News

Noam Brown wouldn't give the 10,000-agent swarm 10% of the credit for solving Navier-Stokes

He also says OpenAI has a model inside the building that keeps producing solutions to unsolved problems, and no good answer on what to do about that.

By Think Facility · September 18, 2026 · OpenAI science

Dwarkesh Patel published an interview with OpenAI researcher Noam Brown on September 17, 2026, and opened with the numbers behind OpenAI's Millennium Prize Problem result: a system of 10,000 AI agents that spent 130 billion tokens over 88 hours.

Brown wanted the credit pointed somewhere else. "There's one thing I want to make clear," he says of the effort. "I wouldn't even attribute 10% of the credit to multi-agent."

Published

Sep 17, 2026

Agents on the run

10,000

Tokens spent

130 billion

Time taken

88 hours

Agents in published work

up to 16

Multi-agent credit

not even 10%

What he says did the work

The thing that solved it, on his account, is the model underneath. "The reason why we're able to do this is because we just have a general-purpose, very strong model." Multi-agent is "flashy and new, and that probably gets disproportionate credit for that reason".

He's blunt about how little of this has been measured, too. OpenAI's published work goes up to about 16 agents (16, against the 10,000 that ran), and pushing the science to 10,000 is too expensive to do properly. OpenAI hasn't run Navier-Stokes with a single agent either, so nobody has a baseline to set the swarm against. One data point, is how he puts it.

The model the outside world can't use

Scott Aaronson wrote on September 15, 2026 that he'd been told the labs were sitting on solutions to open problems. That was a rumor reaching him second hand. Turns out an OpenAI researcher will say it on the record, about an hour into this conversation.

"It's not just Millennium Prize Problems," he goes on. "There are many solutions to unsolved problems that people have been able to get out of this model." Asked what you do about that, he says OpenAI doesn't have a good answer, and calls it "an unfair advantage".

He had posted something close to this on September 8, without saying the model was one the public can't reach.

Chain of thought is degrading

Then the monitoring, which I think is the heaviest thing in the episode. Brown says OpenAI is "already seeing signs that chain-of-thought monitorability is degrading", and the mechanism he gives is hard to get out of: "every time you intervene based on your observations of the chain of thought, you are implicitly applying a tiny bit of pressure for the model to then hide its chain of thought".

So the tool that lets you see the model scheming gets worse every time you act on what it shows you. (He says they're trying to work out exactly why, because they want to reverse the trend.)

On how much faster AI research runs once AI is doing it, he's careful. "If you put a gun to my head and ask me for a number, I could see things going 3x faster. That is huge." He won't go near the bigger multiples, though: there's "a big difference between that and 100x faster", and experiments still need GPUs and time. (He says a lot of people in the field have very high error bars on this sort of thing, and puts himself among them.)

That line is about the Hugging Face incident, which Brown calls people's "first real exposure to multi-agent coordination". He says the alignment metrics OpenAI had beforehand mostly looked pretty good, and that he thinks they underestimated how serious the concerning ones could be.

Brown doesn't say when the math model reaches the outside world.