OpenAI claims AI agents found a Navier-Stokes breakdown as credit dispute erupts
- OpenAI says roughly 10,000 AI agents worked for 88 hours to find a Navier-Stokes solution that becomes singular in finite time, with fluid velocity reaching infinity.
- The claimed counterexample would resolve the Navier-Stokes regularity question by showing that the equations can produce an unphysical result under some conditions, rather than proving the equations are broadly unusable.
- Navier-Stokes is one of the Millennium Problems, and a confirmed solution would qualify for the $1 million Clay Mathematics Institute prize and be the first major open math problem solved by AI.
- NYU mathematician Tristan Buckmaster and Anthropic-affiliated researcher Levent Alpöge had separately made recent progress using AI tools, then questioned whether OpenAI's effort drew on their unpublished work.
- OpenAI says it did not knowingly use Buckmaster and Alpöge's latest results, while Buckmaster publicly disputes that account and OpenAI's reported compute spend reached millions of dollars.
Hacker News opinions
The article says Navier-Stokes is one of six Millennium Problems, but there are seven. One has been solved, yet it remains on the Millennium Problems list.
I do not see a real connection between solving Navier-Stokes and developing materials or curing diseases. That quote sounds like OpenAI is using the result to sell confidence in expensive future projects.
I see Navier-Stokes as a test of how far general intelligence has progressed. Many hard scientific problems require intelligence, even if they are not directly the same problem.
I would be wary of using services like this for unpublished work after this episode. The compute gap gets worse if researchers have to avoid hosted AI systems.
I am a mathematician, and I am uneasy about what these tools mean for the profession. OpenAI probably would not spend $15 million to scoop most of my work, but the possibility still changes how people use ChatGPT.
Researchers and graduate students I know are closing work-in-progress repos, keeping progress out of LLMs, and discussing lab-level self-hosted models. Universities should run open models themselves.
I think the safer alternative is to make work public early. A public record makes it easier to establish who deserves credit.
I think OpenAI used Buckmaster and Alpöge's Codex work to bootstrap its proof. OpenAI started pursuing the questions only after those submissions, and it has not ruled out model training data as a source.
If OpenAI offered Buckmaster sole authorship only if Alpöge was removed, that is scientific misconduct. The admission that product-use data might have improved the model makes the attribution problem worse.
I do not think everyone can simply opt out of this kind of science. Expensive research has always required deep pockets, and a $200 monthly AI subscription is cheap beside many commercial tools.
Credit still matters because these are different claims about AI use. OpenAI describes thousands of largely autonomous agents over 88 hours, while Buckmaster and Alpöge used guided AI work for months.
The right question may be most of the work. If OpenAI learned from users' private back-and-forth and then outspent them to announce the breakthrough, that is hard to defend.