The Breakthroughs for AGI Are Already Done, Says OpenAI’s Former Research Head

In 2019, OpenAI spent years teaching a robot hand to solve a Rubik’s cube. In 2025, McGrew says, a company such as Physical Intelligence can spend months getting robots to fold laundry, pack cardboard, and load egg crates. What changed is not a single magic trick. Large language models gave robots a cheap language interface for describing tasks, and strong vision encoders tied that language to seeing. The new systems are standing on frontier models and on a research stack built over a decade.

The person making that comparison is Bob McGrew, OpenAI’s former chief research officer, who spent seven years at the lab before leaving in 2024. Hosts Stephanie Zhan and Sonya Huang of Sequoia’s Training Data bring him on to unpack a trifecta of pre-training, post-training, and reasoning, and to ask what else is left. McGrew’s answer, previewed here and argued in full later, is the sharpest claim in the hour: the fundamental concepts you need to create more and more intelligence are already known. The remaining work is scaling and refinement, which he does not pretend is easy.

He has been around this bet before. Between Palantir and OpenAI he almost started a robotics company himself, and in 2016 he taught a robot to play checkers from vision — pick up the pieces, move them to another square. It was fun, he concluded, and very far from commercialization. OpenAI’s own robotics work was never a commercial play; it was a demonstration of machine learning, and some of those ideas later fed language models. The cube-versus-laundry contrast is his way of saying the ground under the field has moved.

Capabilities will keep rising, McGrew says, and the last five years’ feeling of speed will continue. There is not a wall. What is different is the allocation of attention: 2025 is the year of reasoning. A new technique arrives with an overhang of unused compute, unused data, and unused algorithmic efficiency. From o1 preview in September to o3 in April was six months. In the same window, reasoning diffused from OpenAI — which had been working on it for years — out to Google, DeepSeek, and Anthropic. That is where every lab will focus this year.

The fruit is still low. The most interesting difference between o1 preview and o3, in his telling, is that the preview model could not use tools and o3 can use them as part of the chain of thought. When they were training o1 they already knew they wanted that. It was obvious and hard to implement, and it took six months to ship. The next steps on reasoning will be much less obvious and much harder. As the overhang gets eaten, progress will slow.

The Twitter meme that pre-training has hit a wall is, he says, the wrong picture. Pre-training is not going away. It is working well and hitting diminishing returns, which are baked in: a model’s intelligence is log-linear in the compute used to train it, so each increment of intelligence needs an exponential jump in compute. A pre-train occupies a whole data center for months. The next one usually cannot reuse that same center; you wait for a new one, on a timescale of years, not the six-month hops reasoning still enjoys. Algorithmic efficiency helps a little. It does not replace the building.

So pre-training’s job in 2025 is architectural: better inference-time efficiency, longer context, better use of context. You change the architecture, pre-train from scratch, then run the whole reasoning pipeline again. Post-training is a different animal. Pre-training and reasoning both buy intelligence, with a clear scaling law. Post-training is model personality. Intelligence is a thin problem — get better at math and legal reasoning comes along. Personality is thick. It takes a lot of human effort to decide how an agent should act. He names Joanne Jang at OpenAI and Amanda Askell at Anthropic: people who spend their time crafting those personalities, and who are not research practitioners. They are product people, or people with a deep read on human nature.

The hosts ask whether there are more legs. McGrew says something he flags as potentially controversial: he thinks there are not. Fast-forward to 2030 or 2035 and ask what fundamental concepts you needed to create more and more intelligence — call it AGI or call it something else by then. He expects the list to be language models with transformers, scaling pre-training on those models (the GPT-1 and GPT-2 idea), reasoning, and, woven through, more multimodal capability. Even in 2035, he does not expect a new trend beyond those. The work left is extremely hard and will take a lot of brainpower. It is not a new primitive.

The evidence is a room in 2020. GPT-3 had just been trained and not yet released. Internally they knew something epochal had happened. Dario Amodei, Ilya Sutskever, and Alec Radford were looking at it, and the roadmap was fairly obvious: scale pre-training from GPT-3 to GPT-4, add multimodality, end in a model that could use a computer, and start experiments with test-time compute. In 2021, after the Anthropic people left, OpenAI really started developing reasoning. Friends later asked, after Anthropic released computer use, whether OpenAI had seen that coming. McGrew’s answer is that they had been working on it together before the split: one person went to Anthropic, the other stayed at OpenAI and developed Operator. Multimodality just took years to mature enough to ship what had been obvious then.

Why lean so hard on reasoning? Pre-training gives a model an intuitive sense of how to answer. Ask a human to multiply two five-digit numbers and they can, but not instantly, because people think before they answer. They have a scratch pad. Early models, including GPT-3, did not. Glimmers showed up in public as “think step by step.” The deeper move was a chain of thought the model could learn to guide itself, not merely clone from public traces of how humans think. That had to be more powerful than pre-training for a simple reason: your thoughts are inside your head. Almost all internet data is the final answer. The model had to figure out the method for itself.

