Jan 31, 2025 · 30m · y-combinator
Bob McGrew: AI Agents And The Path To AGI · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Y Combinator interview hosted by Garry Tan, former OpenAI Chief Research Officer Bob McGrew discusses the evolution of AGI, the shift from pre-training data bottlenecks to test-time reasoning compute, organizational strategies for frontier AI research, and the impending breakthroughs in physical robotics and autonomous agents.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 5.6 Guest teaching 5.2 Guest disagreement 1.4 The partners pushing back 1.6
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:33 McGrew rejects the Her AI girlfriend narrative
McGrew directly dismisses Tan's framing that romantic AI companions are inevitable, arguing that emotional AI girlfriends do not match what people actually want.
Hardest push from the partners ▶ 9:31 Tan challenges why scaling laws were not leveraged earlierTan directly questions McGrew on why AI researchers did not apply scaling law principles to other modalities like robotics sooner.
Biggest teaching moment ▶ 9:35 McGrew explains the 0-to-1 hurdle before scaling lawsMcGrew educates Tan by distinguishing between the difficult zero-to-one algorithmic phase and subsequent scaling, citing the multi-year effort to produce DALL-E's first crude image.
The partners hold their own ▶ 22:39 Tan delivers deep analysis on forward deployed engineeringTan leverages his Palantir background to deliver a thorough operational breakdown of why AI adoption stalls without on-site forward deployed engineers crafting custom software.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Bob McGrew's Path from Palantir to OpenAI | 3 | 5 | 1 | 1 | Tan acts as a welcoming conversationalist, prompting McGrew to recount OpenAI's founding and early pivots. McGrew delivers an expansive retrospective on robotics, Dota II, and Alec Radford's early GPT-1 work. | |
| OpenAI's Research Culture vs. Google Brain and DeepMind | 4 | 6 | 2 | 1 | Tan probes the cultural differences between AI labs and academic paper authorship incentives. McGrew explains OpenAI's startup approach compared to DeepMind and Google Brain, detailing how OpenAI avoided authorship battles. | |
| Scaling Laws Across Domains & The 'Zero to One' Stage | 5 | 6 | 2 | 3 | Tan presses on why scaling laws were not exploited sooner in other domains like vision and robotics. McGrew reframes the question by explaining that a difficult zero-to-one phase must precede any scaling law. | |
| Overcoming the Pre-Training Data Wall with Reasoning | 6 | 6 | 1 | 2 | Tan asks about the ongoing debate over the pre-training data wall and synthesizes Moore's law S-curves. McGrew details how test-time compute and reasoning models unlock a new scaling paradigm beyond pre-training. | |
| Autonomous AI Agents and Increasing System Reliability | 6 | 6 | 1 | 1 | Tan prompts on OpenAI's five levels of AGI and bio lab applications. McGrew explains that autonomous agents need high reliability, which requires order-of-magnitude increases in compute through extended thinking time. | |
| Model Distillation and Strategic Advice for AI Founders | 5 | 6 | 3 | 2 | When Tan suggests the AI companion future of the movie Her is inevitable, McGrew pushes back sceptically on emotional AI relationships. McGrew then outlines the mystery of why widespread laptop automation has lagged behind 2018 predictions. | |
| Forward Deployed Engineers and Reimagining Workflows | 8 | 3 | 1 | 2 | Tan demonstrates deep operational expertise drawing from their shared Palantir background, articulating why Forward Deployed Engineers are essential to link AI models to bespoke customer workflows. McGrew enthusiastically agrees. | |
| Education, Parenting, and Future Human Roles: Genius & Manager | 6 | 4 | 1 | 1 | Tan asks how AI reshapes parenting and education, following up with a historical analogy about photography expanding appreciation for painting. McGrew outlines his framework of future human roles: the lone genius and the manager. | |
| The Impending 'ChatGPT Moment' for Physical Robotics | 7 | 5 | 1 | 1 | Tan connects robotics with hard tech investments in fusion and energy as a triumvirate of abundance. McGrew forecasts a ChatGPT moment for physical robotics within five years as foundation models mature. |
Statements from this episode (24)
McGrew: Solving Dota 2 proved scale enables AI generalization
“And there was real insight that was generated there, which was that it really strengthened our belief that scale was the path to improving artificial intelligence. That with Dota two, the secret idea was that we could take huge amounts of experience and feed i…”
McGrew: Nobody believed next-token prediction would work before GPT-1
“And the core idea behind GPT-ONE is that if you have a transformer and you apply this super simple objective of guessing the next token, guessing the next word, that that would be enough signal that you could actually have something that would be able to gener…”
McGrew: Ilya Sutskever was the guiding light shaping early OpenAI research
“Early on there were sort of, you know, a couple big projects, as I said, and then some room for exploratory research. And at the very earliest days, the exploratory research was really about you know, it's about what the researcher wanted to do, but also it wa…”
McGrew: DeepMind was top-down while Google Brain replicated unguided academia
“And so early on the deep mind culture was I, you know, a caricature is Demis had a big plan and he wanted to hire a bunch of researchers so he could tell them to move forward with his plan. And Google brain said, let's rebuild academia. Let's like bring in all…”
McGrew: OpenAI succeeded by balancing startup agility with opinionated scaling bets
“And we took a different approach, which was really more like a startup where there was no sort of big centralized plan. But at the same time, people didn't have, you know, it wasn't just sort of let a thousand flowers bloom. Instead we had opinions about what …”
McGrew: Academia's obsession with individual credit inhibits collaboration
“I think academia is good for this very narrow thing of, you know, small groups, you know, trying out crazy ideas, but academia has a lot of incentives that prevent people from collaborating. And in particular in academia, there's this obsession with credit.”
