Noam Brown: There's No Hard Wall—Just Eight Orders of Magnitude of Headroom

The episode opens on a 2021 coffee. OpenAI research scientist Noam Brown is telling Ilya Sutskever that superintelligence will not arrive until someone figures out how to scale inference compute in a general way, and that this looks like a decade of research. It took two or three years. Host Jacob Effron then introduces Brown as a key contributor to o1, with an earlier track record at FAIR on poker and the negotiation game Diplomacy, and tees up the questions that have dominated the past month: whether language models have hit a wall, how far test-time compute can scale, and how Brown now thinks about AGI.

Brown is in a good mood about timing. OpenAI is going to release o1 the next day—by the time the podcast airs, the model will already be public—and he thinks the community is going to love it. Effron does not waste the opening. The question of the past month, he says, is whether model capabilities have hit a wall, and the first slice of that question is how much juice is left in scaling pre-training.

Brown has been public about this: there is more room to push across the board, including pre-training. The right way to think about it, he says, is cost. GPT-2 cost somewhere between $5,000 and $50,000 depending on how you measure it. By the GPT-4 generation the models are much better, but the most important thing that changed is the amount of resources poured in—from thousands and tens of thousands of dollars for frontier runs, to hundreds of thousands, millions, tens of millions, and for some labs possibly hundreds of millions.

Throw more money, more data, and more compute at the problem and you will get a better model. He thinks that will keep being true. The catch is that each further 10x in scale multiplies the bill. The next 10x is billions of dollars; the one after that is tens of billions. At some point it is no longer economically worth pushing. You are not going to spend trillions of dollars on a model. There is no hard wall. There is a soft wall, where the economics stop working.

That is why he is excited about test-time compute. It feels like being back in the GPT-2 days, when scaling laws had just been figured out and it was obvious that a 1,000x scale-up would produce a better model. Doing that in pre-training is much harder now. Test-time compute is still early: lots of runway, and a lot of low-hanging fruit for algorithmic improvements, not just bigger bills. Pre-training is not finished. Moore's law should keep bringing unit costs down. The overhang that made it easy to scale very quickly is simply getting smaller.

He prices the ceiling the same way he prices everything else: in dollars. A ChatGPT query today costs, ballpark, a penny. What would you spend on a query you care a lot about? Not a dollar or five dollars. For the most important problems society cares about, he thinks people would pay a million dollars. That is eight orders of magnitude. And it is not just dumping more money into each query. The paradigm itself can get better, so the same compute buys more intelligence.

A month or so earlier, Sam Altman had tweeted that OpenAI basically knows what it has to do to build AGI. Brown had said that matched the median view of OpenAI researchers. Critics heard hype. He is surprised by that reading. The company has been open that it sees progress continuing rapidly; Altman said it, Brown says it, and he thinks it is a common opinion inside OpenAI. That is not a claim that they have cracked the code to superintelligence and now only have to scale. There are still unsolved research questions. He just does not think any of them will be harder than the ones already solved.

Late in 2021, over coffee, Sutskever asked Brown about his AGI timelines. Brown was skeptical they would get there within ten years. He had watched how much difference scaling test-time compute made in games, and language models had no general version of it. The models of that era could not even draw a tic-tac-toe board. GPT-4 can draw the board and make mostly legal moves, but it still makes illegal ones and suboptimal decisions. Spend tens of billions training a model that can barely play tic-tac-toe, he told Sutskever, and you are pretty far from superintelligence. To his surprise, Sutskever agreed that scaling pre-training alone would not get there—and, Brown later realized, was already thinking carefully about test-time compute. The problem Brown had given a decade took two or three years.

