o1 Is a New Generation — and the Day AGI Arrives Will Feel Boring

Bob McGrew spent eight years at OpenAI and left as chief research officer a few months after o1 Preview shipped. This conversation with Jacob Effron and Jordan Segall of Redpoint Ventures is one of his first interviews after stepping down. He had previously led engineering and product at Palantir for its first decade, then spent a year studying robotics before joining OpenAI.

Several judgments in the conversation are ones outsiders rarely hear this bluntly. By a variety of metrics, o1 is a 100x compute increase over GPT-4, even though OpenAI named it o1 instead of GPT-5. He cites estimates that AI has added about 0.1% to GDP growth, and says that bump is from building data centers, not from productivity in use. He also predicts that within two years there will be an AI-generated feature you actually want to watch — because a director used the model to film something they could not have filmed otherwise.

Whether models have “hit a wall” is where the outside view and the inside view diverge most, McGrew says. Outsiders first noticed AI with ChatGPT, then watched GPT-4 arrive six months later and felt acceleration. After a year and a half of GPT-4 as the public flagship, people started asking whether a data wall had arrived. Inside the labs, the accounting is different.

Going from GPT-2 to GPT-3, or GPT-3 to GPT-4, means roughly a 100x increase in effective compute. Algorithmic gains of 2x or 3x would already be amazing; the rest requires more chips and new data centers, a slow multi-year build that Meta, xAI, and other frontier labs are visibly doing. Before a full GPT-4-to-GPT-5 lift, he expects a 10x increment first — people forget the GPT-3.5 step in between.

o1 is that kind of jump, except the 100x comes from reinforcement learning and test-time compute rather than a bigger pre-training run. The technique packs more compute into an answer by generating a longer coherent chain of thought. A model that thinks for a few hours rather than a few seconds can, if the compute is actually used, spend 10,000 times more. OpenAI had been thinking about test-time compute since around 2020; o1 is the first version that does not waste most of it.

That path does not need a new data center, and there is still a lot of room for algorithmic improvement — in theory from 30 seconds of thinking to hours or days. The hard part is the same as GPT-3 to GPT-4: there were no foundational new techniques, and scaling is the work. Stacking the next pre-training generation with this RL process is, he says, something the field will have to wait to see announced.

GPT-4-class chatbots already handle most everyday questions — who the fourth Roman emperor was, how to reheat basmati rice — without much pain. o1’s value shows up on structured, long-horizon work: programming, or writing a policy brief that has to hang together. What McGrew is most excited about is long-running action. He finds the word *agent* overused, and prefers the concrete version: book things, shop, solve problems by interacting with the rest of the world. Nobody has nailed that form factor yet.

The immediate gap is reliability. If you walk away for an hour and the model goes off task, you have wasted the hour. If it is buying something, pushing a PR, or sending email and Slack on your behalf, a mistake becomes money or embarrassment. His rule of thumb is that going from 90% to 99% reliability takes about an order of magnitude more compute, and 99% to 99.9% takes another. Each extra nine is a generation of model work — a year or two.

Enterprise makes the problem thicker. Consumer context is thin; in a company the model needs coworkers, project state, the codebase, and what people have already tried, all of it scattered across Slack, docs, and Figma. One path is a library of connectors — the same integration problem Palantir was built to solve, which is why he finds Palantir’s AIP interesting. The other is computer use: one hammer for every nail. Anthropic has shipped it; Google DeepMind has published papers; OpenAI was discussing these agents before the Anthropic founders left, back around 2020.

Computer use burns perhaps 10x or 100x more tokens than a programmatic API, which again demands a long coherent chain of thought — the problem o1 was built for. He does not buy a Salesforce-specific operating model. The asset is the dataset of how to use the product, and it would be in Salesforce’s interest to share that with Anthropic, OpenAI, and Google so the skill lands in every foundation model, the AI-era equivalent of SEO. That ecosystem has not happened, and he finds it a mystery. Demos that almost work become 10x better in a year, surprisingly usable in two, and still not fully trustworthy; tasks that can tolerate mistakes get automated first. The missing layer between a good computer-use model and an enterprise contract, he says, is for startups to invent.

