Aug 18, 2023 · 43m · american-optimist
Ep 66: Bob McGrew — the Superstar Palantir Alum Leading OpenAI's Transformative Research Projects · Joe Lonsdale
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 episode of American Optimist, host Joe Lonsdale interviews Bob McGrew, VP of Research at OpenAI, covering his career transition from Palantir to leading frontier AI development. McGrew shares key insights on neural network scaling, OpenAI's research culture, AI safety and political neutrality, and how AI will reshape education and the economy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Joe holds 37.2% of the talking time here. How this is scored →
Joe as informed peer 5.8 Guest teaching 4.6 Guest disagreement 1.6 Joe pushing back 2.1
speaking balance: gold is Joe, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:02 Bob rejects the model degradation narrative
Bob directly pushes back against Joe's premise that GPT models are degrading over time, explaining that stochastic variance and user confirmation bias cause this misconception.
Hardest push from Joe ▶ 32:55 Joe presses on asymmetric political outputsJoe challenges OpenAI's safety guardrails by citing concrete examples where the model refused to write praise for Donald Trump while readily writing poems for Joe Biden.
Biggest teaching moment ▶ 10:01 Bob explains transformer attention vs recurrent memoryBob clearly breaks down the mathematical and architectural superiority of transformers over previous memory-constrained recurrent neural networks.
Joe holds their own ▶ 39:20 Joe details startup valuation dynamics and legal tech acquisitionsJoe demonstrates deep venture market command by citing specific portfolio acquisition multiples and hardware financing strategies across the AI landscape.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Joe as informed peer | Guest teaching | Guest disagreement | Joe pushing back | Why |
|---|---|---|---|---|---|---|
| Early Roots: From Small-Town Oklahoma to Stanford and PayPal | 6 | 2 | 1 | 2 | Joe establishes deep personal rapport as a former Stanford peer and co-founder at Palantir, actively contextualizing Bob's early career and technical challenges at Gotham with specific recollections of government data constraints. | |
| Perseverance at Palantir: Key Figures and Holding the Team Together | 6 | 5 | 2 | 2 | Joe and Bob reminisce about early Palantir retention struggles, before Bob educates Joe on the historical trajectory of neural networks from 2005 to AlexNet in 2011, gently pushing back on Joe's overextended brain analogies. | |
| The Transformer Breakthrough and Next-Token Prediction | 5 | 7 | 1 | 1 | Bob explains the core intuition behind the transformer architecture and next-token prediction, explaining how attention mechanisms fundamentally outperform older recurrent memory models. | |
| AI Research Culture, DALL-E, and Prompt Persona Mechanics | 5 | 4 | 1 | 1 | The tone is highly collaborative as Bob details prompt persona mechanics and the solo research culture behind breakthroughs like DALL-E, comparing research management to managing artistic obsession. | |
| OpenAI Five: Beating Dota II and the Breakthrough in AI Robotics | 6 | 5 | 1 | 2 | Joe shares personal chess background when discussing game-playing systems, while Bob narrates OpenAI's technical leap from Dota II reinforcement learning directly into solving robotic manipulation. | |
| AI Capabilities, Remote Work Automation, and Future Scaling Timelines | 6 | 5 | 2 | 3 | Joe probes future model scaling, self-training, and physical robotics limits. Bob clarifies the unpredictability of automating complex workflows and explains why reality-simulating compilers matter. | |
| Code Interpreter, RLHF, and Addressing Model Performance Perception | 6 | 6 | 3 | 4 | Joe presses on widespread perceptions that GPT-4 was degrading in performance and questions neuron-scaling comparisons to animals. Bob firmly dismisses model degradation claims as user stochastic bias. | |
| Civil Rights, Political Neutrality, and Safety in Frontier AI Models | 6 | 5 | 2 | 3 | Joe raises sensitive questions regarding political bias, censorship asymmetries, and civil liberties. Bob grounds OpenAI's alignment strategy in creating an objective, professional workplace standard rather than political dogma. | |
| AI Agent Models, Civilizations, and ChatGPT with Kids | 5 | 4 | 1 | 2 | Bob envisions multi-agent ecosystems interacting to build synthetic civilizations, and both discuss everyday parenting dynamics using ChatGPT as an iterative learning tool for kids. | |
| Redefining Education and Homework in the Age of AI | 5 | 5 | 2 | 2 | Bob shares the internal backstory of ChatGPT's unexpected genesis as John Schulman's low-key side project, while debating whether pedagogical homework should be redesigned around AI critique. | |
| AI Investment Landscape, Infrastructure, and the App Layer | 7 | 4 | 2 | 2 | Joe displays significant venture capital expertise regarding M&A valuations and GPU financing trends, while Bob outlines why he remains skeptical of transient infrastructure bets versus application layers. | |
| Optimism for the Future, Career Advice, and Energy Needs | 6 | 3 | 1 | 1 | Joe highlights nuclear investments like Oklo, while Bob outlines a techno-optimist thesis where AI alleviates classical labor and ideation bottlenecks, culminating in energy constraints around fusion and superconductors. |
Statements from this episode (28)
McGrew: Intelligence software cannot follow standard UX customer-interview playbooks
“And so instead of, you know, all the, you know, all the UX books that you interview your customers, you sort of find out what their problems are and you build them to solve their problems. And we couldn't do that. And so basically what we had to do was we had …”
Lonsdale: Intelligence community spent over $30 billion on databases and data
“They had spent over thirty billion dollars on databases and gathering data.”
