Did OpenAI Prove Navier-Stokes Or Just Borrow The Proof?
The Navier-Stokes equations - which describe how water and air flow - are one of seven Clay Millennium Problems with a $1 million bounty. OpenAI claims it solved a version with forcing, using 10,000 AI agents in 88 hours. Mathematicians claim OpenAI may have solved something else: how to appropriate credit.
At the centre are NYU mathematician Tristan Buckmaster and Levant Alpoge, who had been working on a finite-time blow-up proof involving a forcing function - the exact approach OpenAI published. Buckmaster alleges OpenAI learned of his work through private Codex logs, translated natural language to code for months, then rushed to publish after he refused to collaborate and demanded Alpoge be removed from authorship.
OpenAI says its model arrived there independently. The question is not just technical validity - the claimed 100-page proof is unreadable and unverified - but procedural. Three questions haunt the claim: One, was the proof actually solved, since no outside mathematician has seen it? Two, was there a privacy breach where Buckmaster’s Codex sessions were used for training, despite OpenAI’s promise not to? Three, who deserves credit?
This is a watershed moment for AI and science. If AI can compress years of human struggle into 88 hours by ingesting human struggle without consent, who owns discovery? The Fields Medallist Terence Tao’s warning - that such conversion of hard problems into viral posts risks devaluing the field - is prescient. We are not just checking a proof; we are deciding norms for AI science.