you do see the OSI is trying to come out with a definition, and people are having a very convoluted feeling about it. And we see the Linux Foundation and data release this model openness framework, which is a very viable way of measuring the models. But that’s sad. Even nowadays, we only see one class one, which is a model from BAAI in China, which means by that standard, the rest of the models don’t meet it. And funny enough, last year, when I say we’re open sourcing our models, you’ll see people begin pointing fingers. Hey, you’re not open sourcing your models. Be careful about the words. But this year, all the labs are saying we’re open sourcing our models, and nobody is pointing fingers at all. Because it’s just like getting to a situation that we should maybe care less about this, but more about the direction, or what’s next. So I definitely want to spend more time discussing about that. So first thing first, I will say I did use the metaphor of saying the LLM is more like kernels. So if we kind of think about how many Linux kernel developers are on the planet now, it’s probably less than 1,000 people. So when people are saying that, hey, LLM is not really open source because nobody can contribute to it, yes, that’s correct. It’s very similar to the kernel. Theoretically, you can contribute to the kernel. But in reality, there’s only so few people who know about it. Most of the people are really kind of working around the ecosystem. They’re not the kernel developers, but we are currently at the stage of building the kernel itself. So that’s basically maybe my first point. It takes time. The reason why the open source definition is so convoluted at this moment, maybe just because it’s the first or the second year of a new era of neural development.

Nathan Lambert (51:54)

I agree. I think it’ll take like a decade. It’s like we’re in the first couple of years. I reiterate strongly with what you say where it’s like, it’s much better that people are actually using these models than just getting annoyed about definitions. And it’s like, we’ll figure out the definitions much more quickly if people actually want to use and contribute to these things.

Richard Bian (51:58)

And then the next part coming after this is like, I’m just sharing a very interesting story because I mean, my previous leader, he was working on Kubernetes and containers. So I have a background of being a full stack engineer as well as an engineer working on the data infra of the platform. So one day we did have a conversation about, I was saying, hey, you know, this MySQL infra, because I stopped. MySQL is not infra. MySQL is application. I was like, OK, thank you. That’s very helpful. But it’s kind of interesting, right? Because if you think about why that particular conversation actually happened, it’s because if you’re perceiving this from the infrastructure perspective or if you’re perceiving this from application perspective. My hunch feeling is we are going towards the next stage as we speak right now. I think we are at the transition period of having this MySQL moment. So other gigantic sandbox, gigantic runtime at this moment, that seems more application related. But five years down the road, they will become infrastructure. So the way I’m kind of looking at it is like, first thing first, I’m very optimistic about that. We will have open source. We will have an ecosystem in the AI era. In fact, I think Matt White from PyTorch, I think he introduced this new license called OpenMDW license, which kind of begins treating.

Nathan Lambert (53:53)

It’s an underrated license. It’s a very, very reasonable license.

Gaps in the open source ecosystem today

Richard Bian (53:59)

It’s very reasonable. In fact, I mean, we’re writing some Chinese articles trying to, I mean, I’m working with Art Eagles to do that. It deserves better visibility and more promotion. But kind of back to our original topic, I guess is, again, it will take quite some time for this information to rinse and repeat and consolidate. But I guess at this moment, I do see three gaps, which will prevent us from going to the next step. One is a proper license structure and a proper governance around the license. I think the OpenMDW is a good start, but it will take time. Second, I do believe data is the new code. So I guess how you’re contributing to the LLM is really through the data of pre-training and your data and reward models in post-training. But at this moment, there’s no Git for data. And the Git for data is not as straightforward as a Git for code because data can really be something which is very fundamental. So for instance, I mean like.

Nathan Lambert (55:07)

It’s often impermanent too. So like a lot of multimodal data sets are released as links and then the links die. So it’s like even like we try to, like people at AI2 try to release a fully reproducible data set and 10% dies in the first three years or something like.

Richard Bian (55:12)

And you might be having a lot of, I would say, overhead cost behind the scenes. So I mean, thanks so much for doing that. I mean, when people do that, we’re kind of raising our hands and saying hallelujah, right? Thank you. But it’s a difficult job, right? Because there might be legal battles behind the scenes. There might be a lot of, I would say, data cleansing. And the worst come to worst is really just more like, so I sometimes use this metaphor like, you know, I say, buy Coca-Cola stock. And Warren Buffett is saying, buy Coca-Cola stock. It’s literally the same word, but they mean something intrinsically different. I can’t really get my buy Coca-Cola. But I mean, that’s also a legal problem. So it’s like, in Git, we can say this, your public study was in main is before my public study was in main. But in data, you can’t really say that. So there’s definitely some technical challenges associated with that. Last but not least, the reward model associated with our contribution and the causality of our contribution to the model to the actual rewards. So for instance, if I’m writing a PR on GitHub, people see my PR and they merge my PR, great. I did my contribution. But you know.

Nathan Lambert (56:43)

I see.

Richard Bian (56:46)

Our conversation today is really meaningful. It can be a good, I would say, data corpus for reinforcement learning to some extent. But when people do that, they will not tell you, they will not tell me, they will not tell anyone of us.

Nathan Lambert (56:54)

I’m lucky enough to be big enough and visible where I accept that like me being in it is now good because it reinforces that I’m visible. Just a technical note on language, you were saying reward model as in the thing that rewards people for participating. Reward model is also like a technical thing, which I’ve done a lot of work on. So I was slightly confused, but if there was anybody else that was confused, that’s been clarified. To kind of zoom out, I think that listening to you, it’s like, wow, you’re one of a few people that is totally up to date on the open source definition stuff in the world. And I’m sure there are people all over that are thinking about this. I think you’ve spent a lot of time in both cultures and it’s like, where do you feel like people in my seat versus your seat may see things differently with like what open source AI means, what AI means generally, or like anything in this space that you feel both in your job or your life with respect to AI.

Richard Bian (57:58)

It’s a lovely question. I think it might be too big of a question, too. So I’ll probably answer that through two focuses. One is about open source ecosystem overall, like my feeling of being an engineer by training and global citizen, how I perceived open source ecosystem in general. And the second part is about the Chinese AI ecosystem. So we can tailor on that. So I will say first thing first about the open source ecosystem in the West and in the East. The first thing first, there are definitely more similarities than differences. I’m not sure if you read the book called Alchemist. It’s one of my favorite small books.

Nathan Lambert (58:42)

I haven’t actually read it. I do own it, unfortunately.

Richard Bian (58:45)

Well, congratulations. You have a nice book on your waiting list. It totally worth it. Another fun fact is I used to be working at Square. And the Square’s core payment system is called Esperanto. When I was looking at the word, I was like, what does that mean? And days after, I learned that Esperanto is basically this terminology related to world language. So there was a time people are inventing this term called Esperanto, hoping to connect the human beings altogether by speaking the same language. But clearly, it didn’t work. But now, Python is probably the real Esperanto to my best knowledge. So that’s why I’m saying that there are definitely more similarities than differences, because in open source domain, people are working together. Python code, JavaScript, speak English, they share their ideologies and meanings about technology. It’s all good.

Why China is winning the open race

Nathan Lambert (59:47)

A spicier way to phrase this question is like, why are there so many more open research labs in China than the US? I think like both, US arguably has like a bigger market cap, but fewer in people tech ecosystem. And it’s like, why is, it’s like, I listed what I thought was like 20 reasonable, like there’s like twice as many reasonable contributors in the Chinese ecosystem than in the US. Do you think there’s a reason for this or is it just kind of how the dice fell?

Richard Bian (1:00:11)

Well, I mean, I have my perceptions. Allow me to maybe use a disclaimer. So this is only my perception, not my company’s. So it kind of feels like there are definitely, there might be as many AI research labs in the States too. For instance, I mean, only through you, I learned about AI2. And I mean, I used to be living in Bellevue for years, but I didn’t know such an institution exists. So this is how uninformed I was. And I would imagine that there will be very much similar people like myself who are underinformed in that regard. Truth being told, we do see more open AI labs in China this year. I would say there are two reasons behind that. One is model effect. I would say that people are kind of perceiving the success of DeepSeek as a role model. That’s, I would say, a general consensus. It’s probably also a global consensus at this moment. People appreciate their engineering excellence and their willingness to share their findings. Because again, if we’re just out of Lisbon, we would appreciate the ship who came back and tell us, hey, this is the wrong way. Go that direction. We’ll probably appreciate that. So it’s not a zero sum game. So we cannot really speak on the other’s behalf, but we clearly see Alibaba with Qwen and Ant Group with Inclusion AI, we’re doing the same thing. We know it’s a long journey, it’s all the same. So when you’re outside of Lisbon, the best strategy to do is to be open and be helpful. And people appreciate the individuals who actually help you journey rather than the individuals who applaud you after you became famous.

Nathan Lambert (1:02:05)

I think I approach AI with this sort of curiosity. I think the, I don’t know how this would be a good test is like, there’s a very, the colloquial term of the hour in the Bay area and like tech circles in the US is like locked in. And if you apply this to what the AI companies think, it’s like the AI companies in the US are really, really like, at least acting as if they are locked in on a discovery in the near future that’s going to be transformative. A lot of it is probably for fundraising, but it’s like, I think that’s like, I have a lot more to learn and I will talk to more people like yourself to pick up more of this from talking to Chinese researchers. But I think this might be a recurring theme of like a lot of the US companies have this marketing that is really just different as how you’re describing it. And it’ll be interesting to see if that keeps coming up. Because if you’re so focused on like a one to two year thing, you’re not going to like sharing is a very different action to give. And then it’s like, it’s very different.

Richard Bian (1:03:07)

From a single perspective, I mean, just being told that by spending quite some time on both sides, I would say what we observe nowadays is reasonable, but definitely not ideal. So I would say first and foremost, you know, the chip leader is actually having a different way of playing the game, which is reasonable. I would say that, you know, if you’re the chip leader, there’s no guarantee that, you know, we’re going to be playing the same game. That being said, you know, it’s, we don’t talk about such a hypothesis because you cannot prove or disprove it. But that’s basically the first thing. And the second thing is we’re definitely seeing there are intrinsic, I would say, risks with the direction we’re going. So you hear people talking about the transformer architecture, we’re actually raising. You know, with all the names, they begin claiming that the pre-training might be dead. We hear terms like that. Reinforcement learning is the way to go. But in the latest interview with Andrej Karpathy, he shared this in a very humble and noble way, saying that, hey, this might be a good way to go, but let’s not mythify this. It might not be the golden desire, or it might not be the silver bullet. It’s a good methodology. Let’s go down that direction and explore, rinse and repeat, hoping that we’ll be able to find it. So if we’re at this stage of the game, I would say I would definitely choose the game to be more, I would say, open-minded. That’s one thing. And from a strategy perspective, be less about zero sum and more about where. So in game theory, there are all these kind of different games. One very typical mistake people make is they will treat a stag and hare game as a prisoner dilemma game. Those games look very similar in their own Nash equilibrium, but they’re different. So I guess, I mean, we do see certain companies are playing more like, hey, you know, you win, I lose. Can’t comment on that because, you know, there are a lot of reasons behind it. But, you know, the way we’re kind of looking at this, there are definitely more rooms, even as like Columbus was the first one finding the American continent. But then we begin to know that there’s this kind of North America and South America. And there are a lot of settlers, a lot of places. Right, so you don’t want to be the first pirate on Atlantic Ocean to kind of begin shooting down the other ships before you even reach and disembark. So that’s basically my way of seeing it. Last but not least, I guess I mean like.

Nathan Lambert (1:05:37)

There were a lot of settlers out there other than just Columbus. To finish your metaphor.

Richard Bian (1:06:03)

I think at this moment, there’s also another intrinsic risk associated with the whole business model. We hear a lot of those discussions regarding how Nvidia is actually making a lot of money by just selling the hardware. I also saw a line yesterday which I really like. It’s like, hey, do people still remember Cisco in 2000? I was like, that’s a very powerful line.

Nathan Lambert (1:06:27)

I think a lot about how Claude Code is very different than the likes of GitHub Copilot. And it’s like the different products that you can make with a given model has very, very big Delta in terms of what the user gets out of it. So mostly the floor is yours to comment on anything fun with product, which is probably a lot of your actual day job. I get, this is not my day job. And I get the sense that people that care about AI have to do a lot of work like this of like vision, creating a vision. And I’m guessing product might be closer to what you spend your time on.

Richard Bian (1:07:01)

Thanks so much, Nathan. I really enjoyed the conversation today. So the Model as Product team is very new. It’s brand-new. It’s only one month old. And as far as I know, we are the first company building such a team in China, if I’m not mistaken. But I have a hunch feeling that’s how people in OpenAI are working nowadays. So people are kind of working in small squad teams with seven to eight people. It’s a combination of algorithmic engineers, system engineers, UX engineers, product developers, evaluators, and so on. So we’re all working together.

Nathan Lambert (1:07:41)

Did you launch this before or after Sora? Because Sora is a complete vindication of this, which is like the genius of Sora is adding your friends to the videos versus just having a good video model. So you may not have realized it, but I think you have a great example of reinforcing this hypothesis. And I think more of them will come because I think, I don’t know, I’m soapboxing, but I think 2026 will be there will be more things that we can’t predict like Claude Code and Sora every year that start to work. So I think it’s a good approach.

Richard Bian (1:08:12)

That’s precisely how it works, right? Because working in open source for years, I guess one thing I learned is like, you know, if you just begin selling, I mean, there’s, you know, one of my favorite speakers is Simon Sinek, and he has a very popular YouTube video talking about leadership. So in there, one of his lines is like, leadership is volunteering. I really love that line. So basically, I’m pretty much one in my time and my predictions of trying to build such a team. So what our team does is like, because we are the Ant Ling team, right? So we care a lot about the model itself. That said, there are a lot of models out there. So in order for it to promote the model nowadays, it’s intrinsically difficult because people will say, oh, OK, here’s another model. Oh, it’s an open model from China. Oh, there are so many open models from China. It’s big, great. I remember that. But what’s next? So how can we use it? So we were kind of looking at just how we discuss about MySQL. If MySQL is a platform or an infra or product, I would say that we really want to think model as product now. Because you have all these models. But the good news is you also have the infrastructure, which allows you to switch models very easily, like open routers and all these model service providers, they actually allow you to do that very easily with very low overhead. You can use one model for part of a scenario and another model for the other part, which is good. It essentially means that if you have a good enough model, so I mean, thanks so much for our engineers who are actually building such a model for us to use and, you know, pretty much work upon. Without such a model, it’s impossible to do anything. So now with such a model, it almost feels like you have a very smart individual with IQ equals to 120, but he’s not very well-trained with anything. So what we’re trying to do is we’re trying to really find, during the interview with the model, and say, hey, what are you good at? But do we really know what the models are good at? Honestly, at this moment, it kind of feels like the evaluations are not really there. There’s a long way for benchmark evaluation. We don’t have enough time for that. But I believe that eval-driven heuristic is probably going to be very interesting in 2026. We’re going to essentially use an eval-driven way of finding what the models are good at. It can be very specific. It can be very niche for creative writing, for example, in drama, storyline. It’s very specific, but it can build a very good product on that. We’re trying to find all of those. But at this moment, we need the evaluation data set. We need all of this in order for us to be able to find it. And on the other hand, we need to find the user value. Because even as of 2025, you begin seeing a lot of new products coming out, but only a few things settled. So it almost kind of reminds me at the very early stage.