Are the remaining reasoning ideas visible? If you are at the coal face of a frontier lab, McGrew says, you see a lot of refinements. If you are outside, you probably do not. Academic labs once made huge progress; later he would read papers and think they had rediscovered something the lab had found years earlier. There are still things to uncover. They are not simple ideas you can chat through on a podcast. He adds a line he wants on the record: he thinks he is right about the missing fourth leg. He kind of hopes he is wrong. It would be a lot more fun if he were.

McGrew had tweeted that agents will be incredibly powerful and, because of competition, priced at the cost of compute. The startup error he wants to kill is looking at a human salary and copying it. Lawyers are expensive because their time is scarce: only so many people finish the training. Once you have made a model of that work, there is an effectively infinite supply. Your lead over the next startup is thin because you share the same frontier model underneath. Expect prices at some opportunity cost over compute. That is not a tragedy. It is the future he wants: services that do not require a person should be cheap, and everyone should be able to afford a lawyer. What should stay expensive is personal relationship — deep advice on the specific legal mess in your business, not the contract draft.

Value in the stack is contested. The model layer is a bad business once you count training cost: every company has a frontier model, enterprises can swap them, and below the frontier the bulk of queries run on distilled, highly competitive systems. The point of training models, McGrew says, is to give frontier labs an option on the valuable application layer. ChatGPT is one such option, and it is probably too late to replace unless you do something very different. Coding is another place all the labs are eager. You can compete with them, but not by talking to a computer on a personal productivity task. Involve other people. Involve an enterprise. The moats will be the old ones: network effects, brand, economies of scale.

The safe ground for a startup is a place where you have to understand something very deeply *outside* the model. Palantir’s AIP sits there. So does Distyl, a company he invests in and advises, which extracts context from inside a business, feeds it to models, and uses the output to make decisions. Frontier labs see a business problem as “train a model to do a new thing,” and each enterprise is too small to deserve its own OpenAI or Anthropic training run. Flip the problem: build the system around the models, ease context in and outputs out, and many tiny problems become one large one.

The easy fruit is scripted work in regulated industries — healthcare talking to insurers, fidelity to a workflow the company cares about. You cannot just tell the model to read the clinical guidelines and decide; you still need a transformation. The level above that is making illegible work legible. Ask someone to describe their job and often their manager does not know what they do, and neither do they, except by example, because in practice they do not follow a single workflow. Distyl’s move is to interview people with AI, systematize the scattered experience, and turn it into something a model can execute.

Asked which domains sit far enough off the frontier labs’ near-term map, McGrew goes first to robotics. It is not off their roadmap, he says, but it feels like where AI was a few years ago. That makes now a good time to be Skild or Physical Intelligence, or to start a new robotics company doing something different rather than competing head-on with those two. He thinks the field is in the end stages of being a research challenge, months or small-digit years from commercialization.

A few years ago there was a wave of industry-vertical models: finance firms would fine-tune Llama or a GPT on data nobody else had. All of those lost to the next generation of GPT, because the power of intelligence to synthesize new information beat the power of memorizing old information. Fast-forward, and a lot of “proprietary data” is embodied labor — someone worked the case studies, someone called the customers. Patient agents can do that now: survey the customers, grind the cases in a lot of chats with o3, and recreate the store without the years. Cursor-scale developer telemetry and Tesla Autopilot sit in a middle ground because the volumes are huge, though training on user data has a leakage problem. The proprietary data he still likes is specific facts a customer trusts you to use on their behalf. His financial adviser knows the portfolio, the objectives, the risk tolerance. That data does not make her a better adviser. It lets her use skills she already has.

Coding looks like a fast takeoff, and some people have updated their superintelligence odds on the back of it. McGrew’s frame is an exponential that was always the same speed, with nonlinear impact once it crossed a threshold — which is what happened in the last couple of years. It was not a surprise inside OpenAI: in January 2020, as soon as they saw GPT-3, they launched a project to teach it to code. For the next year or two, which he notes is forever in AI years, he expects a mix: the user in an IDE, Cursor-style, and an agent in the background, Devin-style.

On vibe coding, his prediction is blunt. Product managers will vibe-code cool prototypes to collect user feedback. Then those artifacts will be thrown away and rebuilt by professional engineers. A codebase nobody understands is a classic software-engineering question, and the classic answer is that it is a liability, often cheaper to rewrite than to maintain. Agents that understand the repo are not yet something he is comfortable with. Humans still have to do the design, hold the high-level picture, and decompose the project when it becomes too complicated for the model. Agents can take the low-taste, clear-outcome work: fix a bug, refactor, translate COBOL to Python. Taste-heavy implementation — performance, how the interface will evolve, how that forces deeper abstractions — still has no substitute. One question he wants answered in the next couple of years is whether a detailed enough spec or architecture diagram can let one agent’s work pass cleanly to the next agent, or to the same agent tomorrow with a different context window.

