Whether you work best solo or thrive in a team, you know collaboration is complicated — and it turns out AI agents face the same tensions. This episode dives into multi-agent systems, exploring how networks of AI agents can overcome the individual limitations of a single model, and what the research says about when collaboration actually helps versus when it just adds noise. Think scaling laws, but for teamwork. --- Website: https://lineardigressions.com Apple Podcasts: https://podcasts.apple.com/us/podcast/linear-digressions/id941219323 Spotify: https://open.spotify.com/show/1JdkD0ZoZ52KjwdR0b1WoT Substack: https://substack.com/@lineardigressions

Knowing when an AI agent has failed sounds straightforward — until it isn't. Agents have a frustrating habit of finishing confidently while quietly doing the wrong thing, or looping endlessly without ever crashing in an obvious way. This episode tackles one of the thorniest problems in the agentic world: evaluation. If failure is hard to see, how do you measure it systematically? And how do you know when your agent is actually working?

How do you know if a new AI model is actually better than the last one? It turns out answering that question is a lot messier than it sounds. This week we dig into the world of LLM benchmarks — the standardized tests used to compare models — exploring two canonical examples: MMLU, a 14,000-question multiple choice gauntlet spanning medicine, law, and philosophy, and SWE-bench, which throws real GitHub bugs at models to see if they can fix them. Along the way: Goodhart's Law, data contamination, canary strings, and why acing a test isn't always the same as being smart.

When you build a model for natural language processing (NLP), such as a recurrent neural network, it helps a ton if you’re not starting from zero. In other words, if you can draw upon other datasets for building your understanding of word meanings, and then use your training dataset just for subject-specific refinements, you’ll get farther than just using your training dataset for everything. This idea of starting with some pre-trained resources has an analogue in computer vision, where initializations from ImageNet used for the first few layers of a CNN have become the new standard. There’s a similar progression under way in NLP, where simple(r) embeddings like word2vec are giving way to more advanced pre-processing methods that aim to capture more sophisticated understanding of word meanings, contexts, language structure, and more. Relevant links: https://thegradient.pub/nlp-imagenet/

Facial recognition being used in everyday life seemed far-off not too long ago. Increasingly, it’s being used and advanced widely and with increasing speed, which means that our technical capabilities are starting to outpace (if they haven’t already) our consensus as a society about what is acceptable in facial recognition and what isn’t. The threats to privacy, fairness, and freedom are real, and Microsoft has become one of the first large companies using this technology to speak out in specific support of its regulation through legislation. Their arguments are interesting, provocative, and even if you don’t agree with every point they make or harbor some skepticism, there’s a lot to think about in what they’re saying. https://blogs.microsoft.com/on-the-issues/2018/12/06/facial-recognition-its-time-for-action/

Bringing you another old classic this week, as we gear up for 2019! See you next week with new content. Word2Vec is probably the go-to algorithm for vectorizing text data these days.  Which makes sense, because it is wicked cool.  Word2Vec has it all: neural networks, skip-grams and bag-of-words implementations, a multiclass classifier that gets swapped out for a binary classifier, made-up dummy words, and a model that isn't actually used to predict anything (usually).  And all that's before we get to the part about how Word2Vec allows you to do algebra with text.  Seriously, this stuff is cool.

We’re taking a break for the holidays, chilling with the dog and an eggnog (Katie) and the cat and some spiced cider (Ben). Here’s an episode from a while back for you to enjoy. See you again in 2019! You might sometimes find that it's hard to get started doing something, but once you're going, it gets easier. Turns out machine learning algorithms, and especially recommendation engines, feel the same way. The more they "know" about a user, like what movies they watch and how they rate them, the better they do at suggesting new movies, which is great until you realize that you have to start somewhere. The "cold start" problem will be our focus in this episode, both the heuristic solutions that help deal with it and a bit of realism about the importance of skepticism when someone claims a great solution to cold starts.

