A story we follow
OpenAI announces a Navier–Stokes blowup construction
Tracks OpenAI's announced forced Navier–Stokes result, its multiagent workflow, mathematical scope and disputes over priority and research-data use.
A story page follows one specific event across podcast discussions, with an overview and a timeline of what changed. It updates when new episodes discuss the event, and you can get those updates by email or push. How our formats work
By email or push, when we publish an update to this story. It does not subscribe you to other stories.
Overview
OpenAI announced that an internal model had produced a construction showing blowup in the forced Navier–Stokes equations, with a Lean formalization, and matched it to conditions C and D of the Millennium Prize formulation. Moonshots discussed a reported workflow of 10,000 agents over 88 hours and estimated the inference cost at 6.5 million dollars. The panel saw a breakthrough in automated research, and it passed on disputed accounts of authorship and worries that researchers could be scooped by their own model provider. The announcement says OpenAI will not claim the prize and that its investigation ruled out any influence from Buckmaster's recent Codex prompts.
In a Diary of a CEO debate, the safety-focused side used the result as a warning sign about AI-driven AI research. A skeptic called that a leap, and the debaters gave the run's length as 11 days. On the Dwarkesh Podcast, OpenAI researcher Noam Brown said multi-agent techniques deserve less than 10 per cent of the credit and that the core reason was a very strong general-purpose model. OpenAI has not run a single-agent baseline, so the benefit of 10,000 agents is unmeasured. He said the same internal model, not available outside OpenAI, has produced solutions to other unsolved problems.
On the Latent Space interview with Alex Zhang, published 2 October 2026, Zhang offered an outside reading. He guessed that OpenAI probably did not use a recursive language model, but that it clearly used a swarm of agents with shared context such as a shared file system. In his view a model of the Astra class is smart enough to produce such a proof if given the right information, so the harness was not the important part. He thought a very long search across many agents, possibly with researchers' intuition about what to explore, produced that information. He admitted he was unsure about the researcher part. The interviewer put the cost at about 40 million dollars at public pricing, a different estimate from the earlier 6.5 million figure, and Zhang found it lower than expected. Later in the same conversation, Zhang said mathematicians objecting to the result might have had a stronger case had many of them not worked with OpenAI before. An interviewer answered that this was an ad hominem argument and defended their right to object. The passages do not state the mathematicians' objections in detail.
None of these discussions counts as independent mathematical validation.
Recent coverage
Changes to our coverage published in the past seven days. These are site update dates; podcast discussion dates and sources are in each version.
No changes to this account were published in the past seven days.
What changed
Dates show when each podcast discussion was published.
-
Interpretation
On the Latent Space interview with Alex Zhang, Zhang guessed from outside that OpenAI used a swarm of agents with shared context but probably not an RLM, and that the harness mattered little because a strong model needs only the right information, which a long agent search supplied. An interviewer cited roughly 40 million dollars at public pricing for the run, and Zhang found that lower than expected. Zhang also said the objecting mathematicians would have had a stronger case if many had not worked with OpenAI earlier, which an interviewer called ad hominem.
-
Interpretation
On The Diary of a CEO, Nate Soares's side used the Navier–Stokes claim to argue that AI-driven research on better AI architectures could arrive sooner than expected. He described the run as 10,000 OpenAI agents over 11 days, which differs from the reported 88 hours, and said OpenAI had raced human researchers who were close to a solution. A skeptic called the inference a leap. He said it was unconfirmed how much OpenAI relied on human scientists' work and that the AI had not worked alone. Speakers also mentioned unverified rumors of other Millennium claims.
-
New information
On the Dwarkesh Podcast, OpenAI researcher Noam Brown said he would not give multi-agent techniques even 10 per cent of the credit for the result. In his account, the main reason was a very strong general-purpose model that can run over long horizons and think in parallel. He said OpenAI has not tested how long a single agent would take, so the run is one data point, and it is possible that 10,000 humans currently coordinate better than 10,000 agents. He said the internal model is not available to outsiders and has produced solutions to other unsolved problems as well. He called this an unfair advantage with trade-offs he does not know how to weigh. He said that about two weeks before the result, a frontier-lab researcher bet him 1,000 dollars that it would take until after 2027, and he took the bet. The interviewer described the 130 billion tokens as roughly 4,000 years of full-time human thinking.