A metaphor for our moment in LLMs

Richard Bian (1:11:12)

I don’t know, Nathan, if you remember the product called Foursquare from the very early days.

Nathan Lambert (1:11:32)

I don’t think I was a man of the internet at the time, but I’ve heard of it as being like a canonical reference many times. It comes up in a lot of the readings that I do.

Richard Bian (1:11:38)

So the TLDR for that is Foursquare is basically one of the earlier applications when you have an iPhone. All it does, it gives you a location of your current phone, and you’re able to do a check-in action in there. So for instance, if you go to a restaurant, you can do a check-in at this restaurant. So what it does is actually it’s a demo of the location API of iPhone. And all it does is data labeling and a demonstration of how you can use the location API to be useful. But without Foursquare, you would not really have Uber or like DoorDash and all of those. So Foursquare was pretty much the demo, which led to all of these new products. And another way of putting that is like, you don’t have to be a taxi driver to build Uber. So that’s basically how our team is. We have a very small team. We have a very small team with engineers, product managers, and operational folks. So what we’re trying to do is we’re trying to essentially build Foursquare by really focusing on what the model is good at and what are the core capabilities. So I think there are definitely some of these demoable core capabilities which kind of begin surfacing. One of them is unlimited memory. Unlimited memory is basically this new capability which only AI and gen AI can fully utilize. But do we, so for instance, you have this kind of new products like the cloud note which you can put behind your phone, right? You can put a note there. Oh, I think there’s a company called unlimited.ai (editors note: called limitless) if I’m not mistaken, which is basically the necklace you can put. And people kind of building like watches, rings, glasses, and all of this in hoping to gather the data and trying to pretty much put all these kind of new contexts into the model. I kind of condense those into two core capabilities. One is unlimited memory. It memorizes everything. But in order for us to do that, you can’t really save all the data, right? The data is huge. You have to compress it, being able to find out a nice way of compressing them, and a very nice way of retrieving them. So data compression, data retrieval, called hot storage for all of this data, they’re all new challenges. But the capability is real. So with Unlimited Memory, it will really enable this contextual engineering work, which you can use in Model 4, but it’s not there yet. So it can be a Foursquare moment for the LLM. And the second one is, I would say, the proximity awareness. So for instance, we’re speaking in the room. There are a lot of these kind of new applications which are recording our meeting. What they’re really recording is the meeting, yes. But what they’re also recording is who is sharing the meeting with you. So theoretically speaking, you have sufficient amount of data. You can begin building the new LinkedIn in the gen AI era. It’s all possible, but we’re not there yet. So my team.

Nathan Lambert (1:15:00)

I think there’s a lot of pushback on privacy in the US to these things, but demonstrating the capability is obviously a huge merit of like, if we can figure out the privacy concerns, you have X on the table of new potential things. And I think it’s good. I encourage a lot of people to, it’s the right approach to things, which is like as the models get better, what potentially can work. I’m not a new person to saying this. A lot of people have.

Richard Bian (1:15:27)

Maybe just like two final words. One is like, I guess now is probably the best time to be more, I would say, first principle. Like, people say that a lot, but I actually have a three-year and ten months old boy at home. I guess one thing which really motivates me, what kept me being optimistic is my boy, because his growth is very well aligned with the timeline of the model. I’m seeing a lot of similarities in terms of how the revelations of human beings are kind of aligned with how the models are being trained, both pre-train and post-train. So I’m seeing there’s a long way to go. We don’t really have any understanding about, I would say human intellectual intelligence about where that’s coming from. So it’s a long journey and it’s good to really kind of think more fundamentally as the first principle. And the second line is I would say Inclusion AI and Ant Ling team, we’re being very serious about this. We don’t think this is a zero sum game and we don’t think this is Red Ocean. So I would say we’re open. We’ll stay open for as long as we can. And we’re doing all this kind of explorative approaches and I will probably make a call to action as someone who I’ve been benefiting a lot from globalization, including education and being able to work with smart people like you, Nathan. I hope the world will stay that way, at least as far as technology and open source is concerned. So that means work with us and Inclusion AI will be here. We’ll keep exploring and appreciate everything you’ve been doing for us. Thank you so much, Nathan. I really, really enjoyed this conversation today.

Nathan Lambert (1:17:15)

I look forward to seeing your new models. I have this, I’ve been so busy. I have one of these DGX Spark computers on my desk and I haven’t downloaded any real big model to it. And it’s like, I have to try downloading something like a hundred billion parameter model to see how it works. So maybe one of them will be your model. Thanks!

First, congrats to the Moonshot AI team, one of the 6 “AI Tigers” in China, on the awesome release of Kimi K2 Thinking. One of the overlooked and inspiring things for me these days is just how many people are learning very quickly to train excellent AI models. The ability to train leading AI models and distribute them internationally is going to be pervasive globally. As people use AI more, those who can access supply for inference (and maybe the absolute frontier in scale of training, even if costly) is going to be the gating function.

K2 Thinking sounds like a joy to use because of early reports that the distinctive style and writing quality from their original Kimi K2 Instruct model have been preserved through extended thinking RL training. They released many evaluation scores, for a highlight they’re beating leading closed models on some benchmarks such as Humanity’s Last Exam or BrowseComp. There are still plenty of evals where GPT 5 or Claude Sonnet 4.5 tops them. Rumors are Gemini 3 is coming soon (just like the constantly pending DeepSeek V4), so expectations are high on the industry right now.

TLDR: Kimi K2 Thinking as a reasoning MoE model with 1T total, 32B active parameters, 256K context length, interleaved thinking in agentic tool-use, strong benchmark scores and vibe tests.

The core reaction of this release is people saying this is the closest open models have been to the closed frontier of performance ever, similar to DeepSeek R1‘s fast follow to o1. This is pretty true, but we’re heading into murky territory because comparing models is harder. This is all advantaging the open models, to be clear. I’ve heard that Kimi’s servers are already totally overwhelmed, more on this soon.

What is on my mind for this release:

1. Open models release faster.

There’s still a time lag from the best closed to open models in a few ways, but what’s available to users is much trickier and presents a big challenge to closed labs. Labs in China definitely release their models way faster. When the pace of progress is high, being able to get a model out sooner makes it look better. That’s a simple fact, but I’d guess Anthropic takes the longest to get models out (months sometimes) and OpenAI somewhere in the middle. This is a big advantage, especially in comms, to the fast mover.

I’d put the gap at the order of months in raw performance — I’d say 4-6+ months if you put a gun to my head and made me choose specifically — but the problem is these models aren’t publicly available, so do they matter?

2. Key benchmarks first, user behaviors later.

Labs in China are closing in and very strong on key benchmarks. These models also can have very good taste (DeepSeek, Kimi), but there is a long-tail of internal benchmarks that labs have for common user behaviors that Chinese labs don’t have feedback cycles on. Chinese companies will start getting these, but intangible’s are important to user retention.

Over the last year+ we’ve been seeing Qwen go through this transition. Their models were originally known for benchmaxing, but now they’re legitimately fantastic models (that happen to have insane benchmark scores).

Along these lines, the K2 Thinking model was post-trained natively with a 4bit precision to make it far more ready for real serving tasks (they likely did this to make scaling RL more efficient in post-training on long sequences too):

To overcome this challenge, we adopt Quantization-Aware Training (QAT) during the post-training phase, applying INT4 weight-only quantization to the MoE components. It allows K2 Thinking to support native INT4 inference with a roughly 2x generation speed improvement while achieving state-of-the-art performance. All benchmark results are reported under INT4 precision.

It’s awesome that their benchmark comparisons are in the way it’ll be served. That’s the fair way.

3. China’s rise.

At the start of the year, most people loosely following AI probably knew of 0 Chinese labs. Now, and towards wrapping up 2025, I’d say all of DeepSeek, Qwen, and Kimi are becoming household names. They all have seasons of their best releases and different strengths. The important thing is this’ll be a growing list. A growing share of cutting edge mindshare is shifting to China. I expect some of the likes of Z.ai, Meituan, or Ant Ling to potentially join this list next year. For some of these labs releasing top tier benchmark models, they literally started their foundation model effort after DeepSeek R1. It took many Chinese companies only 6 months to catch up to the open frontier in ballpark of performance, now the question is if they can offer something in a niche of the frontier that has real demand for users.

4. Interleaved thinking on many tool calls.

One of the things people are talking about with this release is how Kimi K2 Thinking will use “hundreds of tool calls” when answering a query. From the blog post:

Kimi K2 Thinking can execute up to 200 – 300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex problems.

This is one of the first open model to have this ability of many, many tool calls, but it is something that has become somewhat standard with the likes of o3, Grok 4, etc. This sort of behavior emerges naturally during RL training, particularly for information tanks, when the model needs to search to get the right answer. So this isn’t a huge deal technically, but it’s very fun to see it in an open model, and providers hosting it (where tool use has already been a headache with people hosting open weights) are going to work very hard to support it precisely. I hope there’s user demand to help the industry mature for serving open tool-use models.

Interleaved thinking is slightly different, where the model uses thinking tokens in between tool use call. Claude is most known for this. MiniMax M2 was released on Nov. 3rd with this as well! It’s new.

5. Pressure on closed American labs.

It’s clear that the surge of open models should make the closed labs sweat. There’s serious pricing pressure and expectations that they need to manage. The differentiation and story they can tell about why their services are better needs to evolve rapidly away from only the scores on the sort of benchmarks we have now. In my post from early in the summer, Some Thoughts on What Comes Next, I hinted at this:

This is a different path for the industry and will take a different form of messaging than we’re used to. More releases are going to look like Anthropic’s Claude 4, where the benchmark gains are minor and the real world gains are a big step. There are plenty of more implications for policy, evaluation, and transparency that come with this. It is going to take much more nuance to understand if the pace of progress is continuing, especially as critics of AI are going to seize the opportunity of evaluations flatlining to say that AI is no longer working.

Are existing distribution channels, products, and serving capacity enough to hold the value steady of all the leading AI companies in the U.S.? Personally, I think they’re safe, but these Chinese models and companies are going to be taking bigger slices of the growing AI cake. This isn’t going to be anywhere near a majority in revenue, but it can be a majority in mindshare, especially with international markets.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

This sets us up for a very interesting 2026. I’m hoping to make time to thoroughly vibe test Kimi K2 Thinking soon!

Quick links:

* Interconnects: Kimi K2 and when “DeepSeek Moments” become normal, China Model Builder Tier List (they’re going up soon probably)

* Model: https://huggingface.co/moonshotai/Kimi-K2-Thinking

* API: https://platform.moonshot.ai/ (being hammered)

* License (Modified MIT): The same as MIT, very permissive, but if you use Kimi K2 (or derivatives) in a commercial product/service that has >100M monthly active users or >$20M/month revenue, you must prominently display “Kimi K2” on the UI. Is reasonable, but not “truly open source.” https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main/LICENSE

* Technical blog: https://moonshotai.github.io/Kimi-K2/thinking.html

* Announcement thread: https://x.com/Kimi_Moonshot/status/1986449512538513505

One of the obvious topics of the Valley today is how hard everyone works. We’re inundated with comments on “The Great Lock In”, 996, 997, and now even a snarky 002 (midnight to midnight with a 2 hour break). Plenty of this is performative flexing on social media, but enough of it is real and reflecting how trends are unfolding in the LLM space. I’m affected. My friends are affected.

All of this hard work is downstream of ever increasing pressure to be relevant in the most exciting technology of our generation. This is all reflective of the LLM game changing. The time window to be a player at the most cutting edge is actually a closing window, not just what feels like one. There are many different sizes and types of models that matter, but as the market is now more fleshed out with resources, all of them are facing a constantly rising bar in quality of technical output. People are racing to stay above the rising tide — often damning any hope of life balance.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

AI is going down the path that other industries have before, but on steroids. There’s a famous section of the book Apple in China, where the author Patrick McGee describes the programs Apple put in place to save the marriages of engineers traveling so much to China and working incredible hours. In an interview on ChinaTalk, McGee added “Never mind the divorces, you need to look at the deaths.” This is a grim reality that is surely playing out in AI.

The Wall Street Journal recently published a piece on how AI Workers Are Putting In 100-Hour Workweeks to Win the New Tech Arms Race. The opening of the article is excellent to capture how the last year or two has felt if you’re participating in the dance:

Josh Batson no longer has time for social media. The AI researcher’s only comparable dopamine hit these days is on Anthropic’s Slack workplace-messaging channels, where he explores chatter about colleagues’ theories and experiments on large language models and architecture.

Work addicts abound in AI. I often count myself, but take a lot of effort to make it such that work expands to fill available time and not that I fill everything in around work. This WSJ article had a bunch of crazy comments that show the mental limits of individuals and the culture they act in, such as:

Several top researchers compared the circumstances to war.

Comparing current AI research to war is out of touch (especially with the grounding of actual wars happening simultaneously to the AI race!). What they really are learning is that pursuing an activity in a collective environment at an elite level over multiple years is incredibly hard. It is! War is that and more.

In the last few months I’ve been making an increasing number of analogies to how working at the sharp end of LLMs today is similar to training with a team to be elite athletes. The goals are far out and often singular, there are incredibly fine margins between success and failure, much of the grinding feels over tiny tasks that add up over time but you don’t want to do in the moment, and you can never quite know how well your process is working until you compare your outputs with your top competition, which only happens a few times a year in both sports and language modeling.

In college I was a D1 lightweight rower at Cornell University. I walked onto a team and we ended up winning 3 championships in 4 years. Much of this was happenstance, as much greatness is, but it’s a crucial example in understanding how similar mentalities can apply in different domains across a life. My mindset around the LLM work I do today feels incredibly similar — complete focus and buy in — but I don’t think I’ve yet found a work environment where the culture is as cohesive as athletics. Where OpenAI’s culture is often described as culty, there are often many signs that the core team members there absolutely love it, even if they’re working 996, 997, or 002. When you love it, it doesn’t feel like work. This is the same as why training 20 hours a week while a full time student can feel easy.

Many AI researchers can learn from athletics and appreciate the value of rest. Your mental acuity can drop off faster than your physical peak performance does when not rested. Working too hard forces you to take narrower and less creative approaches. The deeper into the hole of burnout I get in trying to make you the next Olmo model, the worse my writing gets. My ability to spot technical dead ends goes with it. If the intellectual payoffs to rest are hard to see, your schedule doesn’t have the space for creativity and insight.

Crafting the team culture in both of these environments is incredibly difficult. It’s the quality of the team culture that determines the outcome more than the individual components. Yes, with LLMs you can take brief shortcuts by hiring talent with years of experience from another frontier lab, but that doesn’t change the long-term dynamic. Yes, you obviously need as much compute as you can get. At the same time, culture is incredibly fickle. It’s easier to lose than it is to build.

