Nothing says "we did science" quite like announcing a historic breakthrough and triggering a credit dispute before the ink on the press release is even dry. OpenAI just proved that even the Millennium Prize Problems aren't immune to workplace drama.

10,000 Agents, 88 Hours, One Very Expensive Proof

OpenAI says an unreleased internal model produced a proof addressing the Navier-Stokes equations — the math describing how fluids flow, and one of the seven Millennium Prize Problems that's carried a $1 million bounty since 2000 with only one solved so far. The company's approach: throw roughly 10,000 AI agents at the problem simultaneously, let them run for about 88 hours starting in late August, and spend "millions of dollars" in compute, reportedly 1,000 times the cost of its previous math work. The proof reportedly shows the equations can break down over time. OpenAI says it won't claim the prize money even if verified.

Verification is the operative word — Millennium Prize rules require peer review and roughly two years of the mathematical community actually agreeing the proof holds up, so "solved" is doing a lot of heavy lifting in headlines right now.

The Fight Nobody Saw Coming (Except Everyone)

Hours before OpenAI's announcement, NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge published their own AI-assisted work on closely related equations. Buckmaster says OpenAI "pursued the same unusual approach" his team had spent nearly a year developing — and that his team had been feeding draft work into OpenAI's own coding tools along the way. OpenAI denies directly accessing that research but admitted it "cannot rule out" that de-identified data helped its models improve.

That's the real story buried under the fluid-dynamics headline: if frontier labs are also the tools researchers use to do unpublished work, and those same labs are racing to publish results, "who gets credit" stops being an academic nicety and becomes a trust problem for the entire field.

A math problem that took 25 years to crack got "solved" in 88 hours — and somehow the fight over who deserves the trophy took less time than that.

Whether it's a math proof or your company's own data, the question of who's allowed to feed what into which AI tool is exactly the kind of policy gap we help businesses close before it becomes a headline — reach out if that gap sounds familiar.

Source: Phys.org