McGrew: Palantir took three years to get its first active user
“I think it took three years before we had anyone using it to solve.”
McGrew: Great engineers struggle with uncertainty on long unproven projects
“This is one of those things about the really great engineers, is that a really great engineer wants certainty, right? They don't want to, you know, they've been doing something for three years, they, all they can see are all the problems that are in it”
Lonsdale: Thiel funded Palantir's second round because McGrew joined
“You joining was really, really critical to Peter giving us like our money for the second time. Cause he was like, you guys got Bob to join. How'd you do that?”
McGrew: Vision models have specific neurons that activate for individual people
“Yeah, and at some point, there's a Joe Lonsdale neuron that, you know, is out there, you know, and individual people, you can actually find these neurons in the networks.”
McGrew: OpenAI's 2017 thesis was that neural networks could reach AGI
“So I think that the thesis behind open AI when I joined in 2017 was that neural networks were the final architecture that could take AI all the way to human level intelligence or AGI.”
McGrew: We understand neural networks much better than the human brain
“I, at this point, I think we understand neural networks a lot better than we understand the brain. So whenever someone talks to me about the brain, I always think, well, what does a neural network do? And then probably actually the brain does that.”
McGrew: GPT-3 and GPT-4 must be prompted to act smart or dumb
“Well, that's actually one of the cool things that we've discovered with GPT-III and GPT-IV, Is that the model, so the models are trained on the internet and they're reading everything, right? Then they try to act like what they've seen, right? They're predicti…”
McGrew: Nobody but creator Aditya Ramesh believed DALL-E would work
“There's this researcher, Aditya Ramesh. He built a model called DALI, which allowed you to take a description and build an image out of it. He spent two years working on that. Wow. For most of those two years, no one other than him believed it would work.”
McGrew on Palantir CEO Karp's rule: Great engineers are artists
“And I remember, you know, carp always used to talk about this, that the best engineers were artists, right? And if you want to keep an engineer happy, what you actually have to do is just let him do his art, just get out of his way, let him do his thing.”
McGrew: OpenAI solved simulated robotic Rubik's cube in two weeks
“The Dota II team, Jakub Pachoki, who was doing this, had developed a
A technique that allowed you to put lots of data and compute into solving the problem. And it was really working for Dota. And he came over and started doing this for robotics. And we went f…”
McGrew: There is no limit to neural network capability with enough compute and data
“If you could figure out how to pour enough compute and pour enough data into a neural network, there's really no limit to how, how good it could get.”
McGrew: Robotics will probably require AI to figure out robot manufacturing
“Robotics are probably going to need AI to write that, to figure out how to build robots for us.”
McGrew: OpenAI models are improving, contrary to user reports of degradation
“Ah, we think it's actually getting better over time that, you know, some of the reports are sort of confounding various different things.”
McGrew: AI scaling will keep working without hitting an asymptote
“The way I think about it is that we are still pretty far away from the level of scale that is the human brain. And so there's no reason for there to be an asymptote. I think it just keeps working.”
McGrew: Recent AI progress is just a series of engineering tricks
“Like the difference between neural networks and linear regression was really big. And the difference between neural networks now and neural networks two years ago is just a sequence series of tricks.”
McGrew: AI bias mitigation should prioritize workplace professionalism over political correctness
“And so I think a lot of, like near-term safety, a lot of these discussions about bias, like they really are about the fact that the model has learned all of these things, which it's learned a lot of true facts, it's learned a lot of things that are inaccurate,…”
McGrew: OpenAI uses thousands of human trainers scoring prompts to counter model bias
“Well, there's actually thousands of AI trainers behind the scenes who are, you know, we, we're, we give them prompts and they have to score the prompts. And so, you know, I think for better or worse, they bring their biases to it, but we try to make it so the …”
McGrew: Human-level multi-agent AI will create its own civilization
“When you push this to the level where AI is at the human level, I mean, they're basically going to be creating their own civilization.”
McGrew: Teachers should design homework that requires AI use
“I think I don't want my kids using it for their homework, but what I want is I want teachers figuring out how to make homework that It requires you to use AI.”
McGrew: OpenAI secretly trained GPT-4 in first half of 2022
“We had already then secretly trained GPT-IV. So this was the first half of last year.”
McGrew: ChatGPT launched as a side project with zero press outreach
“And again, it was a side project. We called it a low-key research preview. We didn't do any press on it, and it just blew up.”
Lonsdale: Portfolio legal tech startup sold to Thomson Reuters at 8x valuation
“One of them was a legal tech company. You probably don't even know this. That's partnered with open AI and suddenly sold for like eight times the last round valuation to Thomson Reuters.”
McGrew: GPT-5 will render current AI infrastructure solutions completely obsolete
“I'm always a little worried about the infrastructure work, because I think, you know, you're solving the problems as they exist today, and when GPD 4.5 and GPD five come out, you know, they're gonna have fundamentally different use cases and fundamentally diff…”
McGrew: It is very hard to bet against Nvidia in AI hardware
“I think what everybody has seen so far is that Nvidia has just been able to continually do better and better and better because they have, you know, the big market and a lot of capital to throw at it. So I think it's pretty hard to bet against Nvidia.”
McGrew: AI scaling will ultimately be constrained by power and chips
“You know, the interesting thing about AI is, in the end, it's going to be limited by things like power. You know, if you're, and, you know, chips, and if you're trying to build these really massive clusters in 10 years or whatever.”