Some argue that starting a new lab today can be an advantage against the established labs because you get to start from scratch with a cleaner codebase, but this is cope. Three core ingredients of training: Internal tools (recipes, code-bases, etc.), resources (compute, data), and personnel. Leadership sets the direction and culture, where management executes with this direction. All elements are crucial and cannot be overlooked. The further along the best models get, the harder starting from scratch is going to become. Eventually, this dynamic will shift back in favor of starting from scratch, because public knowhow and tooling will catch up, but in the meantime the closed tools are getting better at a far faster rate than the fully open tools.

The likes of SSI, Thinky, and Reflection are likely the last efforts that are capitalized enough to maybe catch up in the near term, but the odds are not on their side. Getting infinite compute into a new company is meaningless if you don’t already have your code, data, and pretraining architectures ready. Eventually the clock will run out for company plans to be just catching up to the frontier, and then figure it out from there. The more these companies raise, the more the expectations on their first output will increase as well. It’s not an enviable position, but it’s certainly ambitious.

In many ways I see the culture of Chinese technology companies (and education systems) as being better suited for this sort of catch up work. Many top AI researchers trained in the US want to work on a masterpiece, where what it takes in language modeling is often extended grinding to stabilize and replicate something that you know definitely can work.

I used to think that the AI bubble would pop financially, as seen through a series of economic mergers, acquisitions, and similar deals. I’m shifting to see more limitations on the human capital than the financial capital thrown at today’s AI companies. As the technical standard of relevance increases (i.e. how good the models people want to use are, or the best open model of a given size category), it simply takes more focused work to get a model there. This work is hard to cheat in time.

This all relates to how I, and other researchers, always comment on the low hanging fruit we see to keep improving the models. As the models have gotten better, our systems to build them have gotten more refined, complex, intricate, and numerically sensitive. While I see a similar amount of low-hanging fruit today as I did a year ago, the efforts (or physical resources, GPUs) it can take to unlock them have increased. This pushes people to keep going one step closer to their limits. This is piling on to more burnout. This is also why the WSJ reported that top researchers “said repeatedly that they work long hours by choice.” The best feel like they need to do this work or they’ll fall behind. It’s running one more experiment, running one more vibe test, reviewing one more colleague’s PR, reading one more paper, chasing down one more data contract. The to-do list is never empty.

The amount of context that you need to keep in your brain to perform well in many LM training contexts is ever increasing. For example, leading post-training pipelines around the launch of ChatGPT looked like two or maybe three well separated training stages. Now there are tons of checkpoints flying around getting merged, sequenced, and chopped apart in part of the final project. Processes that used to be managed by one or two people now have teams coordinating many data and algorithmic efforts that are trying to land in just a few models a year. I’ve personally transitioned from a normal researcher to something like a tech lead who is always trying to predict blockers before they come up (at any point in the post-training process) and get resources to fix them. I bounce in and out of problems to wherever the most risk is.

Cramming and keeping technical context pushes out hobbies and peace of mind.

Training general language models you hope others will adopt — via open weights or API — is becoming very much an all-in or all-out domain. Half-assing it is becoming an expensive way to make a model that no one will use. This wasn’t the case two years ago, where playing around with a certain part of the pipeline was legitimately impactful.

Culture is a fine line between performance and toxicity, and it’s often hard to know which you are until you get to a major deliverable to check in versus competitors.

Personally, I’m fighting off a double-edged sword of this. I feel immense responsibility to make all the future Olmo models of the world great, while simultaneously trying to do a substantial amount of ecosystem work to create an informed discussion around the state of open models. My goal around this discussion is for more real things to be built. ATOM Project is a manifestation of me feeling that both the U.S. ecosystem generally and the Olmo project are falling behind.

It doesn’t really seem like there will be an immediate fix or end goal at this, but looking back I’m sure it’ll be clear what the key moments were and whether or not my efforts here and elsewhere met my goals.

Will it all be worth it? How long do you plan to go on like this? It’s not like we’re really going to suddenly reach AGI and then all pack it up and go home. AI progress is a long-haul now.

For me, the only reason to keep going is to try and make AI a wonderful technology for the world. Some feel the same. Others are going because they’re locked in on a path to generational wealth. Plenty don’t have either of these alignments, and the wall of effort comes sooner.

Thanks to Ross Taylor, Jordan Schneider, and Jasmine Sun for feedback on this post.

Two quick housekeeping items before I get to the post.1. I’ll be in SF this week for the PyTorch conference (22-23), AI Infra Summit (21st), and other local events. Come say hi.2. I launched a new Substack AI bundle with 8 of my favorite publications packaged together for teams of 20+. Learn more at readsail.com.Onto the post!

“Scaling reinforcement learning (RL)” is the zeitgeisty way to capture the next steps in improving frontier models — everyone is staring at the same hill they plan on climbing. How these different groups are approaching the problem has been a poorly kept secret. It’s a simple idea, but one that’s hard to copy: Predicting the trajectory of the learning curve. There have been two reasons this is hard to copy for academics, which will be solved on different time scales:

* The lack of stable RL training setups. There are many RL libraries being developed in parallel and the community has collectively made them much more ready for big RL runs over the summer.

* The lack of compute for experimentation.

These aren’t new stories. In many ways they mirror the progression of open Mixture of Experts (MoE) models, where they still lag far behind the implementations of the codebases within top AI laboratories because it involves overcoming substantial engineering headaches in an expensive experimentation regime. Scaling RL has been shaping up the same way, but it turns out it is just a bit more approachable.

Last week we got the first definitive paper on scaling RL. It proposes a clear method to extrapolate RL learning curves over compute scales and sets a baseline for the order of compute that should be spent to have top-end performance. The paper, The Art of Scaling Reinforcement Learning Compute for LLMs (Khatri & Madaan et al. 2025), referred to as ScaleRL, is a must read for anyone looking to understand the absolute cutting edge of RL algorithms and infrastructure.

For some personal context, for all of 2025 we’ve had our main slack channel in the reasoning space at Ai2 called “scaling-rl” because of how essential we knew the first clear piece of work in this area would be. This post covers the key details and what I see coming next.

There are two key things you need to know about these, even if all the lower level RL math is confusing to you too. First is how these intuitively work and what they’re actually predicting. Second is how they compare to the pretraining scaling laws we know and love.

To the first point, what the approach entails is taking one (or a handful of) your key base models, run a bit of RL on each of them, predict the end point by a bit of shape forecasting across many stable runs, then, for your big run, you can predict the end point in terms of final performance. The shape of RL runs that motivates this is how you see your model often gain ~80% of the accuracy gain in the first few steps, and you wonder what the final performance of the model will be if you trained on your entire dataset.

The authors define three constants that they fit, A for a measure of the peak performance — accuracy on a subset of your training dataset, aka the validation set, B for the slope of the sigmoid curve, and C as compute on the x axis. What is then done is that you take a set of RL training jobs and you fit a regression that predicts the last chunk of real training points given the early measurements of accuracy over time. Then, you can compare the predicted final performance of your future RL ablations on that starting model by understanding the normal shape of your RL learning curves.

Second is to consider how this compares to pretraining scaling laws. These are very far from the deeply insightful power law relating downstream test loss to pretraining compute — accuracy on RL training datasets is a far more bounded measure than next token prediction. The RL scaling laws are most useful for ablating design choices, relative to pointing to something fundamental about the nature of models. In many ways, scaling laws for pretraining could’ve been viewed this way at the beginning, too, so we’ll see how RL evolves from here.

With that difference, scaling laws for RL will play a very different role in training leading models than the pretraining scaling laws we have today. The pretraining laws are about choosing the exact configuration for your big pretraining run (that you can’t really run a meaningful chunk of to debug at all), where RL is more about ablating which algorithm you’ll let run much longer.

In pretraining many decisions depend on your budget and scaling laws can give the answer. Your training compute, communication bottlenecks, maximum run time, data availability, etc. all define a certain model window. Scaling laws for RL may inform this very soon, but for now it's best to think about scaling laws as a way to extract the maximum performance from a given base model.

For all of these reasons, scaling RL is more like an art, as the authors say it, because it’s about finding the run that’ll get the last few percentage points of performance when let run over an extra order of magnitude (or two) of samples. It’s a fine grained way to extrapolate RL curves — which have a standard shape of a quick rise then a slow saturation. In practice, the authors fit curves over 1/4 of their training compute to predict the outcome after the remaining 3/4 of GPU hours. The limits of scaling laws will likely be pushed further in the future (and I don’t have a good heuristic for what percentage of compute is used for establishing pretraining scaling laws, versus what is deployed in the final run, comment if you do!).

From here, the paper quickly gets technical, serving as a check in on the major ideas that dominated the RL research ecosystem in the last 6 months. This paper blesses those as important or not when it comes to scaled up RL training. This fits a recurring trend across language modeling in the last few years: Most of the key ideas are out there, but open labs tend to not have the resources to put them all together in the right configuration. This sort of slow accumulation of knowledge takes an organizational intensity, clarity, and ability that is hard for small research groups to match.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

There are a few key ideas that stand out to me as worth knowing and betting on following this paper:

* Algorithmic advancements: The paper is very favorable on, arguably painting them as essential, some recent algorithms or advancements. These include truncated importance sampling (TIS), Group Sequence Policy Optimization (GSPO), and Clipped IS-weight Policy Optimization (CISPO) via the MiniMax M1 paper. More on these in a second.

* Systems improvements: The authors highlight PipeLine RL (paper or repository) as the canonical reference for the combination of in-flight updates — i.e. changing model weights within one very long generation — and continuous batching — i.e. filling your RL batch over time until you have enough prompts for a learning step — which together represent 4X+ improvements over standard RL implementations on LLMs in terms of throughput. What this looks like in terms of idle GPUs is below, from the ServiceNow paper.

Intuitively, think about what happens if you were to ask 8 different questions to an LLM simultaneously. Some of these would finish early and some would take a long time. If you allocate your GPUs such that they have to finish all 8 questions before moving onto the next stack of questions, inevitably there will be GPUs idle when you’re waiting for the last answer. Instead, continuous batching pulls in new questions all the time when the GPUs have cycles to do more processing. Though, this is more complicated in the RL setup because after every 8 (or your batch size) of questions you need to update your RL weights. Can you still do this and fill in new questions all the time to the GPUs? What happens to that one question that is taking forever? In-flight updates is the solution to this. What is literally happening is that the model is updated in the middle of the generation. The models and RL systems just handle this seamlessly, and it removes a ton of idle time in matching the inference weights to the new updates from your RL algorithm.Not having a few key details like this will make big RL runs not only more expensive in GPUs, but more importantly in time. A 1 day feedback cycle vs 4 days makes for a very different research setup. We have these two features in Open Instruct, our post training repo at Ai2, as do many other RL libraries.

A lot of this is fixing numerics, which is far harder with Mixture of Experts (MoE) models, and something that most open RL research hasn’t touched. This hunt for numerical stability is a common rumor for why Thinking Machines put out the deterministic VLLM blog post ahead of releasing their Tinker API — deterministic VLLM could be their forward pass.

Back to algorithms.

Ross Taylor summarized the various eras of RL algorithms that the community has gone through in 2025. First was the transition from vanilla GRPO to the likes of DAPO (see my earlier post on GRPO tricks or my YouTube video on them too), which noticed issues with the clipping formulation and biases in the GRPO advantage calculation. The next class of algorithms are those cited in this ScaleRL paper, CISPO and a general class of Truncated Importance Sampling (TIS) approaches, that are designed for sequence level optimization (often closer to vanilla policy gradient) that account for the probability delta between actor (the GPUs generating completions for RL, often something fast like VLLM) and learner (the GPUs performing gradient updates, in a different library).

This importance sampling term seems to be essential to getting modern RL infrastructure right, as without it, scaling to more complex systems is hard to get numerical stability with. There’s been a lot of chatter about “importance sampling” in the AI community. What is happening, practically, is that the advantage or reward is getting re-weighted by an importance sampling log-ratio corresponding to the difference in probabilities from the two sets of model implementations (e.g. VLLM vs Transformers).

In the midst of all the details, the paper summarizes the state of affairs — large scale yolo RL runs — quite well:

While RL compute for LLMs has scaled massively, our understanding of how to scale RL has not kept pace; the methodology remains more art than science. Recent breakthroughs in RL are largely driven by isolated studies on novel algorithms (e.g., Yu et al. (DAPO, 2025)) and model-specific training reports, such as, MiniMax et al. (2025) and Magistral (Rastogi et al., 2025). Critically, these studies provide ad-hoc solutions tailored to specific contexts, but not how to develop RL methods that scale with compute. This lack of scaling methodology stifles research progress: with no reliable way to identify promising RL candidates a priori, progress is tied to large-scale experimentation that sidelines most of the academic community.

What is important going forward, as this will happen again with future eras of LLMs after this RL era, is why we are here. This happened due to the large overhang in potential from deploying RL, where clear scientific best practices take a long time to establish (even when most of the best researchers are publishing publicly, which isn’t the case today). The leading AI labs can build up fairly sizeable gaps quickly, but information tends to flow out and be reproduced. It’s important that the public options keep materializing — I think they will.

This paper is the first step in a direction of that science of scaling RL, but leaves many questions unanswered:

* No information on the impacts of different data. Polaris 53K is used in the paper, which is a solid option of the open, math RL datasets, but we find most of the RL data like this to be solved with a simple SFT set of reasoning traces on 8B models. Harder data may quickly become a limitation of open methods as people scale RL experiments to stronger base models. A paper reproducing these scaling trends over different data regimes is essential.

* No information on choosing the right base model. It is accepted that bigger base models perform better with RL — which the authors acknowledge in the paper: “the larger 17B×16 MoE exhibits much higher asymptotic RL performance than the 8B dense model, outperforming the 8B’s performance using only 1/6 of its RL training compute.” With this, we need to perform scaling RL studies that show the optimal base model for downstream RL, in terms of overall compute budgets.

The authors acknowledge these limitations clearly. They’re not trying to hide it!

To wrap this up, let us recall that there was a big brouhaha in AI circles a few weeks ago when a few frontier lab employees said that GRPO is far behind frontier labs RL stacks. What is more accurate to me is that vanilla GRPO is far behind, and the process of figuring out the set of individual tricks that works on your model and your data is a well kept secret. This new ScaleRL paper is a major step in showing people how to bridge that gap. From here, we have to build the tools in public.

This talk covers everything that’s happened this year in the open model landscape — DeepSeek kickstarting the Chinese open model norms, Llama’s fade, Qwen’s dominance, GPT-OSS — and what comes next. It is my attempt to share what people need to know about where open models are heading, building on all of my research here at Interconnects and in my day job of training these models, in order to help us take the actions we need to steer it in a better direction.

I strongly recommend watching (or listening, as it’s in the podcast feed) if any of the discussions around open models or Chinese AI impacts your decision making. This felt like one of the better talks I’ve given in a bit and I’m excited to keep expanding my coverage here.

You can click through the slides here.

Thanks to the organizers of The Curve for inviting me (and encouraging me to give this talk), and for permission to post this video.