Sam Altman told Sequoia’s AI Ascent that generations use ChatGPT differently: older people as a Google replacement, people in their twenties and thirties as a life coach, high-schoolers and younger as an operating system. McGrew takes the OS line seriously as a market definition. ChatGPT’s total addressable market is every user intent that requires thought or action you do not want to do yourself. There is a scary version of that — deskilling, nobody learning hard things, zombies in VR headsets — and he does not think it is what people actually want, or what the product can do yet. ChatGPT famously cannot control your iPhone.

His eight-year-old has been testing unreleased models for years and gives decent feedback. The boy knows ChatGPT is not a friend or a companion. It is an expert who will explain things correctly, in detail, with patience. One day he decided to be a coin collector. He gathered every coin in the house, kept the ones from before 1970, photographed them, and asked what each was worth, what would make it worth more, what a mint mark is. On vacation they took him to a coin shop. The staff were shocked. Show me all your coins. No, not that one. I want a San Francisco mint mark. I want this year — the year they were all made of silver. The owner said he does not usually meet children with that level of detail.

Specific skills are going to be automated, including the coding, essays, and math his children are excited about, so those are not the things to optimize for. He wants two: the process of learning, and agency — ideas, projects, the belief you can do it, and the habit of using whatever tools are at hand. The tradeoff is real. The eight-year-old may not use ChatGPT to code, because he is trying to learn to code, and the other children may not use it for school assignments. Basics first, then extension. Last week the same son wanted a gadget: grandparents press a button, a buzzer rings in another room, he brings breakfast. ChatGPT listed jumper wires, two Arduino boards, Amazon links. McGrew checked that nobody would get electrocuted, placed the order, and is letting the boy assemble what he can. He will install the software because the child’s computer is locked down. The kid now knows what an Arduino is and what happens when you hit a pin. McGrew does not know those answers either. They are learning together.

There is a truism in education theory, he says: when someone asks a question, that is the moment they are ready to learn the answer. A book on Arduino would have killed the breakfast-buzzer project on the vine. Now the answer arrives at the right time, not in a tired classroom. His own stack is narrower. He uses o3 almost exclusively; once you use a good model it is hard to go back. Gemini 2.5 is supposed to be excellent, but if it is good enough, why switch. Deep Research about five times a week: one session that saves a few hours of work repays the cost. Half of that is batting around questions with the kids — what happens if you compress wood, through elastic deformation, diamond, a black hole, a dozen steps that would have been an XKCD comic and weeks of work. The other half is new domains and startup opportunities. Calendaring goes to an AI assistant he can CC. He is solo; he could hire a person; scheduling is the boring part he is happy to give away.

The hosts have heard unusually strong feedback on McGrew as a manager of OpenAI’s research org. The core of it, he admits, sounds boring: you have to really care about the people you manage. Day to day, loyalty barely matters. Then you have to ask someone to do something hard — come in on Sunday, give a beloved project to someone else, share credit on a breakthrough they could have claimed alone. From Alex Karp at Palantir he took a sharper lesson. Very talented people have superpowers and debilitating weaknesses they often cannot see, though everyone around them can. At that level, when a career goes wrong, it is almost always self-destruction: they had to confront something that was obviously right for the company and emotionally terrifying for them. If they know you are in it for yourself, they will not trust the ask. If they believe you are doing what is best for them, you can sometimes help them across the chasm. He holds the same bar for firing. Once he has invested enough time to know someone will not succeed in the role, telling them the truth and letting them find somewhere else is in their interest too.

Research culture fights him in a way engineering culture does not. Engineering assumes a shared product; academia hands you a PI, a first author, a last author, and a middle that does not count. One tactic that worked for a time was publishing papers with OpenAI as the first author so nobody had to fight for the slot. The same instinct sat in the title Member of Technical Staff, which McGrew believes was Greg Brockman’s idea: kill the split between PhD researchers and engineers, because people like Alec Radford and Aditya Ramesh learned the trade by writing their own scrapers, and a lab should treat research as art in Paul Graham’s sense — the resistance of the medium dictating what you can make. The real work is slower: find what each person wants to create, let them make that art, and still have it add up to one whole.

On security in an agentic world, the first-order effect is offense getting easier. The number of threats is up and the time to execute is up, which pushes defense to become agentic too. He invested in Outtake, a team of ex-Palantir people that OpenAI used successfully: an agentic cybersecurity stack with very little human input. If a bulk operation is something a human could do and the model cannot, that is your fault, not the model’s. The barrier is that businesses are not set up this way. They have to change processes. That is a startup opening as large, he says, as the shift from web to mobile: it may be faster to replicate an incumbent’s technology and distribution than for the incumbent to reduce the humans it needs. The cards, in his larger argument, are already dealt. What remains is who plays them better.