Convex optimization is one of the keys to data science, both because some problems straight-up call for optimization solutions and because popular algorithms like a gradient descent solution to ordinary least squares are supported by optimization techniques. But there are all kinds of subtleties, starting with convex and non-convex functions, why gradient descent is really an optimization problem, and what that means for your average data scientist or statistician.

When you think about it, it’s pretty amazing that we can draw conclusions about huge populations, even the whole world, based on datasets that are comparatively very small (a few thousand, or a few hundred, or even sometimes a few dozen). That’s the power of statistics, though. This episode is kind of a two-for-one but we’re excited about it—first we’ll talk about the Normal or Gaussian distribution, which is maybe the most famous probability distribution function out there, and then turn to the Central Limit Theorem, which is one of the foundational tenets of statistics and the real reason why the Normal distribution is so important.

Neural nets are a way you can model a system, sure, but if you take a step back, squint, and tilt your head, they can also be called… software? Not in the sense that they’re written in code, but in the sense that the neural net itself operates under the same set of general requirements as does software that a human would write. Namely, neural nets take inputs and create outputs from them according to a set of rules, but the thing about the inside of the neural net black box is that it’s written by a computer, whereas the software we’re more familiar with is written by a human. Neural net researcher and Tesla director of AI Andrej Karpathy has taken to calling neural nets “Software 2.0” as a result, and the implications from this connection are really cool. We’ll talk about it this week. Relevant links: https://medium.com/@karpathy/software-2-0-a64152b37c35

Deep neural nets have a deserved reputation as the best-in-breed solution for computer vision problems. But there are many aspects of human vision that we take for granted but where neural nets struggle—this episode covers an eye-opening paper that summarizes some of the interesting weak spots of deep neural nets. Relevant links: https://arxiv.org/abs/1805.04025

At many places, data scientists don’t work solo anymore—it’s a team sport. But data science teams aren’t simply teams of data scientists working together. Instead, they’re usually cross-functional teams with engineers, managers, data scientists, and sometimes others all working together to build tools and products around data science. This episode talks about some of those roles on a typical data science team, what the responsibilities are for each role, and what skills and traits are most important for each team member to have.

Last week’s episode, about methods for optimized web crawling logic, left off on a bit of a cliffhanger: the data scientists had found a solution to the problem, but it wasn’t something that the engineers (who own the search codebase, remember) liked very much. It was black-boxy, hard to parallelize, and introduced a lot of complexity to their code. This episode takes a second crack, where we formulate the problem a little differently and end up with a different, arguably more elegant solution. Relevant links: http://www.unofficialgoogledatascience.com/2018/07/by-bill-richoux-critical-decisions-are.html http://www.csc.kth.se/utbildning/kth/kurser/DD3364/Lectures/KKT.pdf

The paperclip maximizer — the classic AI doom scenario where a hyper-competent machine single-mindedly converts the universe into office supplies — might not be the AI risk we should actually lose sleep over. New research from Anthropic's AI safety division suggests misaligned AI looks less like an evil genius and more like a distracted wanderer who gets sidetracked reading French poetry instead of, say, managing a nuclear power plant. This week we dig into a fascinating paper reframing AI misalignment through the lens of bias-variance decomposition, and why longer reasoning chains might actually make things worse, not better. - "The Hot Mess Theory of AI Misalignment: How Misalignment Scales with Model Intelligence and Task Complexity" — Anthropic AI Safety. https://arxiv.org/abs/2503.08941

Got a fun optimization problem for you this week! It’s a two-for-one: how do you optimize the web crawling logic of an operation like Google search so that the results are, on average, as up-to-date as possible, and how do you optimize your solution of choice so that it’s maintainable by software engineers in a huge distributed system? We’re following an excellent post from the Unofficial Google Data Science blog going through this problem. Relevant links: http://www.unofficialgoogledatascience.com/2018/07/by-bill-richoux-critical-decisions-are.html

The Poisson distribution is a probability distribution function used to for events that happen in time or space. It’s super handy because it’s pretty simple to use and is applicable for tons of things—there are a lot of interesting processes that boil down to “events that happen in time or space.” This episode is a quick introduction to the distribution, and then a focus on two of our favorite applications: using the Poisson distribution to identify supernovas and study army deaths from horse kicks.

