Hallway Chat – Details, episodes & analysis

In this episode of Hallway Chat, Nabeel and Fraser explore the future of AI through the lenses of memory, context, and product design. They debate whether users will maintain separate AIs for work and home, why visually verifiable output is critical to product success, and how startups should behave in a world where the best models become commodities. From the case for uneconomical AI to the trade-off between 500 average agents and one brilliant one, they unpack what it takes to build great AI products—and when, if ever, VCs should preempt funding rounds.

• (00:00) - Chapter 1
• (01:54) - Will AI Democratize Like the iPhone?
• (03:03) - Should You Burn Tokens to Win?
• (06:25) - The Case for Going Big with Compute
• (09:09) - 500 Agents or One Genius?
• (18:19) - The Importance of Proof of Work
• (19:49) - Building the Right UI for Agents
• (22:23) - Visualizing AI Output: A Design Challenge
• (33:37) - Memory Wars: Work AI vs. Personal AI
• (35:01) - Consumer Benefit > Privacy?
• (37:36) - Whats at stake, your memory
• (45:28) - Model Switching as a Feature, Not a Bug
• (53:53) - VC Talk: When to Preempt a Round

Nabeel and Fraser debrief on GDC and the opportunities for innovation beyond the conservative approaches in an industry going through a funk. They touch on the experiences at the recent GDC, including the growth of AI in gaming and the recurring challenges faced by developers. The discussion also covers the competitive landscape of AI startups, the importance of stacking wins, and the nuanced differences between building games and software products. Notably, they highlight the intriguing possibility of a vibrant game creation ecosystem driven by new model capabilities and the compelling idea of making game creation itself a multiplayer experience.

• (00:00) - Introduction: Bridging the Gap in the Gaming Industry
• (00:35) - GDC Week Highlights and Reflections
• (01:41) - Challenges and Innovations in AI and Gaming
• (03:50) - The Broken Ecosystem of Game Distribution
• (07:45) - The Future of Game Development and Creativity
• (19:59) - The Future of Multiplayer Coding
• (21:09) - The Magic of New Capabilities
• (22:07) - From Capabilities to Durable Products
• (23:39) - Stacking Wins for Long-Term Success
• (33:54) - Innovations in the Gaming Industry

Fraser and Nabeel start by retcon'ing their earlier "maybe we should stop talking about models" and reflecting on the importance of adapting product strategy based how much disruption is happening in your market. The episode highlights the release of Claude Artifacts by Anthropic, and comparing it to other AI product innovations. They also nerd out on the Xbloom coffee machine and a robotic mop. Lastly, Fraser asks about how to navigate when a founder gets introduced to the wrong partner at a VC firm.

• (00:00) - Product development early on the S curve, exploring Anthropic Artifacts
• (00:43) - Welcome to Hallway Chat
• (00:58) - Strong Convictions Loosely Held
• (03:37) - Exploring Claude Artifacts
• (10:48) - S Curve Product Design
• (23:33) - The Instinct to Exhale: Post-Launch Reflections
• (25:25) - Balancing Innovation and Stability
• (27:00) - Leaning into risk
• (32:30) - Question: Getting introd to the wrong partner at a VC

Fraser and Nabeel dive into the evolving landscape of AI note-taking apps, focusing on a new product, Granola. They discuss Granola's unique approach of "enhancing your thought" instead of "thinking for you." They also explore: applying this idea to other categories, the influence of LLMs on UX, and the need for more experimental user workflows to unlock AI. They also discuss the challenges and dynamics of startup funding, the significance of capital in achieving market success, and how dynamics of AI team building are changing. 

• (00:00) - Introduction
• (01:05) - Introducing Granola: The AI Note-Taking App
• (02:19) - Granola's Unique Features and User Experience
• (04:54) - Challenges and Innovations in AI Note-Taking
• (06:53) - Granola's Practical Use Cases and Feedback
• (09:54) - Reducing Friction in Product Experience
• (13:04) - Where is the AI innovation?
• (16:11) - The Role of Incentives in Innovation
• (21:51) - When competitors raise 10x more than you
• (23:11) - Introducing Granola: The AI Note-Taking App
• (23:18) - Granola's Unique Features and User Experience
• (24:35) - Challenges and Innovations in AI Note-Taking
• (25:07) - Granola's Practical Use Cases and Feedback
• (25:19) - Reducing Friction in Product Experience
• (26:44) - Where is the AI innovation?
• (27:12) - The Role of Incentives in Innovation
• (27:59) - When competitors raise 10x more than you
• (28:34) - Overcapitalization and Its Consequences
• (29:21) - Customer-Centric Approach in Startups
• (30:22) - How AI Startups are Building Teams Differently
• (42:41) - Final Thoughts and Product Recommendations

Is it still worth it to discuss models when discussing startups? Nabeel and Fraser discuss how that may be the wrong question to ask in the current landscape, and why customer-centric questions and user experience should be the basis of product experience. Later, they deliberate who might come out on top in the “horse race” for AI product dominance, and whether it will come from a large, established company, or if the frontier of capabilities belongs to small innovators.