The recording landed on Sora’s public release, which McGrew treats as the delayed last step of a multimodal arc. Transformers moved from language around 2018 into vision, image output, and audio in and out, first as side models like DALL-E and Whisper, then into the main model. Video is the modality that resisted that merger the longest. An image is often one prompt; a video is a story over time, which is why Sora’s storyboard — checkpoints every five or ten seconds — matters as much as the pixels.

Training and running video models is extremely expensive. Instantaneous quality is already good — reflections, smoke — but extended coherent generation, from a few seconds to an hour, is the next-generation fight. He expects the LLM cost curve to copy over: a GPT-3-quality token is now about 100 times cheaper than at launch, and Sora-class video should become nearly free the same way. Sora already ships unlimited slow generations on OpenAI’s $200-a-month Pro plan. Winning an award, he says, is too low a bar; the real question is whether it is a movie you want to watch. He guesses two years, and the reason you watch will be a director’s vision, not the generator. Once something can be demoed, progress is fast; before that, it is invisible.

Robotics is more personal. After Palantir, in 2015, he spent a year writing early deep-learning vision code and guessed the field was five years away — badly wrong. He now thinks the five-year call is finally right: widespread if limited adoption in work settings, and a good time to start a robotics company. Foundation models give vision-to-action almost for free, and the ecosystem is mature enough that one founder told him you can just talk to the robot instead of typing at a workstation. OpenAI’s two-year robotics effort showed that simulator training can transfer to the real world, but simulators are good at rigid bodies and bad at cloth and cardboard, so anything truly general still needs real-world demonstration data.

He is bearish on mass consumer robots at home. Robot arms can kill people, including children, and the home is unconstrained. In five years he expects robots in warehouses and other work environments — Amazon already solved mobility and is working on pick-and-place — rolling out domain by domain. Frontier labs will keep shipping one model that does the best of everything on the data they have; specialization is mostly price-performance, via the now-standard pattern of generating traces on a frontier model and fine-tuning a much smaller one.

The right mental frame on AI progress, McGrew says, is deep pessimism: ask why it is so slow. Some people put AI’s contribution to GDP growth at 0.1%, and that is from data-center capital expenditure, not from using the models. Productivity statistics still do not show AI, the same complaint people made about the internet in the 1990s. The 2018 intuition — once it can talk and write code, everyone is automated — was the engineer who says two weeks, then discovers the feature is two months.

AI automates a *task*. A job is many tasks, and most real jobs have some piece that will not automate. Programming is the same pattern: boilerplate goes first, and the last residue is giving direction — what you are even trying to do. The startups he is most excited about apply AI to something boring. If a procurement org were full of Elon Musk-quality people comparison-shopping every dollar, a company would save a lot of money; nobody staffs it that way because those people would hate the job. AI is infinitely patient. It does not have to be infinitely smart.

The first productivity studies showing 20% to 50% gains showed up among consultants, which he finds unsurprising given that models are good at producing fluent output and consulting is in the business of producing it. The bigger lift is in the bottom half of performers. He reads that as hopeful: those people often knew what they were trying to do and lacked the implementation, and the model fills that gap.

McGrew has watched two kinds of great researcher. Alec Radford, who built the GPT series and CLIP, does his best work alone with a computer at odd hours of the night. Ilya Sutskever and Jakub Pachocki, OpenAI’s first and second chief scientists, work through other people with big ideas and a roadmap. The common trait is grit. Aditya Ramesh spent 18 months to two years trying to prove neural nets were creative by generating something that could not be in the training set: a pink panda skating on ice. A year in, Ilya showed McGrew a blur and insisted the pink was gathering at the top. McGrew could not see it. Aditya kept going until it worked. Every researcher who succeeds on a foundational problem treats it as the hill they will die on.

