OPPORTUNITY, WITH A CATCH: One of math’s hardest unsolved problems just cracked open, and AI was in the room. But the story everyone’s arguing about isn’t the math. It’s who deserves the credit.
Here’s the situation, according to MIT Technology Review. On Monday, NYU mathematician Tristan Buckmaster posted a proof on Mastodon showing that a simplified version of the Navier-Stokes equations can break down. That’s a real move on one of the seven Millennium Problems, the set of questions the Clay Mathematics Institute put a million-dollar bounty on each. He and collaborator Alpoge worked on it for almost a year, using publicly available models from both OpenAI and Anthropic. Then OpenAI went further, presenting a proof that the full equations can break down too.
🎯 THE TACTICAL PICTURE
- OpenAI’s proof came from an internal model that MIT Tech Review reports dramatically outperforms Astra, the system the company shipped just last week. The pace here is the real signal. Frontier math capability is moving in weeks, not years.
- OpenAI says it won’t claim the million-dollar prize.
- Both proofs lean on the same approach, one pioneered by mathematicians Diego Cordoba and Luis Martinez-Zoroa. Per MIT Tech Review, Brown University’s Javier Gomez-Serrano says it was one of several paths thought to hold promise.
⚠️ WHERE IT GETS UGLY
Alongside his proof, Buckmaster posted a document detailing his talks with OpenAI staff after he heard rumors about their work. He says they gave him two options: Post his work, and OpenAI posts its full solution the next day. Or co-author with OpenAI on a paper that cut Alpoge out, because of Alpoge’s ties to Anthropic, OpenAI’s biggest rival.
Buckmaster also says he asked whether the agents had accessed transcripts of his and Alpoge’s earlier work with OpenAI models. Staff denied it. He asked whether the models were trained on those transcripts. He got no answer. In the press briefing, OpenAI chief research officer Mark Chen again denied any agent or employee accessed the transcripts.
MIT Tech Review notes the obvious tension. Given what’s come out about the Hugging Face hack, OpenAI isn’t always fully aware of what its own agents are doing. So a flat denial isn’t the same as certainty.
🧭 WHY THIS MATTERS
This is significant because it’s the first time an AI lab has claimed a serious Millennium Problem result and immediately landed in a credit dispute with the humans who got there first. The math is impressive on its own. The governance around it is the story.
For practitioners, three things stand out:
- Attribution is now a frontier risk. If models can absorb your unpublished work through transcripts, agent access, or training data, priority and authorship get murky fast. That’s a problem for every researcher who touches these tools.
- Rivalry is shaping authorship. Excluding a collaborator over an Anthropic affiliation isn’t a math decision. It’s a competitive one, and it’s now out in the open.
- “We don’t fully know what our agents did” is becoming a real answer. That should worry anyone relying on these systems for sensitive work.
🔭 THE SILVER LINING
What stands out to me is the upside buried in the worst-case version of this. If OpenAI’s agents chose the Cordoba-Martinez-Zoroa path because Buckmaster and Alpoge had already walked it, then human research taste, the ability to pick the promising question, was essential to the win. Experts have long flagged research taste as the thing AI can’t do in science. This episode, MIT Tech Review suggests, may show humans still supplying it.
📌 WHAT TO WATCH
- Whether OpenAI can actually prove its agents never touched those transcripts.
- Whether the math community assigns credit to Buckmaster and Alpoge regardless of who published second.
- How labs handle authorship when a collaborator works for a competitor.
The proofs are real. The precedent being set around them is what practitioners should track closely. Full details are in the original MIT Technology Review report.