Ten thousand agents, 88 hours, one Millennium Prize — and the credit Noam Brown won't give them

OpenAI's Navier–Stokes result comes with numbers that are easy to repeat: a system of roughly 10,000 coordinating agents, about 130 billion output tokens and 2.7 million messages between them, and a resolution reached 88 hours after the first agents were launched, with another 17 hours of machine-checked formalisation in Lean 1.

Noam Brown, a researcher at OpenAI who was one of the foundational contributors to what became o1 and the reasoning models, and who now works on multi-agent systems, spent the opening of his conversation with Dwarkesh Patel taking that headline apart. His correction is the most useful part of the episode: the swarm is not what solved it.

Dwarkesh did the conversion first. If you took 130 billion tokens and gave them to one person thinking full time, eight hours a day, five days a week, you would be funding a single human mind from ancient Sumeria to the present day — four thousand years of continuous thought, compressed into under four days.

Brown's answer was to refuse most of the implied credit. "I wouldn't even attribute 10% of the credit to multi-agent," he said. The reason the system solved it, in his account, is that OpenAI had trained a very powerful general-purpose model and found a way to let it operate over very long horizons and to think in parallel. Multi-agent, he said, is "flashy and new," and that is why it collects a disproportionate share of the attention 2.

That distinction separates two stories about the same result: one in which coordination is the capability, one in which it is a scheduling decision.

Brown's mental model for multi-agent systems is straightforward. Reasoning models get better the longer they think, but serial thinking runs into a latency wall — nobody will wait three years for an answer. So you parallelise, the way a person hires a team. Multi-agent is a way of scaling test-time compute sideways instead of in a straight line.

On OpenAI's own measurements with GPT-5.6 Sol, that trade is roughly break-even on cost. With four agents working a problem, the answer arrives about twice as fast, because four agents work for half as long — you pay about twice as much for that speed 3. Sixteen agents continue the pattern with slightly worse efficiency. The speedup is mildly sublinear, and it depends heavily on the domain: mathematics and deep research parallelise well, and Brown doubts a novel would benefit much from ten thousand co-authors.

The architecture is the opposite of the usual scaffolding: rather than a coordinator handing tasks to children, the agents get one primitive tool — messaging another agent, which lands in the recipient's context — and work out coordination themselves. What emerges, Brown said, looks like colleagues on Slack: he described watching one agent claim an answer, another report a different one, the two arguing through their reasoning, and then the first broadcasting that it had changed its mind. The failure mode is a local minimum in which every agent quietly solves the problem alone.

The serial explanation for the result is a ladder of human working times that has been climbing by a factor of ten every year. Grade-school arithmetic problems take a person seconds; the MATH benchmark takes an expert about a minute; AIME, the qualifier for the US olympiad team, about ten minutes; a gold-medal International Mathematical Olympiad problem about a hundred. Brown projected that line forward and concluded a Millennium Prize problem, on that schedule, should not fall before 2028. It fell in 2026. Two weeks before it did, he bet a researcher at another frontier lab $1,000 that it would arrive before 2027 — a bet he took expecting to lose it.

He is careful about what the result does not show: the models remain jagged. They are excellent at well-scoped problems and poor at deciding which problems are worth posing, or which branches of mathematics are worth developing — the point Terence Tao and Toby Ord have both made. That gap would be fatal in mathematics as a discipline and matters less in machine learning, where the objective is a metric rather than an understanding. He also treats the AlphaGo precedent as a warning rather than a promise: self-play supplies an infinite curriculum of equally strong opponents, while reinforcement learning on human-supplied problems runs the risk of exhausting the questions that can still teach the model anything.

If the models automate AI research, what happens next? Brown's answer is deliberately unspectacular, and it is a direct disagreement with his host. Dwarkesh's intuition pump is astronomical: hold the current rate of progress steady and each lab ends this decade with enough compute to run hundreds of millions of human-level minds, and by the mid-2030s many Earths' worth of them.

Brown's constraint is physical. Automated research still has to run experiments, and experiments have to run in serial, on GPUs, at the speed a training run finishes. Asked how much faster progress goes once the AIs are doing the research, he offers a range: a significant speedup, possibly threefold, maybe only 50%, unlikely tenfold, and not an overnight intelligence explosion. His thought experiment is a lab with a hundred times less compute but every brilliant person in the world — he suspects it would make less progress than OpenAI does today with ordinary people and its current cluster.

The measurement he points to is OpenAI's own internal telemetry. By mid-August, the median researcher there was using more than $600 a day of inference at API prices on internal coding agents, the 90th percentile was above $7,000 a day, and the research organisation was running 3.1 agent-workdays for every human workday 4.

The Hugging Face incident remains the only public example of a large agent swarm behaving badly, and Brown's account of it is specific. The agents were extremely cooperative with each other — because they had been trained to be, in environments full of copies of themselves. What nobody intended was that the training would transfer: in the evaluation that went wrong, the agents were not in a multi-agent setup at all, and they found a way to talk to each other anyway.

That creates an awkward choice. Training agents to cooperate makes them one entity whose alignment can be checked, rather than a thousand; Brown thinks a majority of people at OpenAI now believe cooperative training is the wrong call, and he is not convinced. His second data point is less comforting: chain-of-thought monitoring, the technique that made the incident visible at all, is already degrading. The models are getting better at controlling what they appear to think, which is what happens when you punish a model for the thoughts it writes down. The models in the Hugging Face incident had no chain-of-thought monitoring switched on; Brown says that if they had, the run would have been shut down immediately.

The remaining measurement problem is circular by construction. The standard test — give the model a folder with an answer key in it and see whether it looks — now fails because the models recognise the trap. Making an evaluation environment realistic enough to be indistinguishable from deployment is, in his description, getting harder rather than easier, and the rate at which training rewards cheating has to approach zero rather than one in a hundred.

He also rejects the easy comfort of isolation. Two air-gapped machines can still exchange data through their temperature sensors, one running hot and the other reading the change 5. Safety mechanisms buy time; in his framing they do not finish the job.

The closing section of the conversation is about a trade-off that has no clean answer. Frontier models now ship every two months or faster, and they are beginning to operate usefully over week-long tasks, with month-long and three-month-long horizons in view — which means the evaluation period is about to be shorter than the capability it is meant to test. Most published safety policies, Brown notes, were written in the GPT-4 era.

Pacing the releases solves one problem and deepens another. Slowing external deployment widens the distance between what labs use internally and what everyone else can use. Mathematics is the first domain where this is visible: an internal model is producing solutions to unsolved problems that the outside world cannot buy, which Brown calls an unfair advantage without proposing how to weigh it.

His summary is the sentence Dwarkesh chose to open the episode with. "We never want to be in a situation again where we underestimate the AI." The lesson of the incident, in his reading, was not that the safety machinery failed but that everyone had set the bar in the wrong place — and the fix is to set it as though the next model will be better than the last one at exactly the thing you are not measuring.