Managing them, he borrows a line from Palantir’s Alex Karp: engineers are artists, 10x engineers are 10x artists, and researchers are 100x the artist of any engineer. You can make them fix bugs, but the artist in them will be sad. Research management is high-touch because the critical failure is snuffing out the artistry. They endure the pain only because of the vision in their heads.

OpenAI’s other cultural fact is how often it was refounded. McGrew joined a nonprofit whose plan was AGI by writing papers — a path that smelled wrong to startup people like Sam Altman, Greg Brockman, and McGrew. Then came the for-profit conversion, the Microsoft partnership, building their own API products, and ChatGPT taking the company from enterprise into consumer. Any one of those would be a definitive startup pivot; at OpenAI it happened every 18 months to two years. Most of it was necessity: the nonprofit ran out of money; Microsoft was not seeing the value, so they built an API to prove demand. ChatGPT was more deliberate. GPT-4 was already trained; the internal bar was a model they all used every day, and in November ChatGPT did not clearly clear it. Cofounder John Schulman wanted outside feedback. McGrew thought a thousand users would count as success. They made the fateful choice not to put it behind a waitlist.

The first days were intense: disbelief, a scramble for GPUs, research compute temporarily diverted, and fear it would fade like DALL-E 2’s viral spike. McGrew bet it would be bigger than the API, and he was right. A less famous, equally painful choice was shutting exploratory work — robotics, games — to double down on language models and generative modeling. Dota 2 had given them the conviction that scale solves hard problems; the math of that game is genuinely harder than Go. Academia, by contrast, he finds badly incentivized around credit and author order. DeepMind tried to put academia inside a company; Google Brain let scholars explore and stationed PMs to catch products. OpenAI’s bet was a lab that works like a startup: freedom to find your hill, collaboration toward one thing rather than a pile of papers.

McGrew has a hard time conceptualizing AGI, and a critique of the concept: there is no single moment. The problems are fractal, and automation will arrive in pieces. The picture in his head is banal. People take self-driving cars to the office, boss around armies of AI, and still feel like the boss is an idiot. The serious version of that view was a long consensus inside OpenAI: reasoning was the last fundamental challenge on the way to human-level intelligence. Pre-training, multimodal, and reasoning are now in place. What remains is scaling — systems, hardware, optimization, data — with very few new foundational ideas. In that sense it is predestined, and the remaining work is hard.

The societal question he cares about is what becomes scarce when intelligence is ubiquitous and free. His guess is agency: knowing the right questions to ask and the right projects to pursue, then actually going and doing them. AI will not solve that for us, and not every human is good at it. Could you just tell a model to build a good company? Sora is the analogy. A vague prompt still yields a video, maybe a cooler one than you would have made, but not the one you wanted. Specific choices are how you get the thing that satisfies you. No matter how advanced the model, coloring in the blank space remains yours.

His own o1 habit is teaching his eight-year-old to code. When the boy asked what a web scraper was, McGrew used o1 to iterate a program short enough not to introduce too many new ideas, but clear enough to teach networking. Playing with the model is itself a test. Evals change every generation: for o1 they watched GPQA until it saturated. Programming has stayed useful for years because it scales from finishing a line of code to writing a whole website, and it is still several orders of magnitude from doing a real software engineer’s job.

He announced his resignation after o1 Preview shipped on purpose. The research program — pre-training, multimodal, reasoning — was solved enough to hand the remaining scaling problems to a next generation of people who are excited about them. He is in no hurry about what is next; after Palantir he spent two years exploring, including a false start on a robotics company, and that wandering is what landed him at OpenAI. Overhyped today: new architectures that look fun and fall apart at scale. Underhyped: o1 itself, which is very hyped and still not hyped enough. He has not changed how he parents, which he calls a failure, and he still cannot fully explain why he teaches an eight-year-old to code when ChatGPT can code. The durable advice is older: put kids on work just at the bound of their capabilities, learn math and coding, write well, read widely. It is not about the syntax. It is learning to think in a structured way about problems.