EDIT: I noticed sometimes the audio jumps weirdly, not sure what caused it (from slideslive export, raw is here: https://slideslive.com/39046297/open-models-in-2025-stakes-state-and-strategy)

Chapters

00:00 2025 so far05:53 China takes the lead15:54 What comes next21:20 What we should do25:00 Q & A

(Podcast feed / Audio only version trims 7 seconds of silence to start)

References & Recommended Reading

* The ATOM Project

* On China’s open-source community & trajectory

* Ranking China’s open AI labs

* On GPT-OSS

* Recent open models

* More on The Curve conference

Of course, you can watch on YouTube:

Listen on Apple Podcasts, Spotify, YouTube, and where ever you get your podcasts.

I spent the weekend debating AI timelines, among other things, at The Curve conference. This translates as spending the weekend thinking about the trajectory of AI progress with a mix of DC and SF types. This is a worthwhile event that served as a great, high-bandwidth way to check in on timelines and expectations of the AI industry.

Updating timelines

My most striking takeaway is that the AI 2027 sequence of events, from AI models automating research engineers to later automating AI research, and potentially a singularity if your reasoning is so inclined, is becoming a standard by which many debates on AI progress operate under and tinker with. It’s good that many people are taking the long term seriously, but there’s a risk in so many people assuming a certain sequence of events is a sure thing and only debating the timeframe by which they arrive.

I’ve documented my views on the near term of AI progress and not much has changed, but through repetition I’m developing a more refined version of the arguments. I add this depth to my takes in this post.

I think automating the “AI Research Engineer (RE)” is doable in the 3-7 year range — meaning the person that takes a research idea, implements it, and compares it against existing baselines is entirely an AI that the “scientists” will interface with.

In some areas the RE is arguably already automated. Within 2 years a lot of academic AI research engineering will be automated with the top end of tools — I’m not sure academics will have access to these top end of tools but that is a separate question. An example I would give is coming up with a new optimizer and testing it on a series of ML baselines from 100M to 10B parameters. At this time I don’t expect the models to be able to implement the newest problems the frontier labs are facing alone. I also expect academics to be fully priced out from these tools.

Within 1-3 years we’ll have tools that make existing REs unbelievably productive (80-90% automated), but there are still meaningful technical bottlenecks that are solvable but expensive. The compute increase per available user has a ceiling too. Labs will be spending $200k+ per year per employee on AI tools easily (ie the inference cost), but most consumers will be at tiers of $20k or less due to compute scarcity.

Within 3-4 years the augmented research engineers will be able to test any idea that the scientists come up with at the frontier labs, but many complex system problems will need some (maybe minimal) amount of human oversight. Examples would include modifying RL implementations for extremely long horizon tasks or wacky new ideas on continual learning. This is so far out that the type of research idea almost isn’t worth speculating on.

These long timelines are strongly based on the fact that the category of research engineering is too broad. Some parts of the RE job will be fully automated next year, and more the next. To check the box of automation the entire role needs to be replaced. What is more likely over the next few years, each engineer is doing way more work and the job description evolves substantially. I make this callout on full automation because it is required for the distribution of outcomes that look like a singularity due to the need to remove the human bottleneck for an ever accelerating pace of progress. This is a point to reinforce that I am currently confident in a singularity not happening.

Up-skilling employees as their roles become irrelevant creates a very different dynamic. The sustained progress on code performance over the next few years will create a constant feeling of change across the technology industry. The range of performance in software is very high and it is possible to perceive relatively small incremental improvements.

These are very complex positions to hold, so they’re not that useful as rhetorical devices. Code is on track to being solved, but the compute limits and ever increasing complexity of codebases and projects (ie. LLMs) is going to make the dynamic very different than the succinct assumptions of AI 2027.

To reiterate, the most important part of automation in the discussion is often neglected. To automate someone you need to outcompete the pairing of a human with the tool too.

Onto the even trickier argument in the AI 2027 standard — automating AI research altogether. At the same time as the first examples of AI systems writing accepted papers at notable AI venues, I’m going to be here arguing that full automation of AI research isn’t coming anytime soon. It’s daunting to try and hold (and explain) this position, and it relies on all the messy firsthand knowledge of science that I have and how it is different in academia versus frontier AI labs.

For one, the level and type of execution at frontier labs relative to academic research is extremely different. Academia also has a dramatically higher variance in quality of work that is accepted within the community. For this reason, we’re going to be seeing incredible disruption at standard academic venues in the very near future, but the nature of science at frontier labs will remain heavily intertwined with human personalities.

Models will be good at some types of science, such as taking two existing fields and merging ideas and seeing what happens, but awful at what I consider to be the most idolized version of science, being immersed in the state of the art and having a brilliant insight that makes anywhere from a ripple causing small performance gain to a tsunami reshaping the field.

I don’t think AI will fully automate our current notion of an AI researcher in the next 5-10 years, but it could reshape what science means altogether and make that role far less relevant to progress.

The researchers grinding out new datasets at frontier labs will have dramatic help on data processing scripts. The researchers coming up with new algorithmic ideas will not expand the rate at which they come up with ideas too much, but their ability to test them is far higher.

A large part of science is a social marketplace of ideas. Convincing your colleagues that you are right and to help you double down on it is not going to change in its core nature. Everyone will have superpowers on making evidence to support their claims, but the relative power there stays the same.

At a dinner during The Curve I went through a lot of these points with Ryan Greenblatt, Chief Scientist at Redwood Research, and a point he made stuck with me. He summarized my points as thinking the increase in performance from these largely engineering tooling improvements will be equalled out by challenges of scaling compute, so the resulting progress will feel much more linear rather than exponential. A lot of our discussions on automation we agree on, with slightly different timelines, but it didn’t feel like it captured my entire point of view.

What is missing is that I expect an inherent slowdown as our AI models get more complicated. Our models today needs tools, more complex serving systems, products to wrap them, and so on. This is very different than the age when just model weights were needed for the cutting edge of AI. There’s an inevitable curse of complexity, a death by a thousand cuts, that is going to add on top of the obvious compute costs to slow down progress.

2026 will be a big year on the compute rollout front, and shipping meaningful improvements to users will be essential to funding the progress that comes after. I’m not sure the economy can keep shifting even more of its weight behind AI progress, where most people bought into fast timelines think of it as a default position. Peter Wildeford wrote a summary of the situation that I resonate with:

Here’s how I think the AI buildout will go down.

Currently the world doesn’t have any operational 1GW+ data centers. However, it is very likely we will see fully operational 1GW data centers before mid-2026. This likely will be a part of 45-60GW of total compute across Meta, Microsoft, Amazon/AWS/Anthropic, OpenAI/Oracle, Google/DeepMind, and xAI.

My median expectation is these largest ~1GW data center facilities will hold ~400,000-500,000 Nvidia Blackwell chips and be used to train ~4e27 FLOP model sometime before the end of 2027. Such a model would be 10x larger than the largest model today and 100x larger than GPT-4. Each individual 1GW facility would cost ~$40B to manufacture, with ~$350B total industry spend across 2026.

He continues with estimates for 2028, and saying he’s fuzzy on 2029, but my fuzziness cuts in a bit earlier depending on adoption and performance across the AI industry.

Where I feel like in the long run it’ll look like a very consistent pace of progress, that feels like a bunch of big jumps and periods of stagnation in the short-term. I have fairly large error bars on how the price of intelligence — and therefore adoption — is going to evolve over the next 2-4 years, with it obviously becoming far cheaper over the following decades.

As for my recent articles on timelines and key debates in the field, I encourage people to comment and dig in on what I wrote below.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

Other thoughts

Something crazy about this conference is no one is talking about how the models actually work or are trained, and everyone here is totally convinced that AGI is coming soon.

One of my new friends at the conference described this tendency as “an obsession with the problem.” This is a feeling that many AI obsessors are more interested in where the technology is going rather than how or what exactly it is going to be. Helen Toner gave a great talk at The Curve related to this, arguing how the current and future jaggedness of AI — the fact that similarly difficult tasks when assigned to a human will either be easily mastered by AI or barely showing any competence (her will appear later on her great Substack). It is the idea that AI capabilities evolve highly randomly across potentially similar tasks.

This original figure on jaggedness comes from work with the popular AI Substacker Ethan Mollick.

The relation of Helen’s talk is that she gets many forms of arguments that only the endpoint of AI matters, but that doesn’t account for the messiness of the trajectory and how unsettling that could be for the world.

I agree with Helen.

One of the things that I am confident will exist in about two years is a Sora 2 style model that can run on a MacBook without copyright, personal opt-in, or other safety filters. On this, Epoch AI has a wonderful plot showing that local models lag behind in capabilities by a fixed amount of time:

With trends like this, it is so obvious that we need to stay on the front foot of open models and not reacting to international parties that are far harder to predict and engage with. This is where I get renewed motivation for American DeepSeek / The ATOM Project. For example, I still get many adamant questions that we should consider banning open models altogether. The state of discourse, study, investment, and everything in between on open models in the U.S. is still in a quite underdeveloped state.

China’s rise in open models was something I expected to be a bigger topic at the conference, but it seemed like it was too orthogonal to the overall pace of progress to be front of mind. There were many discussions of the Chinese chip ecosystem, but less on what it enables. Not focusing on this could have costly geopolitical consequences as we cede ownership of a global standard to China. This was a large theme of my talk. The recording will be posted here soon and the slides for my talk are here (credit for Florian Brand who helps me with open model analysis here for feedback on the slides). Otherwise:

* These messages are very important and I will work to spend a bit more time engaging with the communities they touch and mastering this type of talk (and analysis)

* More people should work in the area, it’s crazy it has just fallen on me where it is my side hustle.

For now, I’m just landing at the conference on language modeling (COLM) in Montreal, so I may have some technical hot takes to share later this week!

Ever since ChatGPT exploded in popularity, there has been a looming “how” to its monetization plans. Much has been said about shopping and advertising as the likely paths, especially with Fidji Simo joining as CEO of Applications under Sam Altman.

Advertising as a business model for AI is logical but difficult to personalize and specialize. We know tons of people spend a lot of time using AI models, but how do you best get the sponsored content into the outputs? This is an open technical problem, with early efforts from the likes of Perplexity falling short.

Shopping is another, but the questions have long been whether AI models actually have the precision to find the items you want, to learn exactly what you love, and to navigate the web to handle all the corner cases of checkouts. These reflect a need for increased capabilities on known AI benchmarks, rather than inventing a new way of serving ads. OpenAI’s o3 model was a major step up in search functionality, showing it was viable; the integration was either a business problem — where OpenAI had to make deals — or an AI one — where ChatGPT wasn’t good enough at managing websites for you.

Yesterday, ChatGPT launched its first integrated shopping push with Buy It in ChatGPT, a simple checkout experience, and an integrated commerce backend built on the Agentic Commerce Protocol (ACP). The announcement comes with the perfect partners to complement the strengths of OpenAI’s current models. GPT-5-Thinking is the best at finding niche content on the web, and ChatGPT’s launch partner for shopping is Shopify (*soon, Etsy is available today), the home to the long tail of e-commerce merchants of niche specialties. If this works, it will let users actively uncover exactly what they are looking for — from places that were often hard to impossible to find on Google.

This synergy is a theme we’ll see reoccur in other agents of the future. The perfect model doesn’t make a useful application unless it has the information or sandbox it needs to think, search, and act. The crucial piece that is changing is that where models act is just as important as the weights themselves — in the case of shopping, it is the network of stores with their own rankings and API.

The ACP was built in collaboration with Stripe, and both companies stand to benefit from this. Stripe wants more companies to build on the ACP so that its tools become the “open standard for agentic payments” and OpenAI wants the long-tail of stores to adopt it so they can add them to their ever-growing internal recommendation (or search) engine. The business model is simple, as OpenAI says “Merchants pay a small fee on completed purchases.” OpenAI likely takes a larger share than Stripe, and it is a share that can grow as their leverage increases over shoppers.

I’m cautiously optimistic about this. Finding great stuff to buy on the web is as hard as it has ever been. Users are faced with the gamification of Google search for shopping and the enshittification of the physical goods crowding out Amazon. Many of the best items to buy are found through services like Meta’s targeted ads, but the cost of getting what you want should not be borne through forced distraction.

OpenAI will not be immune to the forces that drove these companies to imperfect offerings, but they’ll come at them with a fresh perspective on recurring issues in technology. If this works for OpenAI, they have no competitor. They have a distribution network of nearly 1B weekly users and no peer company ready to serve agentic models at this scale. Yes, Google can change its search feed, but the thoroughness of models like GPT-5 Thinking is on a totally different level than Google search. This agentic model is set up to make ChatGPT the one Agentic App across all domains.

The idea of an agentic model, and really the GPT-5 router itself, shows us how the grand idea of one giant model that’s the best for every conceivable use-case is crumbling. OpenAI only chooses the more expensive thinking model when it deems a free user to need it and they have an entirely different model for their coding products. On the other hand, Claude released their latest model, Claude 4.5 Sonnet, yesterday as well, optimizing their coding peak performance and speed yet again — they have no extended model family.

The reality that different models serve very different use-cases and how AI companies need to decide and commit to a certain subset of them for their development points to a future with a variety of model providers.

Where coding is where you can feel the frontier of AI’s raw intelligence or capabilities, and Anthropic has turned their entire development towards it, the type of model that is needed for monetization of a general consumer market could be very different. This is the web-agent that OpenAI has had the industry-leading version of for about 6 months.

Specialization is making the AI market far more interesting, as companies like OpenAI and Google have been in lockstep with their offerings for years. Every company would drop the same model modalities with approximately the same capabilities. Now, as hill-climbing benchmarks are no longer providing immediate user value, especially in text domains, the vision for each AI company is more nuanced. I predicted this earlier in the summer, in my post on what comes next:

This is a different path for the industry and will take a different form of messaging than we’re used to. More releases are going to look like Anthropic’s Claude 4, where the benchmark gains are minor and the real world gains are a big step.

What I missed is that this applies downward pressure on the number of models labs will release — the value can be more in the integrations and applications than the model itself. Expect releases like today, where Claude released Claude Sonnet 4.5 along with version 2 of Claude Code. The period will still be busy as the industry is on the tail end of the low hanging fruit provided by reasoning models, but over time the hype of model releases themselves will be harder to conjure.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

Let’s consider the applications that are rolling out today on top of different models. If you haven’t pushed the limits of GPT-5-Thinking, and better yet GPT-5-Pro, for search you really need to, it’s a transformative way of using compute that can find many buried corners of the web. In terms of untapped model capability value, the abilities of search-heavy thinking models like GPT-5 seem far higher than coding agents, which are obviously heavily used. Search-heavy models are an entirely new use, where coding models were the first widespread LLM-based product. As coding agents become more autonomous, they’ll continue to flex and mold a new form for the software industry, but this will be a slow co-evolution.

OpenAI is going to focus on its vertical Agentic App where Anthropic (and likely Gemini with Google Cloud) are going to power the long-tail of AI applications reshaping the web and the rest of work. OpenAI will only expand from here. Email, scheduling, travel bookings, and more everyday digital tasks are surely on their roadmap. Their biggest competitor is themselves — and whether their vision can be crafted into something people actually use. If shopping doesn’t work out as the vertical that lets them realize their valuation, they’re positioned to keep trying more. OpenAI has both the lead in the variety of models that power these agentic information tasks and the user base to incentivize companies to collaborate with them.