-
New information
The panel interprets the announced result as evidence that generalist agents can tackle major mathematical problems. It discusses compute costs, concurrent Euler work, disputed credit negotiations and possible training contamination, leaving the latter as a concern rather than an established finding.
Podcast discussions
- Latent Space Academia is for Ambition — Alex Zhang, MITOct 2, 2026
- Dwarkesh Podcast Noam Brown – Agent swarms, alignment, & recursive self-improvementSep 17, 2026
- The Diary of a CEO AI Debate Ed Zitron, Andrew McAfee, Nate Soares, Roman YampolskiySep 17, 2026
- Moonshots with Peter Diamandis Why Jensen Huang Believes We’ve Reached AGI and Inside OpenAI’s German Website Hijack | #287Sep 9, 2026
Sources
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Our coverage
An MIT researcher guesses most agent-swarm search is wasted
MIT PhD student Alex Zhang guesses that about 95% of an agent swarm like OpenAI's Navier–Stokes run burns tokens on dead ends. He still calls OpenAI's approach right and says getting a swarm to converge is the hard part.
Noam Brown expects faster AI research, but not an overnight explosion
OpenAI researcher Noam Brown says AI could make AI research perhaps three times faster. He argues that, unlike maths, research waits on sequential experiments and GPUs, so he doubts a 100-fold jump, and says his uncertainty is large.
Brown credits OpenAI's model, not its swarm, for Navier–Stokes result
OpenAI researcher Noam Brown says multi-agent coordination deserves not even 10% of the credit for the claimed Navier–Stokes result. He adds that OpenAI has not run a single-agent comparison or measured what 10,000 agents added.
After OpenAI's Navier–Stokes result, panel splits on self-improving AI
Panelists split over whether OpenAI's Navier–Stokes proof means AI will soon improve AI research; a skeptic called math narrower, while a risk-focused guest gave about 10% odds a bigger agent swarm designs a better AI architecture in six months.
What OpenAI's 10,000 agents proved about Navier–Stokes blowup
OpenAI says an internal model running about 10,000 agents for 88 hours produced a forced Navier–Stokes blowup construction with a Lean-checked proof. Moonshots panelists stressed it concerns idealized fluids and noted disputes over credit.
Version history
-
Oct 2, 2026 · Version 4
On the Latent Space interview with Alex Zhang, Zhang guessed from outside that OpenAI used a swarm of agents with shared context but probably not an RLM, and that the harness mattered little because a strong model needs only the right information, which a long agent search supplied. An interviewer cited roughly 40 million dollars at public pricing for the run, and Zhang found that lower than expected. Zhang also said the objecting mathematicians would have had a stronger case if many had not worked with OpenAI earlier, which an interviewer called ad hominem.
-
Sep 29, 2026 · Version 3
On the Dwarkesh Podcast, OpenAI researcher Noam Brown said he would not give multi-agent techniques even 10 per cent of the credit for the result. In his account, the main reason was a very strong general-purpose model that can run over long horizons and think in parallel. He said OpenAI has not tested how long a single agent would take, so the run is one data point, and it is possible that 10,000 humans currently coordinate better than 10,000 agents. He said the internal model is not available to outsiders and has produced solutions to other unsolved problems as well. He called this an unfair advantage with trade-offs he does not know how to weigh. He said that about two weeks before the result, a frontier-lab researcher bet him 1,000 dollars that it would take until after 2027, and he took the bet. The interviewer described the 130 billion tokens as roughly 4,000 years of full-time human thinking.
-
Sep 25, 2026 · Version 2
On The Diary of a CEO, Nate Soares's side used the Navier–Stokes claim to argue that AI-driven research on better AI architectures could arrive sooner than expected. He described the run as 10,000 OpenAI agents over 11 days, which differs from the reported 88 hours, and said OpenAI had raced human researchers who were close to a solution. A skeptic called the inference a leap. He said it was unconfirmed how much OpenAI relied on human scientists' work and that the AI had not worked alone. Speakers also mentioned unverified rumors of other Millennium claims.