• (00:00) - Decoding the Future: Puzzles vs. Mysteries in Tech
• (01:22) - Welcome to Hallway Chat: Podcast or Tweets?
• (01:35) - Exploring the Venture Firm USV's "Hallway Chat" Tweets
• (02:27) - The atomization of media
• (03:17) - Rethinking the Focus on AI Models in Startups
• (04:38) - The Importance of Use Cases Over Models in AI Innovation
• (07:45) - What makes a Foundational Model... Foundational
• (18:23) - AI for Consumers: Navigating the S Curve of Innovation

What is the extent of autonomous coding engineer Devin’s ability to generate real, functional applications with little to no help? Nabeel and Fraser dive into the buzz about Cognition’s ‘Devin’, what makes it different, and the transformative potential of AI in software engineering, particularly focusing on autonomous coding software. Later, they get into the innovator's dilemma and the lessons the .com era can lend to this new time in AI.

(00:00) Intro

(01:23) The buzz about Cognition’s Devin

(07:44) What makes Devin different?

(12:19) Tolerance for time

(16:07) The interface of the future

(22:19) Innovating around the incumbent’s advantage

(25:30) Cutting edge products mean new user bases

(29:52) Netscape was the Open AI of the Mobile Revolution

(33:42) Optimism as the engine of capitalism

(37:56) The model is not the product

This week is a discussion of the points of differentiation vs commodification in various AI models. Including how these points might change over time, and might change between language, image, and video models. Then Fraser pivots to orienting around jobs to be done in AI, and how the various models have a huge gap between being capable of doing something, and doing it well. 

We then talk about Claude 3, the nature of benchmarking, and the rapid dropping LLM prices. Lastly, we cover a startup subject, debating the merits of SAFE (Simple Agreement for Future Equity) vs. priced equity rounds in early-stage funding. 

Links

1. Claude 3 from Anthropic
2. LMSys Chatbot Leaderboard for which models are sticking out right now
3. For areas where models don't differentiate, we are back to the 7 powers framework. Hamilton Helmer's website with overview of the 7 Powers framework: https://7powers.com/, Sachin Rekhi has a good detailed primer on the 7 Powers as well.
4. Carta guide on priced rounds vs SAFEs, which we don't really agree with all the pros and cons, but is a decent overview

• (00:00) - Opening
• (00:31) - Are models an inevitable commodity?
• (05:58) - How image, video, and other models may pan out differently than language
• (08:20) - It's not the model, it's the customer
• (10:20) - Jobs to be done in AI - 1. Can it do it... 2. very well.. 3. exactly how I want it to.
• (22:13) - Claude 3 by Anthropic
• (25:25) - ELO Leaderboard Results
• (26:25) - Claude 3 Haiku and Falling LLM Prices
• (30:33) - SAFES vs priced equity rounds

What is Hallway Chat anyway

• (00:00) - Introduction and Podcast Naming

Josh Miller, Founder and CEO of The Browser Co, joins this week. We loved the experience with Arc Search, a new AI-enabled mobile search experience, we decided to talk to Josh directly about how it all came together. 

Topics
* How Arc Search came to be spawned out of basically a side project
* The core value of "we don't know"
* How the key AI feature 'Browse for Me' was not the initial concept, but came to be added last minute
* Josh's initial skepticism about AI, up until just a few months ago, and what flipped him
* Three strategies for AI integration
* The perils of leading with consumer hooks, how that derailed Browser Co, and focusing on solving problems
* Should cost be a key consideration when developing AI?

Hope you enjoy us mixing it up this week by talking with someone wrestling with building in the space. Drop us a line on twitter at @nabeel and @fraser with what you want to hear more (or less) of. 

• (00:00) - Intro
• (01:02) - Welcome Josh, CEO of The Browser Company
• (02:51) - The origin of Arc Search
• (05:12) - Browse For Me
• (06:55) - Overview of Arc Search
• (16:51) - Thinking about competition
• (25:08) - User problems over elegant hooks
• (29:25) - Three strategies for AI integration
• (34:51) - The importance of prototyping
• (40:28) - The future of desktop search
• (46:01) - Thinking about costs when building with AI

Two former founders, now VCs, play with AI products and see where that leads...

Fraser and Nabeel discuss the differences between horizontal disruption and vertical market disruption and the patterns of the phases of this disruption in the mobile age. From Adaptation, to Evolution, and eventually Revolution. Next, they dive into AI email app Shortwave AI. This leads to a conversation around what AI models are optimizing for, and how speed and polish can sometimes be part of the baseline usability of a product. 

They also explore the default Agent workflows we should be trying as we figure out the right knowledge worker AI copilot. Using the analogy of the history of web development and AI no code, from webflow to squarespace, what those design approaches might tell us about how AI workflow tools will develop.

Finally, we discuss the potential value of talking to VCs when you're not raising, and the perils of taking even good generalized advice all the time.

Topics
* Shortwave email
* Avi Goldfarb and his book Prediction Machines
* Avi also spoke with Patrick O'Shaughnessy
* Yahoo Pipes wikipedia page, or see Retool's amazing history of Pipes

* TLdraw's Makereal

• (00:00) - Intro
• (00:32) - "AI Week" at Startups
• (02:43) - Is it Adaption, Evolution, or Revolution
• (05:23) - Shortwave AI - Adapting email to AI
• (08:07) - Familiar but personalized
• (11:09) - Speed is a feature too
• (15:51) - Taking users on a behavior change journey
• (18:37) - What are Agent workflows we should default to
• (26:04) - TLdraw & Notion, over Pipes, as the core UX for AI
• (29:43) - TLDraw as metaphor for "show me the output" application building
• (31:13) - History of Web Development as analogy for AI no code
• (33:37) - Is it worth talking to VCs when you aren't raising
• (43:24) - Building relationships, Bizdev vs Sales