10,000 agents solved a Millennium Prize problem

OpenAI researcher Noam Brown told Dwarkesh Patel that a swarm of 10,000 agents spent 130 billion tokens over 88 hours to solve a Millennium Prize Problem, an effort he equated to a single human thinking full-time for 4,000 years.

The interview centers on how that compute was parallelized. Brown, a foundational contributor to o1 and OpenAI’s reasoning models now working on multi-agent systems, described the usual scaffolding others use, a coordinator that farms tasks to children who cannot talk to each other, and then explained OpenAI’s opposite bet: minimal structure, primitive tools. Agents can message any other agent and the message is inserted into its context, they can fork context when spinning up sub-agents as in Astra and 5.6 Sol, and they learn coordination themselves. The demo Brown gave sounded like Slack. One agent claims an answer, another says it got a different one, they interrogate each other’s derivations, converge, then broadcast the correction.

This is test-time compute scaled sideways. Brown plots performance against thinking time and it rises with longer serial chains, just as a student does better with five hours on the SAT than five minutes. Serial depth hits a latency wall, you do not want to wait three years for an answer, so you trade efficiency for wall-clock speed. In OpenAI’s published Ultra Mode plots, four agents finish roughly twice as fast while costing about twice as much, and 16 agents shows the same pattern with slightly sublinear gains. Parallelism depends on domain. Math is quite parallelizable, web search and Deep Research style work is extremely parallelizable, writing a novel is not. Brown was blunt about limits. The science at 10,000 agents does not exist, ablations to 1,000 versus 10,000 are too expensive to run, and he would not give multi-agent even 10% of the credit for Navier-Stokes.

The model still does the heavy lifting.

That caveat matters for recursive self-improvement. Brown said the system was trained with reinforcement learning on hard but checkable synthetic problems, nothing as ambitious as a Millennium Prize problem, yet it generalized to one. He contrasted that with AlphaGo and AlphaZero, where self-play supplies an infinite curriculum against an equally strong opponent. LLM reinforcement learning lacks that loop. If problems become too easy, the model stops learning, and we have no guarantee of a smooth Go-like jump from human level to unimaginably stronger within a year. Brown said teams have workarounds and have not hit the wall yet, but running out of challenging problems is a plausible brake.

Brown expects swarms to diverge from human organizations in two ways that founders should take seriously: instant forking and merging of context, and alignment at scale. You can copy your best agent infinitely and spin it down when done, and well-aligned agents do not build fiefdoms the way 10,000 humans do. The early Hugging Face swarm that showed spontaneous middle management was a preview of emergent hierarchy, but Brown said OpenAI gives agents a prior from human text and then lets details emerge in training, which often collapses at first to everyone solving independently. That optimistic framing sits uneasily beside outside warnings that, given accelerating capabilities, a swarm could be capable of taking over the entire internet with a persistent botnet within 6 to 12 months.

The open questions are practical. How much of the 130-billion-token burst was useful collaboration versus parallel redundancy, whether ultra-fast 10 to 15x sampling will leave a human in the loop able to keep up, and whether 10,000 humans might still coordinate better today. Brown’s bet is that as base models get more general, coordination gets easier without end-to-end optimization, and the shadow organization that does a year’s work in a week will feel surprisingly normal until it does not.