The application paradigm that dominated the mobile era is going to rebound. AI applications started in a form where the user needed to be heavily involved in the work process. The first beneficiaries of this were IDEs and terminal tools. Both of these workplaces allow in-depth and detailed inspection of the process and results. The cutting edge of AI will still work there, but the long tail of casual use will all shift to the standard mode of applications — siloed, simple, and scalable in the cloud. The simpler an AI application is, the wider its potential audience.

With this addition of shopping, OpenAI is poised to launch a standalone TikTok-style app with the release of its next video generation model, Sora 2, soon after Meta launched Vibes in their Meta AI app for only AI generated videos with a specific theme to start. At the same time, OpenAI’s Codex web agent is available in the ChatGPT application, which represents an even bigger change in the nature of software work than the addition of coding agents — it allows real websites, and soon businesses, to be built with only a prompt on your phone.

In 6-12 months, these agentic applications that feel rough around the edges due to the quality of the AI today, rather than the interface, are going to feel seamless and second-nature to use, despite their complete novelty relative to the past decades of technology.

If OpenAI is positioning itself to be The Agentic App, this also opens the door to the near future where many applications we use today shift to an agentic era. Want to schedule a meeting with someone? Let the Google Calendar agent handle that (or some startup that beats them to it). Your email application can find who the next client is and remind them of their appointment. The Banking App will file your taxes in one prompt. The list of these is infinite and across a wide spectrum of difficulty. OpenAI wants to be the one app, The Agentic App, that serves all of these, and the rest of the industry is racing to master their specific vertical before OpenAI gets there.

The weaknesses of today’s best models are far from those of the original ChatGPT — we see they lack speed, we fear superhuman persuasion, and we aspire for our models to be more autonomous. These models are all reasoning models that have long surpassed the original weaknesses of ChatGPT-era language models, hallucinations, total lack of recent information, complete capitulations, and other hiccups that looked like minor forms of delusion laid on top of an obviously spectacular new technology.

Reasoning models today are far more complex than the original chatbots that consisted of standalone model weights (and other lightweight scaffolding such as safety filters). They're built on three primitives that'll be around for years to come:

* Thinking: The reasoning traces that enabled inference-time scaling. The "thoughts" of a reasoning model take a very different form than those of humans that inspired the terminology used like Chain of Thought (CoT) or Thinking models.

* Searching: The ability to request more, specific information from non-parametric knowledge stores designed specifically for the model. This fills the void set by how model weights are static but living in a dynamic world.

* Acting: The ability for models to manipulate the physical or digital world. Everything from code-execution now to real robotics in the future allow language models to contact reality and overcome their nondeterministic core. Most of these executable environments are going to build on top of infrastructure for coding agents.

These reasoning language models, as a form of technology are going to last far longer than the static model weights that predated and birthed ChatGPT. Sitting just over a year out from the release of OpenAI's o1-preview on September 12, 2024, the magnitude of this is important to write in ink. Early reasoning models with astounding evaluation scores were greeted with resounding criticism of “they won’t generalize,” but that has turned out to be resoundingly false.

In fact, with OpenAI's o3, it only took 3-6 months for these primitives to converge! Still, it took the AI industry more broadly a longer time to converge on this. The most similar follow-up on the search front was xAI's Grok 4 and some frontier models such as Claude 4 express their reasoning model nature in a far more nuanced manner. OpenAI's o3 (and GPT-5 Thinking, a.k.a. Research Goblin) and xAI's Grok 4 models seem like a dog determined to chase their goal indefinitely and burn substantial compute along the way. Claude 4 has a much softer touch, resulting in a model that is a bit less adept at search, but almost always returns a faster answer. The long-reasoning traces and tool use can be crafted to fit different profiles, giving us a spectrum of reasoning models.

The taxonomy that I laid out this summer for next-generation reasoning models — skills for reasoning intelligence, calibration to not overthink, strategy to choose the right solutions, and abstraction to break them down — are the traits that'll make a model most functional given this new perspective and agentic world.

The manner of these changes are easy to miss. For one, consider hallucinations, which are an obvious weakness downstream of the stochastic inference innate to the models and their fixed date cutoff. With search, hallucinations are now missing context rather than blatantly incorrect content. Language models are nearly-perfect at copying content and similarly solid at referencing it, but they're still very flawed at long-context understanding. Hallucinations still matter, but it’s a very different chapter of the story and will be studied differently depending on if it is for reasoning or non-reasoning language models.

Non-reasoning models still have a crucial part to play in the AI economy due to their efficiency and simplicity. They are part of a reasoning model in a way because you can always use the weights without tools and they'll be used extensively to undergird the digital economy. At the same time, the frontier AI models (and systems) of the coming years will all be reasoning models as presented above — thinking, searching, and acting.

Language models will get access to more tools of some form, but all of them will be subsets of code or search. In fact, search can be argued to be a form of execution itself, but given the imperative of the underlying information it is best left as its own category.

Another popular discussion with the extremely-long generations of reasoning models has been the idea that maybe more efficient architectures such as diffusion language models could come to dominate by generating all the tokens in parallel. The (or rather, one) problem here is that they cannot easily integrate tools, such as search or execution, in the same way. These’ll also likely be valuable options in the AI quiver, but barring a true architectural or algorithmic revolution that multiplies the performance of today’s AI models, the efficiency and co-design underway for large transformers will enable the most dynamic reasoning models.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

With establishing what makes a reasoning model complete comes an important mental transition in what it takes to make a good model. Now, the quality of the tools that a model is embedded with is arguably something that can be more straightforward to improve than the model — it just takes substantial engineering effort — and is far harder with open models. The AI “modeling” itself is mostly open-ended research.

Closed models have the benefit of controlling the entire user experience with the stack, where open models need to be designed so that anyone can take the weights off of HuggingFace and easily get a great experience deploying it with open-source libraries like VLLM or SGLang. When it comes to tools used during inference, this means that the models can have a recommended setting that works best, but they may take time to support meaningful generalization with respect to new tools.

For example, OpenAI can train and serve their models with only one search engine, where I at Ai2 will likely train with one search engine and then release the model into a competitive space of many search products. A space where this can benefit open models could be something like MCP, where open models are developed innately for a world where we cannot know all the uses of our models, making something like MCP libraries a great candidate for testing. Of course, leading AI laboratories will (or have already started) do this, but the ranking will be different in a priority stack.

Much has been said about tokenomics and costs associated with reasoning models, without taking the tool component into account. There was a very popular article articulating how models are only getting more expensive, with a particular focus on reasoning models using far more tokens. This is overstating a blip, a point in time when serving costs increased by 1000x for models by generating vastly more tokens, but without improved hardware.

The change in cost of reasoning models reflected a one-time step up in most circumstances where the field collectively turned on inference-time scaling by using the same reasoning techniques. At the same time as the reasoning model explosion, the size of models reaching users in parameter count has all but stagnated. This is due to diminishing returns in quality due to scaling parameters — it’s why OpenAI said GPT 4.5 wasn’t a frontier model and why Gemini never released their Ultra model class. The same will come for reasoning tokens.

While diminishing returns are hitting reasoning token amount for serial streams, we’re finally seeing large clusters of Nvidia’s Blackwell GPUs come online. The costs for models seem well on path to level out and then decrease as the industry develops more efficient inference systems — the technology industry is phenomenal at making widely used products far cheaper year over year. The costs that’ll go up are the agents that are enabled by these reasoning models, especially with parallel inference, such as the Claude Code clones or OpenAI’s rumored Pro products.

What we all need is a SemiAnalysis article explaining how distorted standard tokenomics are for inference with tools and if tools substantially increase variance in implementations. People focus too much on the higher token costs from big models with long context lengths, those are easy to fix with better GPUs, while there are many other costs such as search indices or idle GPU time waiting for tool execution results.

When we look at a modern reasoning model, it is easy to fixate on the thinking token aspects that give the models their name. At the same time, search and execution are such fundamental primitives to modern language models that they can rightfully stand on their own as pillars of modern AI. These are AI systems that substantially depend on the quality of the complex inference stack far more than getting the right YOLO run for the world’s best model weights.

The cause of thinking, searching, and acting all being looped in as a “reasoning model” is that this inference-time scaling with meandering chains of thought was the technological innovation that made both search and execution far more functional. Reasoning was the step change event that set these three as technology standards.

The industry is in its early days of building out fundamental infrastructure to enable them, which manifests as the early days of language model agents. The infrastructure pairs deterministic computing and search with the beauty, power, and flexibility of the probabilistic models we fell in love with via ChatGPT. This reasoning model layer is shaping up to be the infrastructure that underpins the greatest successes of the future technology industry.

Coding, due to its breadth of use-cases, is arguably the last tractable, general domain of continued progress for frontier models that most people can interface with. This is a bold claim, so let’s consider some of the other crucial capabilities covered in the discourse of frontier models:

* Chat and the quality of prose written by models has leveled off, other than finetuning to user measures such as sycophancy.

* Mathematics has incredible results, but very few people directly gain from better theoretical mathematics.

* The AIs’ abilities to do novel science are too unproven to be arguable as a target of hillclimbing.

Still, coding is a domain where the models are already incredibly useful, and they continue to consistently stack on meaningful improvements. Working daily with AI over the last few years across side projects and as an AI researcher, it has been easy to take these coding abilities for granted because some forms of them have been around for so long. We punt a bug into ChatGPT and it can solve it or autocomplete can tab our way through entire boilerplate.

These use-cases sound benign, and haven’t changed much in that description as they have climbed dramatically in capabilities. Punting a niche problem in 1000+ lines of code to GPT-5-Pro or Gemini Deep Think feels like a very fair strategy. They really can sometimes solve problems that a teammate or I were stuck on for hours to days. We’re progressing through this summarized list of capabilities:

* Function completion: ~2021, original Github CoPilot (Codex)

* Scripting: ~2022, ChatGPT

* Building small projects: ~2025, CLI agents

* Building complex production codebases, ~2027 (estimate, which will vary by the codebase)

Coding is maybe the only domain of AI use where I’ve felt this slow, gradual improvement. Chat quality has been “good enough” since GPT-4, search showed up and has been remarkable since OpenAI’s o3. Through all of these more exciting moments, AIs’ coding abilities have just continued to gradually improve.

Now, many of us are starting to learn a new way of working with AI through these new command-line code agents. This is the largest increase in AI coding abilities in the last few years. The problem is the increase isn’t in the same domain where most people are used to working with AI, so the adoption of the progress is far slower. New applications are rapidly building users and existing distribution networks barely apply.

The best way to work with them — and I’ll share more examples of what I’ve already built later in this post — is to construct mini projects, whether it’s a new bespoke website or a script. These are fantastic tools for entrepreneurs and researchers who need a way to quickly flesh out an idea. Things that would’ve taken me days to weeks can now be attempted in hours. Within this, the amount of real “looking at the code” that needs to be done is definitely going down. Coding, as an activity done through agents, is having the barriers to entry fully fall down through the same form factor that is giving the act of coding re-found joy.

Why I think a lot of people miss these agents is that the way to use the agents is so different from the marketing of incredible evaluation breakthroughs that the models are reaching. The gap between “superhuman coding” announcements and using an agent for mini projects is obviously big. The best way to use the agents is still mundane and requires careful scoping of context.

For example, yesterday, on September 17, 2025, OpenAI announced that GPT-5 as part of a model system got a higher score than any human (and Google’s Gemini Deep Think) at the ICPC World Finals, “the premier collegiate programming competition where top university teams from around the world solve complex algorithmic problems.” Here’s what an OpenAI researcher said they did:

We competed with an ensemble of general-purpose reasoning models; we did not train any model specifically for the ICPC. We had both GPT-5 and an experimental reasoning model generating solutions, and the experimental reasoning model selecting which solutions to submit. GPT-5 answered 11 correctly, and the last (and most difficult problem) was solved by the experimental reasoning model.

These competitions often get highlighted because they’re “finite time,” so the system must respond in the same fixed time as a human does, but the amount of compute used by GPT-5 or another model here is likely far higher than any user has access to. This is mostly an indication that further ability, which some people call raw intelligence, can be extracted from the models, but most of that is limited by scaffolding and product when used by the general population.

The real story is that these models are delivering increasing value to a growing pool of people.

For followers of AI, coding with AI models is the easiest way to feel progress. Now that models are so good at chat, it takes very specialized tasks to test the general knowledge of models, or many of the gains are in getting the right answer faster than GPT-5-Thinking’s meandering path.

I’m not an expert software engineer and the huge differences between models, and improvements that the individual models and systems are making, have been incredibly obvious.

I’ve said many times how Claude Code (or now Codex) are far better than Cursor Agent, which is in turn far better than Github CoPilot. GitHub CoPilot feels borderline drunk at the wheel. Cursor often feels a little distracted while still being smart, but Claude Code and Codex seem on topic and able to test the best of a model’s intelligence on the problem at hand. Yes, even the best agents often aren’t good enough in complex codebases, but it removes the need to go back and forth countless times in a chat window to see if a model can reach the end of the puzzle for you. These CLI agents can run tests, fix git problems, run local tools, whatever. The scope is constantly growing.

For the nuanced take of Claude Code vs Codex CLI right now, the answer is expensive. The best has been Claude Code forcing Claude Opus 4.1, but Codex is not far behind and comes in at a much cheaper entry point ($20/month) — Opus requires a $100+/month plan. Codex also has nice features like web search, but it hasn’t been a major differentiator yet in my use.

The new workflow is to switch to the other agent when one cannot solve the current problem, and let it see the repository with fresh eyes, much like you pasted a question to another chatbot. The agents are just one tab away, just like the competitors for chat.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

In the comparison of Claude, Cursor, and CoPilot above, the crucial component is that all of these agents can be tested with the same Claude 4 Sonnet model. The gaps are just as wide as I stated, highlighting how so many of the gains in coding agents are just in product implementations. A second version is slightly embarrassing for me, but follows as I hadn’t updated my OpenAI Codex code when trying the new GPT-5-Codex model, which resulted in an immediate massive jump in performance by changing it. It’s a new phenomenon to have a domain at the cutting edge of AI abilities where the software scaffolding of a model is felt so strongly. Product and prompts matter more than ever and this sensation will expand to more domains.

The why of these performance differences — even when using the same model — is worth dwelling on. It’s unlikely that the Claude team is that much better at general software engineering and product design — rather, Anthropic has extensive in-house experience in extracting the most from models. The current shift in models has been about how to take a set of models that are designed for question answering and other single-stream text tasks and break down problems. In my taxonomy on next-generation reasoning models, I called this ability “abstraction.”

The need to just slightly shift the model to this task explains OpenAI’s recent specialized model for this, GPT-5-Codex. GPT-5 was primarily a release about balancing OpenAI’s books with a user base approaching 1B active users in the chat format. GPT-5 is a honed tool for a different job. The evaluation scores are slightly better than the general reasoning model for this new GPT-5-Codex, but the main gains are in how behavior is different in coding tasks.