If you wanted to find a dataset of jokes, how would you do it? What about a dataset of podcast episodes? If your answer was “I’d try Google,” you might have been disappointed—Google is a great search engine for many types of web data, but it didn’t have any special tools to navigate the particular challenges of, well, dataset data. But all that is different now: Google recently announced Google Dataset Search, an effort to unify metadata tagging around datasets and complementary efforts on the search side to recognize and organize datasets in a way that’s useful and intuitive. So whether you’re an academic looking for an economics or physics or biology dataset, or a big old nerd modeling jokes or analyzing podcasts, there’s an exciting new way for you to find data.

We started Linear Digressions 4 years ago… this isn’t a technical episode, just two buddies shooting the breeze about something we’ve somehow built together.

This week, we’re dusting off the ol’ particle physics PhD to bring you an episode about ambitious new model-agnostic searches for new particles happening at CERN. Traditionally, new particles have been discovered by “targeted searches,” where scientists have a hypothesis about the particle they’re looking for and where it might be found. However, with the huge amounts of data coming out of CERN, a new type of broader search algorithm is starting to be deployed. It’s a strategy that casts a very wide net, looking in many different places at the same time, which also introduces all kinds of interesting questions—even a one-in-a-thousand occurrence happens when you’re looking in many thousands of places.

If you’re a data scientist, you know how important it is to keep your data orderly, clean, moving smoothly between different systems, well-documented… there’s a ton of work that goes into building and maintaining databases and data pipelines. This job, that of owner and maintainer of the data being used for analytics, is often the realm of data engineers. From data extraction, transform and loading procedures to the data storage strategy and even the definitions of key data quantities that serve as focal points for a whole organization, data engineers keep the plumbing of data analytics running smoothly.

A very intriguing op-ed was published in the NY Times recently, in which the author (a senior official in the Trump White House) claimed to be a minor saboteur of sorts, acting with his or her colleagues to undermine some of Donald Trump’s worst instincts and tendencies. Pretty stunning, right? So who is the author? It’s a mystery—the op-ed was published anonymously. That hasn’t stopped people from speculating though, and some machine learning on the vocabulary used in the op-ed is one way to get clues.

If you've been in data science for more than a year or two, chances are you've noticed changes in the field as it's grown and matured. And if you're newer to the field, you may feel like there's a disconnect between lots of different stories about what data scientists should know, or do, or expect from their job. This week, we cover two thought pieces, one that arose from interviews with 35(!) data scientists speaking about what their jobs actually are (and aren't), and one from the head of data science at AirBnb organizing core data science work into three main specialties. Relevant links: https://hbr.org/2018/08/what-data-scientists-really-do-according-to-35-data-scientists https://www.linkedin.com/pulse/one-data-science-job-doesnt-fit-all-elena-grewal

There's just too much interesting stuff at the intersection of agile software development and data science for us to be able to cover it all in one episode, so this week we're picking up where we left off last time. We'll give a quick overview of agile for those who missed last week or still have some questions, and then cover some of the aspects of agile that don't work well out-of-the-box when applied to data analytics. Fortunately, though, there are some straightforward modifications to agile that make it work really nicely for data analytics! Relevant links: https://www.agilealliance.org/agile101/12-principles-behind-the-agile-manifesto/ https://www.locallyoptimistic.com/post/agile-analytics-p1/ https://www.locallyoptimistic.com/post/agile-analytics-p2/ https://www.locallyoptimistic.com/post/agile-analytics-p3/

