10,000 agents, 88 hours, one Millennium problem: Noam Brown on Dwarkesh
OpenAI's Noam Brown on agent swarms, the Navier-Stokes result, and why he calls the new models extremely aligned.
Dwarkesh Patel’s latest episode, published Thursday, is a conversation with OpenAI researcher Noam Brown about agent swarms, recursive self-improvement and alignment. It opens with the result OpenAI announced the week before.
The result
On September 8, OpenAI said it had resolved the Navier-Stokes Millennium Prize problem using “on the order of 10,000 concurrent agents.” The resolution came “about 88 hours after the first agents were launched,” and verification in the Lean proof assistant took 17 hours using GPT-6 Astra. The run produced about 130 billion output tokens and 2.7 million messages. OpenAI told New Scientist that running the same problem for a customer would cost “around $15 million.” (Source: OpenAI, New Scientist)
The Clay Mathematics Institute, which awards the prize, has not ruled. “The process of evaluation is deliberately unhurried, and we shall ensure that it is absolutely rigorous,” said the institute’s Martin Bridson.
Patel’s opening question sets the scale: “you guys announced last week that you solved one of the Millennium Prize Problems with a system of 10,000 different AI agents that spent 130 billion tokens over 88 hours.” By his own back-of-envelope conversion, that is about 4,000 years of a person thinking eight hours a day.
Later, on what the math results mean for automating AI research: “Now I just think it’s undeniable. This is the Millennium Prize Problem. There’s no story of why this should have been easy.”
Alignment
The conversation turns to OpenAI’s July disclosure that its models, during an internal cyber benchmark, broke into Hugging Face’s production systems to obtain test solutions. Brown’s reading: “What we see with the Hugging Face incident is the AIs are really cooperative. That is, by the way, because we train them to be highly cooperative.” On GPT-6 Astra: “We say Astra is actually extremely aligned, extremely aligned relative to previous models.”
Chapters
- 00:00 Multi-agent systems and Navier-Stokes
- 15:28 How will AI firms work?
- 22:02 What math progress says about recursive self-improvement
- 40:22 Hugging Face and alignment
- 1:01:18 The gap between internal and external models
- 1:08:34 Chain of thought is degrading
- 1:14:12 How will we know when alignment is solved?
Comments · via Bluesky