On the job market, the frontier labs were largely receptive. Most of them already believed the current pre-training paradigm would not reach superintelligence. What surprised him was how on board OpenAI was, given that it had pioneered large-scale pre-training and pushed it farther than anyone else. Their motivation, he later learned, was slightly different: getting over the data wall, not scaling test-time compute as such. The techniques lined up. When he joined, the work was still exploratory, with a small group trying things, many of which failed. A few months in, one attempt showed signs of life. People piled on. Leadership recognized something different and valuable and chose to scale it up. That o1 came out of OpenAI is, in his view, underappreciated: it disrupts the paradigm OpenAI itself pioneered, and it is a sign the lab is not trapped in the innovator's dilemma.

The first thing that made him believe the timeline had collapsed was not a benchmark jump. The team had been enumerating behaviors they wanted: try a new strategy when one is failing, break a multi-step problem into pieces, notice a mistake and correct it. Tackling those one by one bothered him. Ideally the model would figure them out on its own. He advocated for a simple move: have it think for longer. When they did, those behaviors started showing up emergently. Performance was better, but not dramatically. The qualitative change was the point—the model was inventing the tactics they had planned to add, and it was clear the approach could scale much further. That was around October 2023.

The deeper shift in mindset came from Diplomacy. After the early-2017 poker results, he had spent years asking how to extend a test-time algorithm from one game to the next—poker and Go, poker and Diplomacy, then as many games as possible. Diplomacy convinced him that was the wrong direction. Porting the poker and Go techniques, even with constraints, only reached strong human-level play, with a visible ceiling. Superhuman performance on the full, general game needed something that would work for pretty much anything. The move was to start from the end point: the most general domain, language, accept that the first way of scaling test-time compute would be inefficient, and then make it scale better.

Eventually, Brown thinks, there will be a single model: if a query needs deep thinking it thinks; if it does not, it answers immediately and well. Right now the split is real. GPT-4o is better for many things; o1 is more intelligent, and it is the one you want for a very hard problem. A professor friend of his is a power user because o1 can handle research questions that would normally need a PhD. For creative writing, GPT-4o beat o1 preview; he is not sure how the comparison looks for the full o1. The practical benefit of 4o is speed. o1 already takes images as input, and he does not see blockers to making it as multimodal as 4o.

Specialized models still have a job, but as tools, not as rivals. You can ask o1 to multiply two large numbers and it will grind through the arithmetic. That is a waste. It should call a calculator or write a Python script. The calculator sits at one extreme—specialized, simple, fast, cheap. o1 sits at the other—general, capable, expensive. Between those poles he expects a lot of things that act as tools, saving the model and the user a great deal of cost. Some of those tools may simply do a better job than o1. He thinks about it the way he thinks about a person: sometimes you should just pick up a calculator.

The use case he is most eager to watch in the wild is coding. o1 preview impressed people in some ways and not others; it was not strictly dominant. He thinks the full o1 will do a lot better. His own split is simple: easy work goes to 4o; a hard problem, or a lot of code, goes to o1, which usually just does the whole thing. He is also clear about a limit: o1 is not doing core AI research yet. The next milestone he cares about is making these models more agentic. After ChatGPT, people kept asking why he was not working on agents. His answer was that the models were too brittle for long-horizon tasks. Prompting could sort of get you there, but it was fragile and not general. o1 is a proof of concept: give it a hard problem and it invents the intermediate steps and executes them, without the excessive prompting 4o would need.

What about all the orchestration people are building around today's limits—chained calls, extra checkers, fine-tuned verifiers? Brown reaches for Richard Sutton's essay *The Bitter Lesson*. He still polls rooms at AI events, and surprisingly few people have read it. The history of chess is the argument: encoding human knowledge lost to methods that scale with compute and data. The same pattern is playing out with language models. Scaffolding and prompting tricks can push the frontier a little further. They do not scale with more data and more compute. Something like o1 does. For startups the warning is pointed. You can solve a here-and-now problem with a custom workflow. You do not want to spend six months on scaffolding that a new model then does out of the box. One reason OpenAI keeps saying the models will progress quickly is so builders do not get stuck there.