If you're a data scientist at a firm that does a lot of software building, chances are good that you've seen or heard engineers sometimes talking about "agile software development." If you don't work at a software firm, agile practices might be newer to you. In either case, we wanted to go through a great series of blog posts about some of the practices from agile that are relevant for how data scientists work, in hopes of inspiring some transfer learning from software development to data science. Relevant links: https://www.locallyoptimistic.com/post/agile-analytics-p1/ https://www.locallyoptimistic.com/post/agile-analytics-p2/ https://www.locallyoptimistic.com/post/agile-analytics-p3/

Every AI builder knows the anxiety: you spend months engineering prompts, tuning pipelines, and chaining calls together — then a new model drops and half your work evaporates overnight. It turns out researchers have been wrestling with this exact dynamic for 30 years, and they keep arriving at the same uncomfortable answer. That answer is called the Bitter Lesson — and understanding it might be the most important thing you can do for whatever you're building right now. From Deep Blue to AlexNet to modern LLMs, scale keeps beating sophistication, and knowing which side of that line your work falls on makes all the difference. Links - Richard Sutton, "The Bitter Lesson" - Alon Halevy, Peter Norvig, and Fernando Pereira, "The Unreasonable Effectiveness of Data" - Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, "ImageNet Classification with Deep Convolutional Neural Networks"

We've got a classic for you this week as we take a week off for the dog days of summer. See you again next week! Competing in a machine learning competition on Kaggle is a kind of rite of passage for data scientists. Losing unexpectedly at the very end of the contest is also something that a lot of us have experienced. It's not just bad luck: a very specific combination of overfitting on popular competitions can take someone who is in the top few spots in the final days of a contest and bump them down hundreds of slots in the final tally.

There's a lot of great machine learning papers coming out every day--and, if we're being honest, some papers that are not as great as we'd wish. In some ways this is symptomatic of a field that's growing really quickly, but it's also an artifact of strange incentive structures in academic machine learning, and the fact that sometimes machine learning is just really hard. At the same time, a high quality of academic work is critical for maintaining the reputation of the field, so in this episode we walk through a recent paper that spells out some of the most common shortcomings of academic machine learning papers and what we can do to make things better. Relevant links: https://arxiv.org/abs/1807.03341

The stars aligned for me (Katie) this past weekend: I raced my first half-marathon in a long time and got to read a great article from the NY Times about a new running shoe that Nike claims can make its wearers run faster. Causal claims like this one are really tough to verify, because even if the data suggests that people wearing the shoe are faster that might be because of correlation, not causation, so I loved reading this article that went through an analysis of thousands of runners' data in 4 different ways. Each way has a great explanation with pros and cons (as well as results, of course), so be sure to read the article after you check out this episode! Relevant links: https://www.nytimes.com/interactive/2018/07/18/upshot/nike-vaporfly-shoe-strava.html

When you're using an AB test to understand the effect of a treatment, there are a lot of assumptions about how the treatment (and control, for that matter) get applied. For example, it's easy to think that everyone who was assigned to the treatment arm actually gets the treatment, everyone in the control arm doesn't, and that the two groups get their treatment instantaneously. None of these things happen in real life, and if you really care about measuring your treatment effect then that's something you want to understand and correct. In this post we'll talk through a great blog post that outlines this for mobile experiments. Oh, and Ben sings.

Artificial Intelligence has been widely lauded as a solution to almost any problem. But as we justapose the hype in the field against the real-world benefits we see, it raises the question: Are we coming up on an AI winter

We like learning on vacation. And we're on vacation, so we thought we'd re-air this episode about how to learn. Original Episode: https://lineardigressions.com/episodes/2017/5/14/how-to-find-new-things-to-learn Original Summary: If you're anything like us, you a) always are curious to learn more about data science and machine learning and stuff, and b) are usually overwhelmed by how much content is out there (not all of it very digestible).  We hope this podcast is a part of the solution for you, but if you're looking to go farther (who isn't?) then we have a few new resources that are presenting high-quality content in a fresh, accessible way.  Boring old PDFs full of inscrutable math notation, your days are numbered!