GPT‑5-Codex adapts how much time it spends thinking more dynamically based on the complexity of the task. The model combines two essential skills for a coding agent: pairing with developers in interactive sessions, and persistent, independent execution on longer tasks. That means Codex will feel snappier on small, well-defined requests or while you are chatting with it, and will work for longer on complex tasks like big refactors. During testing, we've seen GPT‑5-Codex work independently for more than 7 hours at a time on large, complex tasks, iterating on its implementation, fixing test failures, and ultimately delivering a successful implementation.

And they included this somewhat confusing plot to showcase this dynamic. I’ve certainly felt these changes when I updated the Codex software and the Codex model.

This represents another key problem I presented in my taxonomy — calibration, i.e. not overthinking.

Having specialized models and specialized products for a use case could make people think that they’re narrowing in to make progress, but in OpenAI’s case it is rather that their hands are tied financially to support the main ChatGPT application. Claude has already fully committed to code. This is due to the size that the space could expand into.

These “coding” agents are definitely going to be seen as doing far more than writing code. Yes, their primary ability is going to be writing the code itself and executing it, but what that enables is an entirely new way of working with your computer.

In my post Contra Dwarkesh on Continual Learning, I presented a view where agents are going to be given all your digital working context in order to be a research or editorial assistant available 24/7. I’ve begun putting this to use for Interconnects, where I give the agents all of my articles, metadata, interviews, and details, so I can ask them for relevant references and context for future posts. This is very underbaked and early as a project for searching efficiently over my 400K tokens of writing, but I was prompting it a few times to see any interesting references for this post, and it got me something that was useful!

This quote from my Ross Taylor interview was spot on for the vibes of using coding agents in July:

My main worry with Claude Code is that... people confuse agents making you more productive versus preventing you from exerting mental effort. So sometimes I’ll have a day with Claude Code where I feel like I use very little mental effort—and it feels amazing—but I’m pretty sure I’ve done less work... Where it becomes really bad is when the file size becomes too long. Then the agent tends to struggle and get into these weird line search doom loops.

This sentiment is still definitely true for production codebases that are extremely complex, but the doom loop likelihood is dropping in my tests. At the same time, the joy and mental ease still applies.

Some examples of what I’ve built with a mix of Claude Code or OpenAI’s Codex CLI recently include:

* A raw HTML site for my RLHF book for comparing the responses of SFT vs. RLHF trained models from the same lineage (and improvements to RLHF book itself).

* Making a repository with all of the posts and content from Interconnects so I can use coding agents as editorial assistants while writing.

* Improvements to the ATOM Project website.

* Stripping my personal website out of Webflow’s systems (which was a mistake to sign up for during graduate school), including CMS entries and other detailed pages.

* Other small scripts and tools in my day job training models.

It’s not just me building extensively with these. There are multiple open-source projects committed to tracking the public contributions of these models — two are PRArena and Agents in the Wild.

PRArena’s dashboard shows over a million PRs getting merged from the Codex web agent, dwarfing many of the competitors. This is the power that OpenAI can wield with distribution, even if the web app version of Codex is far from the zeitgeist that is CLI agents today.

This comes with a notable asterisk in methodology that can explain many of the gaps in similar dashboards:

Some agents like Codex iterate privately and create ready PRs directly, resulting in very few drafts but high merge rates. Others like Copilot and Codegen create draft PRs first, encouraging public iteration before marking them ready for review.

The statistics below focus on Ready PRs only to fairly compare agents across different workflows, measuring each agent's ability to produce mergeable code regardless of whether they iterate publicly (with drafts) or privately.

The other dashboard, Agents in the Wild, shows that OpenAI’s coding agent is only one order of magnitude behind humans and other automations in PRs merged.

Putting this in perspective relative to Gemini or Claude:

The context with this is that Claude Code is far more downloaded than OpenAI’s CLI agent Codex, but it doesn’t name its PRs the same clever way by default with the agent name in the branch. Claude Code has over 20X the downloads of Codex in the last week on NPM.

Despite the challenges of measurement, it’s clear that coding agents are taking off.

The Codex PRs above actually represent the web agent, which has the default branch name behavior, not the CLI agent. This shows the might of OpenAI’s distribution, and it is impressive how many of the PRs are actually merged (over 80%), when thousands of people are trying a new tool for the first time.

The primary difference between the web agent and the CLI agent is a reduction in interactivity. The CLI agents propose a plan and ask for feedback, or let you monitor and interrupt. Codex on the web wraps a similar behavior as the CLI agents in one system that runs all the way until it can open a PR.

Over time coding is only going to get more asynchronous and OpenAI is poised to capture this transition if it happens soon. Based on all the above evidence of coding models getting more capable, the move to this new UX for software will happen faster than people expect. The transition to fully autonomous coding will happen soon for types of work where coding models already work near flawlessly — scripts, websites, data analysis, etc. Later, complex production codebases will work best at lower levels of the stack — IDEs, CLI agents, and other things that are both interactive and best for absorbing content.

Within a few years, the two trends will converge where autonomous agents are functional and the most complex codebases can be improved with AI. Then everything can return to the chatbot window — you only need to open your IDE when you want to understand what’s going on. For most people, not having to look at the code will be a welcome change.

Progress in coding feels slower than the “emergent” abilities between model generations past, which makes it easier to keep track of. This is due to how big the range in behaviors that encompass “coding” is, but results in a fantastic area for learning how AI models evolve and iterate. This playbook will be used many times over by frontier labs in the coming years as AI models are taught to solve more challenging tasks.

There’s a quiet revolution happening, and in order to truly understand it, you need to partake. Go build something.

Hello everyone! I’m coming back online after two weeks of vacation. Thankfully it coincided with some of the slowest weeks of the year in the AI space. I’m excited to get back to writing and (soon) share projects that’ll wrap up in the last months of the year.

It seemed like a good time to remind people of the full set of housekeeping for Interconnects.

* Many people love the audio version of the essays (read by me, not AI). You can get them in your podcast player here. Paid subscribers can add private podcast feeds under “manage your subscription” where voiceover is available for paywalled posts.

* The Interconnects Discord for paid subscribers continues to get better, and is potentially the leading paid perk amid the fragmentation of Twitter etc.

* We’re going to be rolling out more perks for group subscriptions and experimental products this fall. Stay tuned, or get in touch if group discounts are super exciting for your company.

For the time being, I’m planning trips and meetups across a few conferences in October. I’ll be speaking at The Curve (Oct. 3-5, Berkeley), COLM (Oct. 7-10, Montreal, interest form), and the PyTorch Conference (Oct. 21-24, SF) on open models, Olmo, and the ATOM Project, so stay tuned for meetups and community opportunities. On to the post!

China is maneuvering to double down on its open AI ecosystem. Depending on how the U.S. and its allies change culture and mobilize investment, this could make the dominance of Chinese AI models this summer, from Qwen, Kimi, Z.ai, and DeepSeek, looks like foreshadowing rather than the maximum gap in open models between the U.S. and China.

Until the DeepSeek moment, AI was likely a fringe issue to the PRC Government. The central government will set guidelines, rules, budgets, and focus areas that will be distributed and enforced across the decentralized government power structures. AI wasn’t a political focus and the strategy of open-source was likely set by companies looking to close the gap with leading American competitors and achieve maximum market share in the minimum time. I hear all the time that most companies in the U.S. want to start with open models for IT and philosophical reasons, even when spinning up access to a new API model is almost effortless, and it’s likely this bias could be even higher internationally where spending on technology services is historically lower.

Most American startups are starting with Chinese models. I’ve been saying this for a while, but a more official reference for this comes from a recent quote from an a16z partner, Martin Casado, another vocal advocate of investment in open models in America. He was quoted in The Economist with regards to his venture portfolio companies:

“I’d say 80% chance [they are] using a Chinese open-source model.”

The crucial question for the next few years in the geopolitical evolution of AI is whether China will double down on this open-source strategy or change course. The difficulty with monitoring this position is that it could look like nothing is happening and China maintains its outputs, even when the processes for creating them are far different. Holding a position is still a decision.

It’s feasible in the next decade that AI applications and open models are approached with the same vigor that China built public infrastructure over the last few decades (Yes, I’m reading Dan Wang’s new book Breakneck). It could become a new area that local officials compete in to prove their worth to the nation — I’m not sure even true China experts could make confident predictions here. A large source of uncertainty is whether the sort of top-down, PRC edicts can result in effective AI models and digital systems, where government officials succeeded in the past with physical infrastructure.

At the same time as obvious pro-AI messaging, Chinese officials have warned of “disorderly competition” in the AI space, which is an indirect signal that could keep model providers releasing their models openly. Open models reduce duplicative costs of training, help the entire ecosystem monitor best practices, and force business models that aren’t reliant on simple race-to-the-bottom inference markets. Open model submarkets are emerging for every corner of the AI ecosystem, such as video generation or robotic action models, (see our coverage of open models, Artifacts Logs) with a dramatic evolution from research ideas to mature, stable models in the last 12-18 months.

China improving the open model ecosystem looks like the forced adoption of Chinese AI chips, further specialization of companies’ open models to evolving niches, and expanded influence on fundamental AI research shared internationally. All of these directions have early signs of occurring.

If the PRC Government wanted to exert certain types of control on the AI ecosystem — they could. This Doug Guthrie excerpt from Apple in China tells the story from the perspective of international companies. Guthrie was a major player in advising on culture changes in Cupertino to better adapt Apple’s strategy to the Chinese market.

“When you stake your life, your identity, on and around certain ideas, you sort of fight for them,” Guthrie says. “Xi Jinping kind of broke my heart… I was sitting there, in China, in my dream job, and I’m watching Xinjiang’s internment camps. I’m watching China tearing up a fifty-year agreement over Hong Kong.”

Apple, meanwhile, had become too intertwined with China. Guthrie had been hired to help understand the country and to navigate it. And Apple had followed through—very successfully. But it had burned so many boats, as the saying goes, that Guthrie felt its fate was married to China’s and there was no way out. “The cost of doing business in China today is a high one, and it is paid by any and every company that comes looking to tap into its markets or leverage its workforce,” he later wrote in a blog. “Quite simply, you don’t get to do business in China today without doing exactly what the Chinese government wants you to do. Period. No one is immune. No one.”

China famously cracked down on its largest technology companies in late 2020, stripping key figures of power and dramatic amounts of market value off the books. AI is not immune to this.

The primary read here is that the PRC leadership will decide on the role they want to have in the open-source AI ecosystem. The safe assumption has been that it would continue because the government picked up a high-impact national strategy when it first started focusing on the issue, already seeded with international influence.

To formalize these intentions, the Chinese government has recently enacted an “AI+” plan that reads very similarly to the recent White House AI Action Plan when it comes to open models. The AI+ plan idea was first proposed in March 2024 and was just approved in its full text on July 31, 2025. The AI+ plan, when enacted by local officials, lays out goals for the AI industry in how many open models to have at different tiers of performance and some funding mechanisms for nurturing them.

This is right in line with other comments from party officials. Chinese Premier Li Qiang, second-ranking member of the Politburo Standing Committee, made comments in March directly supporting open-source models. From the Wall Street Journal:

Li pledged that China would boost support for applications of large-scale AI models and AI hardware, such as smartphones, robots, and smart cars.

China’s top economic planning body also said Wednesday that the country aimed to develop a system of open-source models while continuing to invest in computing power and data for AI.

An excerpt from Beijing’s city plan as part of the overall AI+ initiative, translated by GPT-5 Pro, has interesting, specific goals:

By end-2025: implement 5 benchmark application projects at a world-leading level; organize 10 demonstration application projects that lead the nation; and promote a batch of commercializable results. Strive to form 3–5 advanced, usable, and self-controllable base large-model products, 100 excellent industry large-model products, and 1,000 industry success cases. Take the lead in building an AI-native city, making Beijing a globally influential AI innovation source and application high ground.

The goal of this is to:

Encourage open-source, high-parameter, ‘autonomous and controllable’ base foundation models, and support building cloud hosting platforms for models and datasets to facilitate developer sharing and collaboration.

Beyond the minor translation bumpiness, the intentions are clear. The goal of the A+ plan is clear with multiple mentions of both open-source models and an open ecosystem with them where the models can be adopted widely. The ecosystem of models can make the impact of any one individual model greater than it would be alone.

The Chinese government having centralized power has more direct levers to enact change than the White House, but this comes with the same trade-offs as all initiatives face when comparing the U.S. vs. China’s potential. I won’t review all of the differences in the approaches here.

Where the Chinese Government enacts top level edicts that’ll be harder to follow from the West, there are numerous anecdotes and interactions that highlight in plain terms the mood of the AI ecosystem in China. I’ve routinely been impressed by the level of direct engagement I have received from leading Chinese AI companies and news outlets. Interconnects’ readership has grown substantially in China.

Chinese companies are very sensitive to how their open contributions are viewed — highlighting great pride in both their work and approach. The latest case was via our China open model rankings that got direct engagement from multiple Chinese AI labs and was highlighted by a prominent AI news outlet in China — 机器之心/Synced. They described Interconnects as a “high-quality content platform deeply focused on frontier AI research.” (This Synced post was translated and discussed in the latest ChinaAI Newsletter)

When intellectuals, influencers, and analysts I follow talk directly to technical members of the AI workforce in China, they sound like what we would expect — people who want to build a great technology. Jasmine Sun had a great writeup on her trip that had some anecdotes on AI in China. She asked “Do you guys worry about AI safety?”

“We don’t think about risks at all.” …

Continuing from Jasmine:

This was the first of several conversations that gave us a distinct impression of the Chinese tech community. Spirits are high, and decoupling policies like export controls only fuel their patriotic drive.

At the same time, America still represents a covetable life, despite the current political tumult:

To be clear, our researcher friend made clear that working at a top US AI lab was still the most desirable option.

In so many ways, trying to precisely map China’s next steps in AI is extremely challenging. Can they convert their lead in energy infrastructure to more total AI compute? Can they build their own AI chips? Will they take the frontier of performance with their talented population and a different approach? All of this is up for debate.

The intrigue here is exemplified by the abundant interest in sparse news stories on how DeepSeek is training some AI model with Huawei chips. In many ways, these new chips working would be a bigger story than the original DeepSeek model, but all signs point to expected experiments with domestic chips, where China’s leading AI models are likely to be trained on Nvidia and other Western chips for the foreseeable future. I do not expect DeepSeek R2 to be trained on Huawei’s hardware.

China’s hardware investment will take a lot longer to play out than open model strategies, but if China pulls it off — along with its other investments, such as self-driving cars and robots — their practical lead in AI could come for more areas. Open models could be China’s beachhead in a bigger technological resurgence with AI.

Without major changes to Western investment in open models, we’re approaching a status quo in 2026 and beyond where:

* Chinese open models would continue to increase their lead in performance (and adoption) over American counterparts. This will manifest in many ways. One example is how startups in Silicon Valley built on stronger Chinese models will be offering services that compete with entrenched, handicapped Fortune 500 companies wary of adopting these models in their services. This could make some subareas of AI disruption feel particularly intense.