PhD students, he says, are in a tough spot. Pushing the frontier depends on data and compute they do not have, so the temptation is to take a frontier model, add clever prompting, and eke out 0.1% on an eval. He does not blame the students. Academia rewards a paper at a prestigious conference, and a small win on a leaderboard is an easy way to get one. In the long run that is not the most impactful work. His advice is not to compete with industry labs on frontier capabilities. Investigate novel architectures or approaches that scale, and show the trend line as you add data and compute. Casual observers may miss it. Industry research labs will not. If the curve looks promising, they will put in the resources to see whether it holds at large scale.

His own sniff test for a new model is still tic-tac-toe. It is shocking, he says, how hard the game is for some of them. He jokes that there are not enough five-year-olds on the internet posting tic-tac-toe strategy on Reddit. He also just asks the models his day-to-day questions, and the jump from GPT-4o to o1 preview to o1 is obvious. The bigger update in his own thinking is speed. He started graduate school in 2012, watched the deep learning revolution, and in 2015–2017 was on the skeptical side of AGI talk: AlphaGo being superhuman at Go, chess, and shogi did not mean it could play poker, and two-player zero-sum games with unlimited self-play are a special case the real world does not offer. The 2021 conversation with Sutskever started shifting that. Seeing a general way to scale test-time compute actually work changed it. He now thinks things will happen faster than he originally expected.

Asked about Amazon's push on Trainium and Anthropic's use of it, Brown is more interested in the hardware implication of o1 itself. The old mindset was massive pre-training runs and then cheap, scalable inference. He does not think that holds. There will be a major shift toward inference compute, and anyone who can optimize for that will have a big win. On a different front, his Diplomacy work makes him curious about social science. Models trained on vast amounts of human data can imitate people, at far lower cost than hiring subjects. Take the ultimatum game: player A has $1,000 and offers a split to B; if B rejects, both get nothing. Experiments show people reject offers below roughly 30 percent. Would you still reject $3,000 out of $10,000 to spite someone? That experiment is expensive, which is why researchers historically went to very poor communities. Models can also run studies that would be unethical with human subjects. GPT-3.5 will not imitate people well, but the match is measurable, and he suspects more capable models do better. They were also trained on a broader slice of the internet than the usual pool of 19-year-old undergraduates at top schools.

A few neighboring questions get shorter answers. Emergent communication—the old problem of teaching AIs to talk to each other—is effectively solved, because the models arrive with a language that humans already use. Robotics will take longer. He did a master's in the field, and the main takeaway was that hardware is hard: iterating on physical robots is slow and expensive. The application he is most excited about is scientific research. Until now models have been broadly capable without surpassing expert humans in hardly any domain. That is going to stop being true, first in a few narrow areas and then in more, which means using them as partners to advance the frontier of knowledge, not as replacements. He does not know whether the current models are already there. That is one reason he wants o1 in the hands of chemists, biologists, and theoretical mathematicians. o1 preview was particularly strong on math and coding; he expects broad improvement, with those two staying ahead.

Quickfire, and he does not love the first question. Overhyped: prompting techniques and scaffolding, the same methods he thinks the long run will do away with. Underhyped: o1 itself. For people paying close attention to the field it was a big update. For the broader world, he is not sure people have recognized what it means yet. Will model progress in 2025 be faster, slower, or the same as 2024? He thinks progress will accelerate.

How does he define AGI? He does not. He has been trying to use the term as little as possible. There will be a lot of things an AI cannot do that humans can, for a long time, and he thinks that is the ideal outcome—especially physical tasks, where humans will keep an edge. The more important standard is an AI that can accelerate human productivity and make lives easier.

He leaves the last word for the skeptics, because he used to be one. He has been in the field a long time, and he thought it would take much longer even to reach this point. Where we are now, he says, is complete science fiction compared with five years ago, let alone ten. Concerns about hitting a wall are reasonable. In his view the test-time compute paradigm addresses a lot of them. Look at the public blog posts and the results. He thinks the evidence is pretty clear.