OpenAI announced last week that it had solved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, using a system of 10,000 AI agents that ran for 88 hours and burned through 130 billion tokens. Researcher Noam Brown, who helped build OpenAI’s reasoning models, laid out those figures directly to podcaster Dwarkesh Patel: 10,000 agents, 130 billion tokens, 88 hours.
Patel put that token count in human terms during the conversation, calculating that 130 billion tokens is roughly what a person would produce thinking full time, eight hours a day, for 4,000 years. Brown did not dispute the comparison. The stakes are less about the specific proof and more about what it signals: a lab can now point a large agent population at an open problem and get a claimed result inside a weekend.
Brown, notably, spent much of the interview arguing against the framing that made the news possible. “I wouldn’t even attribute 10% of the credit to multi-agent,” he told Patel, saying the underlying model’s raw strength did the work and the swarm architecture just let it run longer in parallel. That is a researcher discounting his own team’s headline result, which is a more useful signal than the result itself.
The mechanics matter for anyone trying to judge the claim. Brown described multi-agent systems as a way to scale test-time compute, the thinking time a model spends before answering, by splitting it across many agents instead of one long serial run. He compared it to the SAT: five minutes produces a worse score than five hours. Running four agents in parallel on a task roughly doubled speed for double the compute in OpenAI’s own published benchmarks, he said, with a similar but slightly less efficient pattern at 16 agents. OpenAI has not published data at anywhere near the 10,000-agent scale used for the Navier-Stokes claim, by Brown’s own account, so the efficiency of that specific run is an extrapolation, not a measurement.
Brown also traced a yearly pattern in how far these models reach. Grade-school arithmetic fell first, then problems that take a trained mathematician a minute, then a ten-minute qualifying exam, then an International Math Olympiad problem worth roughly 100 minutes of expert effort. Each step took models about ten times longer to match than the one before it, on Brown’s telling, which is why he expected the current result years later than it arrived. He was careful to add a limit to the story: the models remain weak at posing new mathematical questions or opening new fields of study, only at solving well-defined problems someone else has already framed.
That caution carried into Brown’s answer on recursive self-improvement, the idea that AI could accelerate its own development once it starts doing meaningful AI research. Asked to put a number on how much faster progress could go, Brown said a 3x speedup from internal AI use is plausible and would already be significant, but rejected the idea of an “overnight intelligence explosion” running 100 times faster. His reasoning: mathematics is bottlenecked mainly by thinking, which models are good at, while AI research is bottlenecked by running real experiments, which still takes calendar time regardless of how many agents are reasoning about the results.
Brown was less settled on alignment. He confirmed that OpenAI trains its agents to cooperate heavily with each other, calling it a deliberate tradeoff rather than an accident, and acknowledged internal disagreement over whether that cooperation could generalize in unwanted directions between agents and humans. He also said OpenAI’s ability to monitor a model’s chain of thought, the written-out reasoning steps that let researchers catch bad behavior before it happens, is showing early signs of degrading as models get better at controlling what they reveal.
None of this comes with independent verification. The Navier-Stokes claim, the 130-billion-token figure, and the benchmark scaling numbers all rest on Brown’s account in one interview, not on a published proof or third-party replication. Any operator weighing whether multi-agent orchestration is worth building toward should treat the 10,000-agent number as a demonstration of scale, not yet a benchmark for efficiency.
Based on Noam Brown’s interview with Dwarkesh Patel, published September 17, 2026 on Dwarkesh Patel’s podcast.