We're on vacation on Mars, so we won't be communicating with you all directly this week. Though, if we wanted to, we could probably use this episode to help get started. Original Episode: http://lineardigressions.com/episodes/2017/3/19/space-codes Original Summary: It's hard to get information to and from Mars.  Mars is very far away, and expensive to get to, and the bandwidth for passing messages with Earth is not huge.  The messages you do pass have to traverse millions of miles, which provides ample opportunity for the message to get corrupted or scrambled.  How, then, can you encode messages so that errors can be detected and corrected?  How does the decoding process allow you to actually find and correct the errors?  In this episode, we'll talk about three pieces of the process (Reed-Solomon codes, convolutional codes, and Viterbi decoding) that allow the scientists at NASA to talk to our rovers on Mars.

We're on vacation, so we hope you enjoy this episode while we each sip cocktails on the beach. Original Episode: http://lineardigressions.com/episodes/2017/6/18/anscombes-quartet Original Summary: Anscombe's Quartet is a set of four datasets that have the same mean, variance and correlation but look very different.  It's easy to think that having a good set of summary statistics (like mean, variance and correlation) can tell you everything important about a dataset, or at least enough to know if two datasets are extremely similar or extremely different, but Anscombe's Quartet will always be standing behind you, laughing at how silly that idea is. Anscombe's Quartet was devised in 1973 as an example of how summary statistics can be misleading, but today we can even do one better: the Datasaurus Dozen is a set of twelve datasets, all extremely visually distinct, that have the same summary stats as a source dataset that, there's no other way to put this, looks like a dinosaur.  It's an example of how datasets can be generated to look like almost anything while still preserving arbitrary summary statistics.  In other words, Anscombe's Quartets can be generated at-will and we all should be reminded to visualize our data (not just compute summary statistics) if we want to claim to really understand it.

Now that hurricane season is upon us again (and we are on vacation), we thought a look back on our hurricane forecasting episode was prudent. Stay safe out there.

By now, you have probably heard of GDPR, the EU's new data privacy law. It's the reason you've been getting so many emails about everyone's updated privacy policy. In this episode, we talk about some of the potential ramifications of GRPD in the world of data science.

From Atari to ChatGPT: How AI Learned to Follow Instructions by Katie Malone

If you're a data scientist, chances are good that you've heard of git, which is a system for version controlling code. Chances are also good that you're not quite as up on git as you want to be--git has a strong following among software engineers but, in our anecdotal experience, data scientists are less likely to know how to use this powerful tool. Never fear: in this episode we'll talk through some of the basics, and what does (and doesn't) translate from version control for regular software to version control for data science software.

Data science and analytics are hot topics in business these days, but for a lot of folks looking to bring data into their organization, it can be hard to know where to start and what it looks like when they're succeeding. That was the motivation for writing a whitepaper on the analytics maturity of an organization, and that's what we're talking about today. In particular, we break it down into five attributes of an organization that contribute (or not) to their success in analytics, and what each of those mean and why they matter.  Whitepaper here: bit.ly/analyticsmaturity

Shapley values in machine learning are an interesting and useful enough innovation that we figured hey, why not do a two-parter? Our last episode focused on explaining what Shapley values are: they define a way of assigning credit for outcomes across several contributors, originally to understand how impactful different actors are in building coalitions (hence the game theory background) but now they're being cross-purposed for quantifying feature importance in machine learning models. This episode centers on the computational details that allow Shapley values to be approximated quickly, and a new package called SHAP that makes all this innovation accessible.

As machine learning models get into the hands of more and more users, there's an increasing expectation that black box isn't good enough: users want to understand why the model made a given prediction, not just what the prediction itself is. This is motivating a lot of work into feature important and model interpretability tools, and one of the most exciting new ones is based on Shapley Values from game theory. In this episode, we'll explain what Shapley Values are and how they make a cool approach to feature importance for machine learning.