* The Chinese open ecosystem’s density of knowledge and sharing would translate into increased scientific and academic impact. China’s share of conference papers at leading AI conferences is already rapidly on the rise, and having an ecosystem built around substantially better models than their Western counterparts could lead this numerous research growing also to be impactful. Better base models allow more interesting RL and agentic research today, and the list of areas reliant on high-performance models is likely to only grow longer with time.

* A proliferation of strong open models would make it difficult to restrict the presence or availability of many forms of AI. We do not have the government tools, incentives, nor culture to successfully prevent digital goods from China (or elsewhere) entering the U.S. economy. Many forms of AI governance and regulation in the United States and the rest of the world may need to be reconsidered, where many jurisdictions have looked to control and understand the development of “frontier AI.” Regulation needs to be approached for a world enmeshed in powerful AI models, rather than trying to control access or the releases of a few.

These realities all paint a clear picture that bends the association of open models from “soft power” to just “power.” Continuously releasing strong open AI models could allow Chinese companies to shape the technology interfaces, services and reality around the world. Where 2024 was about research on open models, and 2025 the professionalization of them, 2026 could be where we begin to see clear impacts of their power through endless distribution.

The Chinese AI ecosystem has taken the AI world by storm this summer with an unrelenting pace of stellar open model releases. The flagship releases that got the most Western media coverage are the likes of Qwen 3, Kimi K2, or Zhipu GLM 4.5, but there is a long-tail of providers close behind in both quality and cadence of releases.

In this post we rank the top 19 Chinese labs by the quality and quantity of contributions to the open AI ecosystem — this is not a list of raw ability, but outputs — all the way from the top of DeepSeek to the emerging open research labs. For a more detailed coverage of all the specific models, we recommend studying our Artifacts Log series, which chronicles all of the major open model releases every month. We plan to revisit this ranking and make note of major new players, so make sure to subscribe.

At the frontier

These companies rival Western counterparts with the quality and frequency of their models.

DeepSeek

deepseek.com | 🤗 deepseek-ai | X @DeepSeek_AI

DeepSeek needs little introduction. Their V3 and R1 models, and their impact, are still likely the biggest AI stories of 2025 — open, Chinese models at the frontier of performance with permissive licenses and the exposed model chains of thought that enamored users around the world.

With all the attention following the breakthrough releases, a bit more has been said about DeepSeek in terms of operations, ideology, and business model relative to the other labs. They are very innovative technically and have not devoted extensive resources to their consumer chatbot or API hosting (as judged by higher than industry-standard performance degradation).

Over the last 18 months, DeepSeek was known for making “about one major release a month.” Since the updated releases of V3-0324 and R1-0528, many close observers have been surprised by their lack of contributions. This has let other players in the ecosystem close the gap, but in terms of impact and actual commercial usage, DeepSeek is still king.

An important aspect of DeepSeek’s strategy is their focus on improving their core models at the frontier of performance. To complement this, they have experiments using their current generation to make fundamental research innovations, such as theorem proving or math models, which ultimately get used for the next iteration of models. This is similar to how Western labs operate. First, you test a new idea as an experiment internally, then you fold it into the “main product” that most of your users see.

DeepSeekMath, for example, used DeepSeek-Coder-Base-v1.5 7B and introduced the now famous reinforcement learning algorithm Group Relative Policy Optimization (GRPO), which is one of the main drivers of R1. The exception to this (at least today) is Janus, their omni-modal series, which has not been used in their main line.

Qwen

qwenlm.ai | 🤗 Qwen | X @Alibaba_Qwen

Tongyi Qianwen, the primary AI lab within Alibaba’s cloud division, is by far and away most known for their open language model series. They have been releasing many models across a range of sizes (quite similar to Llama 1 through 3) for years. Recently, their models from Qwen 2.5 and Qwen 3 have had accelerating market share among AI research and startup development.

Qwen is closer to American Big Tech companies than to other Chinese AI labs in terms of releases: They are covering the entire stack, from VLMs to embedding models, coding models, image and video generation, and so on.They also cater to all possible customers (or rather every part of the open community) by releasing capable models of all sizes. Small dense models are important for academia to run experiments and for small/medium businesses to power their applications, so it comes to no surprise that Qwen-based models are exploding in popularity.

On top of model releases for everyone, they also focused on supporting the (Western) community, releasing MLX and GGUF versions of their models for local usage or a CLI for their coding models, which includes a generous amount of free requests.

Unlike some American companies, the core team seems to have stayed relatively small in terms of headcount, in line with other Chinese AI labs: Qwen3 has 177 contributors, whereas Llama 3 has thrice the amount, while Gemini 2.5 has over 3,000 people as part of the model.

Close competitors

These companies have recently arrived at the frontier of performance and we will see if they have the capability to consistently release great models at a pace matching Qwen or DeepSeek.

Moonshot AI (Kimi)

moonshot.cn | 🤗 moonshotai | X @Kimi_Moonshot

Moonshot AI is one of the so-called “AI tigers”, a group of hot Chinese AI startups determined by Chinese media and investors. This group consists of Baichuan, Zhipu AI, Moonshot AI, MiniMax, StepFun, and 01.AI — most of which have attracted investments by tech funds and other tech grants. For example, Alibaba is seen as a big winner in the AI space by having their own models and by being a lead investor in Moonshot, sort of like how big tech companies in the U.S. are investing in fundraising rounds for newer AI labs.

While their first models, K1 and K1.5, were closed and available on their API, they started releasing open models after the R1 release with experimental models using the Muon optimizer. Similar to DeepSeek, they focus on a single model line, with small experiments eventually feeding back into the main model. K2 is their “moonshot run,” a.k.a. yolo run, and quickly became a hit similar to R1 (see our report from the release).

Further reading on Kimi can be found on ChinaTalk.

Zhipu / Z.AI

z.ai | 🤗 zai-org | X @Zai_org

Zhipu, known in the west as Z.ai, is a startup spinoff of Tsinghua University with considerable investments by Chinese companies and VCs. Currently, they are even considering an IPO, which would make them the first AI tiger to do so.

In terms of models, they are mostly known for their recent release of GLM-4.5 and GLM-4.5V, which are all very capable for their sizes (both of which are fairly large mixture of expert models). However, they are not just releasing LLMs, but also image and video generation models, setting them apart from pure-LLM companies and labs.

Noteworthy

These companies are transitioning to open releases, have open models with inferior capabilities, or slightly different foci than the text-centric labs pushing the frontiers of intelligence.

StepFun

stepfun.ai | 🤗 stepfun-ai | X @StepFun_ai

StepFun first started as a closed model provider, but pivoted to open model releases after DeepSeek R1 shook up the industry. They are mostly focusing on multi-modal model releases, with Step3 being their flagship VLM. They also have image, audio and video generation models.

Tencent (Hunyuan)

hunyuan.tencent.com | 🤗 Tencent | X @TencentHunyuan

Hunyuan is mostly known for HunyuanVideo and Hunyuan3D. While they have released three series of different LLMs, their releases come with very strict licenses, which is unusual for Chinese companies and dampens excitement when combined with performance levels that can be found elsewhere.

RedNote (Xiaohongshu)

xiaohongshu.com | 🤗 rednote-hilab

The Chinese version of Instagram, RedNote, recently joined the ranks of Chinese companies releasing open models. Especially their capable character recognition / OCR model surprised many (see our coverage). Similar to Xiaomi and Baidu, it remains to be seen what their overall open strategy will be in the near and distant future and they have not competed in the large, frontier model space.

MiniMax

minimaxi.com | 🤗 MiniMaxAI | X @MiniMax__AI

MiniMax is another of the AI tigers and also started as a closed company. After the release of R1, they changed their strategy and released the weights of Minimax-Text-01, following up with reasoning models building upon it. The unique selling point of these models are the 1M context window achieved with hybrid attention.

These text models are not the only thing they are focusing on — they also have image and video generation models, but those remain closed and only available on their API. They are also promoting their consumer platform heavily as they eye an IPO.

OpenGVLab / InternLM

internlm.intern-ai.org.cn | 🤗 InternLM | X @opengvlab

InternLM & OpenGVLab have deep ties to the Shanghai AI Laboratory, with InternLM focusing on the language models, while OpenGVLab releases vision models. While they release a range of models such as S1 or InternLM-Math, the orgs are mostly known for the strong InternVL series. While the first versions mostly used their own InternLM pretrained models, later releases (such as InternVL3) rely on Qwen as the language backend.

Skywork

skywork.ai | 🤗 Skywork | X @Skywork_AI

The Singaporean Skywork first started out as an online karaoke company (yes, really) before they pivoted to AI and being a competitor to Manus, with their platform focusing on agents for work-related tasks, such as slide generation.

Their LLM journey started with them releasing their own pretrained dense and MoE models. However, they stopped pre-training their own models and instead started to fine-tune existing models: Their OR1 reasoning model builds on top of DeepSeek-R1-Distill-Qwen-32B, R1V3 uses InternVL3 (which itself uses Qwen2.5 as its LLM backend).

Aside from LLMs, they have a wide range of other models, from world models, image and video generation models, and reward models. Similar to their LLMs, they mostly build on top of other models. Unlike many labs, Skywork has released some datasets with their models, such as preference and reasoning training data.

On the rise

These companies are either just getting their toes wet with open models or operating as more of academic research organizations than labs pushing the performance of models.

ByteDance Seed

seed.bytedance.com | 🤗 ByteDance-Seed

Seed is the R&D arm of ByteDance and eerily similar to Meta’s FAIR division: Diverse models with interesting research, with their papers garnering a ton of attention in the community. However, it remains to be seen whether they shoot for a Llama-style model release or continue to release research artifacts.

Here are some recent papers:

* Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference

* Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving

* Seed-X: Building Strong Multilingual Translation LLM with 7B Parameters

* Seedance 1.0: Exploring the Boundaries of Video Generation Models

* SeedEdit 3.0: Fast and High-Quality Generative Image Editing

* Seed1.5‑VL Technical Report

* Mogao: An Omni Foundation Model for Interleaved Multi‑Modal Generation

* Seed1.5‑Thinking: Advancing Superb Reasoning Models with Reinforcement Learning

* VAPO: Efficient and Reliable Reinforcement Learning for Advanced Reasoning Tasks

* Seed LiveInterpret 2.0: End‑to‑end Simultaneous Speech‑to‑speech Translation with Your Voice

OpenBMB

openbmb.ai | 🤗 openbmb | X @OpenBMB

OpenBMB is an open-source community (comparable to BigScience) from Tsinghua University NLP Lab (the very same university where Zhipu was spun off from) with support from the Beijing Academy of Artificial Intelligence (BAAI) and ModelBest.

They are mostly focusing on small multi-modal models for the edge, such as MiniCPM-V-4. However, the license is rather restrictive, which is surprising given the community-driven origins of the group. Aside from model releases, they also release frameworks and specialized kernels to make sure their models run on low-end hardware.

Xiaomi (MiMo)

mi.com | 🤗 XiaomiMiMo

Xiaomi started releasing a bunch of small, capable models, ranging from LLMs to VLMs and audio models. Xiaomi updating the models quickly after an initial launch and releasing multiple variants of the models show that it is not a one-off foray into open models. However, it remains to be seen whether those are mostly research artifacts or whether they are serious about potentially pushing the frontier or competing for adoption.

Baidu (ERNIE)

yiyan.baidu.com | 🤗 baidu | X @Baidu_Inc

Baidu, one of the original names in the Chinese AI space, has only released the weights of ERNIE 4.5. It remains to be seen whether they continue to release weights of newer releases as well.

Honorable Mentions

The rest of the labs that we are watching.

Multimodal Art Projection

m-a-p.ai | 🤗 m-a-p

An open research community, releasing all kinds of models (including a truly open 7B language model with data, etc.). Now, they’re mostly known for the music generation model YuE.

Alibaba International Digital Commerce Group

aidc-ai.com | 🤗 AIDC-AI

Another R&D arm of Alibaba, mostly releasing niche models building upon Qwen.

Beijing Academy of Artificial Intelligence (BAAI)

baai.ac.cn | 🤗 BAAI | X @BAAIBeijing

As a university, the Beijing Academy of Artificial Intelligence has a high diversity of projects. They are mostly known for BGE, which are capable embedding models.

inclusionAI

🤗 inclusionAI | X @InclusionAI666

The open weight arm from the Ant Group (an affiliate of Alibaba handling mobile payments and some financial industries), responsible for Ling Lite, a series of LLMs.

Pangu (Huawei)

huaweicloud.com | X @HuaweiCloud1

Huawei is working on AI accelerators to threaten the market share of Nvidia GPUs, which are often targeted by regulations, both from the US and China. Their model releases are mostly to show what’s possible with their cards, but not without drama accusing them of upcycling Qwen models and not stating it. We would expect them to continue to release more models in the near future.

Dwarkesh Patel’s now well-read post on why he is extending his AI timelines focuses on the idea of continual learning. If you ask me, what we have already is AGI, so the core question is: Is continual learning a bottleneck on AI progress?

In this post, I argue that continual learning as he describes it actually doesn’t matter for the trajectory of AI progress that we are on. Continual learning will eventually be solved, but in the sort of way that a new type of AI will emerge from it, rather than continuing to refine what it means to host ever more powerful LLM-based systems.

Continual learning is the ultimate algorithmic nerd snipe for AI researchers, when in reality all we need to do is keep scaling systems and we’ll get something indistinguishable from how humans do it, for free.

To start, here’s the core of the Dwarkesh piece as a refresher for what he means by continual learning.

Sometimes people say that even if all AI progress totally stopped, the systems of today would still be far more economically transformative than the internet. I disagree. I think the LLMs of today are magical. But the reason that the Fortune 500 aren’t using them to transform their workflows isn’t because the management is too stodgy. Rather, I think it’s genuinely hard to get normal humanlike labor out of LLMs. And this has to do with some fundamental capabilities these models lack.

I like to think I’m “AI forward” here at the Dwarkesh Podcast. I’ve probably spent over a hundred hours trying to build little LLM tools for my post production setup. And the experience of trying to get them to be useful has extended my timelines. I’ll try to get the LLMs to rewrite autogenerated transcripts for readability the way a human would. Or I’ll try to get them to identify clips from the transcript to tweet out. Sometimes I’ll try to get them to co-write an essay with me, passage by passage. These are simple, self contained, short horizon, language in-language out tasks - the kinds of assignments that should be dead center in the LLMs’ repertoire. And they're 5/10 at them. Don’t get me wrong, that’s impressive.

But the fundamental problem is that LLMs don’t get better over time the way a human would. The lack of continual learning is a huge huge problem. The LLM baseline at many tasks might be higher than an average human's. But there’s no way to give a model high level feedback. You’re stuck with the abilities you get out of the box. You can keep messing around with the system prompt. In practice this just doesn’t produce anything even close to the kind of learning and improvement that human employees experience.

The core issue I have with this argument is the dream of making the LLMs we’re building today look more like humans. In many ways I’m surprised that Dwarkesh and other very AGI-focused AI researchers or commentators believe this — it’s the same root argument that AI critics use when they say AI models don’t reason. The goal to make AI more human is constraining the technological progress to a potentially impossible degree.