McGrew: OpenAI authored early robotics paper as 'OpenAI' to avoid credit disputes
“And one of the early robotics papers, we actually said site as open AI because we didn't want to get into a fight. You know, the first author is the one who, you know, gets cited and their name shows up every single time. So we said, you know, we're not going …”
McGrew: Aditya Ramesh spent up to two years building DALL-E 1
“I think Aditya Ramesh, who built that model, spent 18 months, maybe two years, just getting to the first version that, that clearly worked.”
McGrew: Initial AI breakthroughs are completely separate from scaling laws
“Just getting to that point where it's sort of plausibly begins to work is, is a huge, difficult problem. And it's completely separate from using scaling laws. Now, once you get it to work, that's when scaling laws come into play.”
McGrew: LLM pre-training scaling will inevitably hit a data wall
“It is definitely the case that there is a data wall and that if you take the same techniques that we were using to scale LLMs you know, at some point you're going to run into that.”
McGrew: LLMs have entered a pure scaling regime toward AGI
“And now that that has been cracked at this point, I think we actually have a very clear path to just focus on scaling. You know, we were, you know, talking about that you know, the zero to one part, that's not about scaling. I think there's a really strong cas…”
McGrew: Extended reasoning compute will unlock reliable autonomous AI agents
“What we're going to see out of reasoning, out of long thinking is that it's really going to unlock the possibility of agents to do actions on your behalf, which, you know, has sort of always been possible, but it's just never been quite good enough.”
McGrew: Adding a 'nine' of AI reliability requires 10x more compute
“There's a rule of thumb that I like, which is basically, you know, if you want to go, if you want to add a nine, if you want to go from 90 to 99% or 99 to 99.9%, That's maybe an order of magnitude increase in compute.”
McGrew: Frontier labs will increasingly rely on model distillation for smaller models
“I think over the last year, the big frontier labs and a lot of other people have figured out the tricks to take big models. And, you know, take a very particular distribution of user input and train a model that is almost as good as the big model, but much, mu…”
McGrew: AI founders should build with frontier models first, distill later
“Yeah, I would say if you're a founder, the right approach is to start with the very best model you can, because, you know, your startup is only going to be successful if it exploits some, something about AI that realistically is going to be on you know, the fr…”
McGrew: Men are not seeking deep emotional connections with AI girlfriends
“I am a little skeptical of the deep emotional connection, you know, that, that, you know, guys are gonna have AI girlfriends. I think that's not what guys are looking for in a girlfriend, frankly.”
McGrew: Highly personalized workplace AI assistants are a market hole
“I think this is actually a real hole in the market because that's not something I can go out and purchase today.”
McGrew: AI impact is not visible in productivity statistics
“I mean, yes, AI has had some effects, you know, particularly on people who write code, but, you know, I don't think you can see it in the productivity statistics, unless it's about how big the data centers are that we're building.”
McGrew: AI adoption requires redesigning workflows rather than automating existing ones
“I think that's the, like that we need some twist like that for AI that lets people figure out How to use the AI to solve the problem they actually have not just sort of take their existing workflow and have AI do that workflow.”
Garry Tan: Broad AI adoption requires more forward deployed engineers
“We just need more software engineers who are like that forward deployed engineer to link up the intelligence and we're there.”
McGrew: Learning to code builds critical thinking even if AI writes code
“And I think the answer is that like right now we still have to, like, this is how you learn how to do critical thinking. And, you know, I think back to Paul Graham's idea of the resistance of the medium, like even once You know, the computer can do the program…”
McGrew: Future human jobs will bifurcate into 'genius' and 'manager'
“I think that the role that we're going to be playing, You know, one, I think there's going to be two roles. One will be something like a lone genius. You know, the Alec Radford of the world working alone at his computer, coming up with some crazy idea, but now…”