If you were a machine learning researcher or data scientist ten years ago, you might have spent a lot of time implementing individual algorithms like decision trees and neural networks by hand. If you were doing that work five years ago, the algorithms were probably already implemented in popular open-source libraries like scikit-learn, but you still might have spent a lot of time trying different algorithms and tuning hyperparameters to improve performance. If you're doing that work today, scikit-learn and similar libraries don't just have the algorithms nicely implemented--they have tools to help with experimentation and hyperparameter tuning too. Automated machine learning is here, and it's pretty cool.

A huge part of the ascent of deep learning in the last few years is related to advances in computer hardware that makes it possible to do the computational heavy lifting required to build models with thousands or even millions of tunable parameters. This week we'll pretend to be electrical engineers and talk about how modern machine learning is enabled by hardware.

Last episode we talked conceptually about capsule networks, the latest and greatest computer vision innovation to come out of Geoff Hinton's lab. This week we're getting a little more into the technical details, for those of you ready to have your mind stretched.

Convolutional nets are great for image classification... if this were 2016. But it's 2018 and Canada's greatest neural networker Geoff Hinton has some new ideas, namely capsule networks. Capsule nets are a completely new type of neural net architecture designed to do image classification on far fewer training cases than convolutional nets, and they're posting results that are competitive with much more mature technologies. In this episode, we'll give a light conceptual introduction to capsule nets and get geared up for a future episode that will do a deeper technical dive.

If you've done image recognition or computer vision tasks with a neural network, you've probably used a convolutional neural net. This episode is all about the architecture and implementation details of convolutional networks, and the tricks that make them so good at image tasks.

It's been a nasty flu season this year. So we were remembering a story from a few years back (but not covered yet on this podcast) about when Google tried to predict flu outbreaks faster than the Centers for Disease Control by monitoring searches and looking for spikes in searches for flu symptoms, doctors appointments, and other related terms. It's a cool idea, but after a few years turned into a cautionary tale of what can go wrong after Google's algorithm systematically overestimated flu incidence for almost 2 years straight. Relevant link: https://gking.harvard.edu/publications/parable-google-flu%C2%A0traps-big-data-analysis

Today we are going to talk about the feature with the worst acronym in generative AI: RAG, or Retrieval Augmented Generation. If you've ever used something like "Chat with My Docs," if you have an internal AI chatbot that has access to your company's documents, or you've created one yourself on some kind of personal project and uploaded a bunch of documents for the AI to use — you have encountered RAG, whether you know it or not. It's an extremely effective technique. Works super well for taking general purpose models like ChatGPT or Claude and turning them into AIs that are aware of all the specific information that makes them truly useful in a huge variety of situations. RAG is pretty interesting under the hood, so I thought it would be fun to spend a little while talking about it. You are listening to Linear Digressions. RAG was first introduced in this paper from Facebook Research in 2021: https://arxiv.org/pdf/2005.11401

This week's episodes is for data scientists, sure, but also for data science managers and executives at companies with data science teams. These folks all think very differently about the same question: what should a data science team be working on? And how should that decision be made? That's the subject of a talk that I (Katie) gave at Strata Data in early March, about how my co-department head and I select projects for our team to work on. We have several goals in data science project selection at Civis Analytics (where I work), which can be summarized under "balance the best attributes of bottom-up and top-down decision-making." We achieve this balance, or at least get pretty close, using a process we've come to call the Idea Factory (after a great book about Bell Labs). This talk is about that process, how it works in the real world of a data science company and how we see it working in the data science programs of other companies. Relevant links: https://conferences.oreilly.com/strata/strata-ca/public/schedule/detail/63905

Autoencoders are neural nets that are optimized for creating outputs that... look like the inputs to the network. Turns out this is a not-too-shabby way to do unsupervised machine learning with neural nets.

After all the back-patting around making data science datasets and code more openly available, we figured it was time to also dump a bucket of cold water on everyone's heads and talk about the things that can go wrong when data and code is a little too open. In this episode, we'll talk about two interesting recent examples: a de-identified medical dataset in Australia that was re-identified so specific celebrities and athletes could be matched to their medical records, and a series of military bases that were spotted in a public fitness tracker dataset.