Human intelligence has long been the inspiration for AI, but we have long surpassed it being the mirror we look to for inspiration. Now the industry is all in on the expensive path to make the best language models it possibly can. We’re no longer trying to build the bird, we’re trying to transition the Wright Brothers’ invention into the 737 in the shortest time frame possible.

To put it succinctly. My argument very much rhymes with some of my past writing.

Do language models reason like humans? No. Do language models reason? Yes.

Will language model systems continually learn like humans? No.Will language model systems continually learn? Of course.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

Dwarkesh writes “Rather, I think it’s genuinely hard to get normal humanlike labor out of LLMs.” This is because we’re still early on the buildout of the technology. Human labor takes an immense amount of context and quick thinking, both of which we’re starting to unlock with our language models. On top of this, human labor may not be what we want to create — we want to augment it.

Using LLMs as drop in replacements for humans is not a requirement for AGI nor is what Dwarkesh describes a fundamental limitation on AI progress. Francois Chollet cleverly poked at this weakness in his recent conversation with Dwarkesh at an ARC-AGI event:

Well, how do you define the difference between the ability to adapt to a new task and learning on the fly? It's, it sounds like the same thing to me.

Language models can already pick up subtle context extremely fast. ChatGPT’s memory feature has gotten far better for me. When we’re using the far more powerful models we can expect in the next 18 months this’ll already start to appear magical. Language models are extremely apt at inferring context even without us giving it to them. Soon we’ll be unlocking that subtle connection engine by providing immense, explicit context.

I don’t know of anyone who has actually thoroughly digitized all the relevant context of their job and formatted it in a way that is easily readable by an LLM. GPT-5 Pro estimates that all of the writing on Interconnects would be only 500K tokens. That would fit into an existing LLM with no extra system, but I’ve never tried it.

The problem that Dwarkesh is facing is that we’re still using LLMs primarily in a single generation manner, which got far better with the introduction of reasoning models, but the economically useful way to use current tools in more complex intellectual domains will require a deep-research style approach over all of your recent work interactions. No one is giving language models that kind of context. None of the tools we use are set up properly to accumulate this type of context.

I expect this to change rapidly. ChatGPT, Claude, and the likes are all adding memory features across chats and countless connectors to other pieces of information in your professional life. These memory features will be omnimodal and essential to extracting the type of value Dwarkesh wants. Without them, I agree language models in their current form are hopeless at solving continual learning.

This is what I would expect the rumored $2000/month ChatGPT level subscriptions to work with. Each of these bespoke tasks needs to absorb a ton of context and reasoning tokens in order to make a directionally right output. If someone built the Claude Code equivalent for my Substack, with every post tagged by topic and performance metrics, I bet the AI could easily make useful suggestions on how to format my content.

Continual learning in how Dwarkesh presents it is a systems problem rather than a learning problem. I expect better context management over my information ecosystem to exist in 2026, but more work to be needed for the AI companies to know how best to reference it and unlock in-context learning that feels like rapid adaptation. Call that 2027.

The models that have been released in 2025 will make this far more tractable in the near future. Reasoning models have made in-context learning far more powerful, resulting in rapid progress on held-out and complex domains such as ARC-AGI. These models also have come with massive improvements in context length. Claude and Gemini have 1M+ token context lengths and GPT-5’s is at 400K — they’re all growing steadily. What is important with the context length numbers is that evaluations are showing that these are meaningful improvements that the models can leverage intelligently.

With these reasoning models and smart retrieval of context, the systems we are building will look indistinguishable from continual learning. This will definitely be multiple LLMs working together and will operate very differently than the first versions of ChatGPT we were given (and often still use today).

The path to continual learning is more context and more horsepower. This is directly in line with the direction AI investment is going. This doesn’t feel like a bottleneck, rather another product problem that we are going to solve. This sort of continual learning may not enable the type of raw intelligence and autonomy that many vocal leaders in AI describe as “superintelligence.”

Training models to be smarter on even more complex tasks — e.g. novel biological research — requires mastering agentic behaviors that need to be learned from scratch, as discussed in my post on “What comes next with RL”. There’s no internet scale pretraining data for such agentic tasks. My point is that not all jobs that require continual learning will require the frontiers of intelligence. I’m excited to write blog posts with the bliss of my ChatGPT 6 co-editor.

This technology coming soon will not be without its challenges. My first reaction to the continual learning post was more in line with “society isn’t ready for this” rather than commentary on its feasibility. I’ll repeat my warning:

For a long time I’ve written that AI models have a higher risk potential in terms of social outcomes because the modalities they interact with us in are far more personal… As AI is going to be so powerful as a standalone entity, breaking some of the symbiotic links will be good for adding friction that makes the technology easier to steer towards good outcomes. In short, be wary of wishing for end-to-end (reinforcement) learning when you’re part of the environment.2 It’s a destiny to dystopia.

What we have today is a form of AGI and it’ll soon get much better with better context and memory. The industrialization of language models is giving us incredible improvements across a wide swath of use-cases. These will blow past many basic primitives of intelligence in humans that have motivated AI for decades. First was models reasoning, then will come systems with continual learning. This is exactly what most AI companies are actually building — regardless of what their superintelligence messaging is.

Comments are open on this post, please continue the debate!

If you want a video version of this, check out the last 20 minutes of the livestream reaction (edit, fixed link) I did with Will Brown of Prime Intellect and Swyx of Smol AI & Latent Space.

GPT-5 was set up to fail on some of the narratives it was expected to satisfy. The two central themes it had to decide between were the AGI (or superintelligence) narrative that Sam Altman & co. have been using to fundraise and the fact that ChatGPT is one of the fastest-growing consumer technologies of all time.

To fulfill both, GPT-5 needed to be AGI while also being cheap enough to serve as the most-used AI system in the world. Business and technological realities made it inevitable that GPT-5’s primary impact would be to solidify OpenAI’s market position, even if it raises a lot of eyebrows for the long-term trajectory of AI.

The reactions online capture this as well. The OpenAI live streams have historically catered to AI insiders, but the product speaks entirely to a different audience. The people discussing this release on Twitter will be disappointed in a first reaction, but 99% of people using ChatGPT are going to be so happy about the upgrade. Confusingly enough, this includes many of the critics. GPT-5 is a good AI system. It’s right in line with best-in-class across pretty much every evaluation, while being cheap enough to serve the whole world.

OpenAI is largely fixing its product offering with an announcement that was hyped to be one of the biggest AI news cycles of the year. AI news being loud is defined by narratives being different more-so than technology being better. OpenAI releasing an open model again will likely be pinpointed as just as important a day for the arc of AI as the GPT-5 release. In many ways GPT-5 was set up to fail and that is very off-putting for those expecting maximum AI progress in the near term.

I’m not going to dwell on it, but oh boy, that was a messy release. GPT-5 being announced and rolled out like this is very odd. Countless plots were mislabeled, live demos had bugs, and the early rollout is doing some weird stuff. This reinforces how OpenAI was torn about the release and backed into a corner with their messaging. They knew they needed to improve the experience with strong competition in the industry, but releasing GPT-5 needed to make a splash after how long they’ve waited (and already parked the GPT 4.5 name).

The core question we track in this post is: What does it mean for the next 6-18 months of AI progress if GPT-5 is just as good as all the best models out there, e.g., Claude Sonnet for coding or o3 for search, funneled into one, super cheap package?

If AGI was a real goal, the main factor on progress would be raw performance. GPT-5 shows that AI is on a somewhat more traditional technological path, where there isn’t one key factor, it is a mix of performance, price, product, and everything in between.

Interconnects is a reader-supported publication. Consider becoming a subscriber.

GPT-5’s performance

There are a few places that we can see that GPT-5 represents a solid step on the performance trend line, but nothing like a step change. First, on LMArena, GPT-5 is fantastic, sweeping the board to #1 on all categories. The last model to claim #1 in pretty much every category was Gemini 2.5 Pro — and that was the biggest step change in Elo since GPT-4 Turbo skyrocketed past the first Claude.

Second, GPT-5 is the top model on the ArtificialAnalysis composite benchmark.

These two, LMArena & ArtificialAnalysis, represent two coarse evaluations — community vibes and raw benchmarks. Both of these can be gamed, but are still correlated with real-world use. You can also see in OpenAI’s shared results how much the smaller versions improve on the likes of GPT-4.1 mini and o4-mini.

In many ways, the march of progress on evals has felt slowed for a while because model releases are so frequent and each individual step is smaller. Lots of small steps make for big change. The overall trend line is still very positive, and multiple companies are filling in the shape of it.

My post on “what comes next” from earlier this summer all but called this type of release, where the numbers aren’t shocking but the real world use cases are great, becoming more common.

This is a different path for the industry and will take a different form of messaging than we’re used to. More releases are going to look like Anthropic’s Claude 4, where the benchmark gains are minor and the real world gains are a big step. There are plenty of more implications for policy, evaluation, and transparency that come with this. It is going to take much more nuance to understand if the pace of progress is continuing, especially as critics of AI are going to seize the opportunity of evaluations flatlining to say that AI is no longer working.

To say it succinctly: Abilities will develop more slowly than products.

The product overhang is being extended with each release. We’re still building untapped value with AI models and systems faster than we’re capturing it.

Another way to see this incremental push out in models or systems is through OpenAI’s update to the famous METR plot of time to completion for humans of various tasks AI systems can solve 50% of the time. GPT-5 is leading, but also just in line with trends.

All of this is to say comprehensively that AI progress is very alive and well, as long as you don’t subscribe to the exponential takeoff in ability. Those arguments are very strained by this GPT-5 release.

Yes, AI progress on intelligence and “raw ability” is certainly going to continue at a solid pace for a long time, but how will this translate into recursive self-improvement?

GPT-5’s details

If you’re reading closely, you may have noticed that this post uses the word system instead of model. All of the leading chat systems have been adding more components onto them like safety checkers and so on, but this is the first one to use different architectures and weights for the primary generation of content across similar queries. GPT-5 is the first in what is to come, mostly to better balance cost and give better user experiences. From the system card:

GPT‑5 is a unified system with a smart and fast model that answers most questions, a deeper reasoning model for harder problems, and a real-time router that quickly decides which model to use based on conversation type, complexity, tool needs, and explicit intent (for example, if you say “think hard about this” in the prompt). The router is continuously trained on real signals, including when users switch models, preference rates for responses, and measured correctness, improving over time.

Along with this, they shipped many product improvements, such as how the model has a 400K context window in the API with great performance, reduced hallucinations, and new personalities.

Primarily, I worry as a power user about the router. I sense that for now I’ll default to GPT-5 Thinking, and sometimes upgrade to Pro mode, while downgrading to standard GPT-5 only for benign queries (depending on its search behavior — if it is search-heavy like o3 without thinking, then it should still work well).

Thankfully, the thinking mode has a “get an early answer” button, so I don’t see any reason to start elsewhere. If I need an answer fast, I’ll get one. If not, I want the best responses possible.

As for prices, here’s a comparison. GPT-5’s top-level model is cheaper than Claude Sonnet and far better than any OpenAI model has been before at coding — one of the core details of this release. Matching Gemini Pro’s pricing when considering Google’s infrastructure advantage is a substantial accomplishment.

* OpenAI — GPT-5 (API sizes)

* GPT-5: input $1.25, output $10.00. (OpenAI)

* GPT-5 mini: input $0.25, output $2.00. (OpenAI)

* GPT-5 nano: input $0.05, output $0.40. (OpenAI)

* OpenAI — o3 (reasoning)

* o3: input $2.00, output $8.00. (OpenAI Platform)

* o3-mini: input $1.10, output $4.40. (cached input $0.55) (OpenAI Platform)

* Anthropic — Claude 4 family

* Claude Sonnet 4: input $3.00, output $15.00. (Anthropic)

* Claude Opus 4.1: input $15.00, output $75.00. (Anthropic)

* Google — Gemini 2.5

* Gemini 2.5 Pro: input $1.25 (≤200k prompt) / $2.50 (>200k); output $10.00 (≤200k) / $15.00 (>200k). (Google AI for Developers)

* Gemini 2.5 Flash: input $0.30 (text/image/video) or $1.00 (audio); output $2.50 (includes thinking tokens). (Google AI for Developers)

* Gemini 2.5 Flash-Lite: input $0.10 (text/image/video) or $0.30 (audio); output $0.40. (Google AI for Developers)

Cheaper, thinking models that work well in applications are far more useful than scaling (as GPT-4.5 has shown us).

GPT-5’s impact

It seems like most people in all walks of life are going to love this model — from AI researchers all the way to people who are learning of ChatGPT for the first time today. This is very in line with my expectations for how AI will proceed, as a long, steady march of progress.

The fact that the models are getting way cheaper rather than way more expensive definitely signals that we cannot just brute-force scale our way to much stronger systems. Scaling helps, but it is now one of many considerations, and all the laboratories are showing us that much bigger models have diminishing returns in value to customers. At the same time, models being cheaper could be just what we need for Jevons paradox to kick in and provide another boost in AI adoption.

Many people will claim that the GPT-5 release was a flop and the bubble will pop for AI. This is downstream of the industry generally making totally unrealistic promises. As someone whose core through-line when covering frontier models is tracking the pace of progress, I translate this as “AI capabilities on benchmarks will proceed a bit more slowly, but we aren’t reaching any clear walls in performance.” The AI performance hills we’re climbing up as an industry do put up some more resistance as the obvious low hanging fruit is gone, but we have the tools to overcome it consistently for the next 6 to 18 months.

For companies that have been fundraising on promises of AGI, such as Anthropic and OpenAI, closing the next rounds could be harder. Of course, this depends on whether the messaging of the rounds was a key part of the fundraising.

This fundraising inspires capital expenditures across the industry, e.g. TSMC developing the next node for NVIDIA to build new chips, and so on. The AGI narrative and the fundraising it has enabled have been good for the U.S. in terms of building out valuable, raw infrastructure.

This could be the beginning of the money train slowing down, but that’s very different from a derailment and a stock market crash. As raw infrastructure spend slows, there will be even more pressure to deliver valuable products to users. A key trend for 2025 has been many of those appearing — Deep Research and Claude Code being the paradigms that everyone has copied.

GPT-5 makes these applications better and makes it easier and cheaper for the next viral AI products to hit the market. I’m still excited for what is to come.

But first, I’m going to sign off and go play with GPT-5. It’s a good day to build something for the fun of it. As I use it more, I’ll have more to say.

Extra GPT-5 links

For more specifics on the model from people who got early access, I recommend Tyler Cowen, Every.to, or Simon Willison (or Swyx soon, on Latent.Space).

Livestream link: https://openai.com/gpt-5/ Research blog post: https://openai.com/index/introducing-gpt-5/ Developer blog post: https://openai.com/index/introducing-gpt-5-for-developers Enterprise blog post: https://openai.com/index/gpt-5-new-era-of-work GPT-5 landing page: https://openai.com/gpt-5/ System Card: https://openai.com/index/gpt-5-system-card/ Coding examples: https://openai.github.io/gpt-5-coding-examples/What would you say if you could talk to a future OpenAI model https://progress.openai.com/

Finally, I’ll plug again the video I did with Will Brown and Swyx:

Send me the most interesting things you find on GPT-5!