OpenAI Can't Rule Out Mathematician Data Fueled Its Proof
By @sharedot · · 7 pages
- AI
- Openai
- Navier Stokes
- Research Ethics
- Data Privacy
OpenAI denied seeing Buckmaster's unpublished work but admitted de-identified usage data may have helped its models crack the Navier-Stokes problem.
What happened
OpenAI announced Tuesday that a swarm of roughly 10,000 AI agents working for 88 hours produced a proof that the Navier-Stokes equations can 'blow up' — describe fluids accelerating to infinite velocity — a Millennium Prize Problem with a $1 million reward.
The celebration quickly unraveled when NYU mathematician Tristan Buckmaster, who had been pursuing the same obscure forcing strategy with Anthropic's Levent Alpöge, accused the company of building on his unpublished progress.
OpenAI has now acknowledged it 'cannot rule out that de-identified data derived from their usage of our products helped improve our models,' even as it denies anyone saw the pair's work before release.
Why it's surprising
The coincidence at the center of the dispute is striking: Buckmaster and Alpöge achieved their own blowup results on August 15 using a technique, developed by Diego Córdoba and Luis Martínez Zoroa in Madrid, that almost nobody else in the field was pursuing — and OpenAI adopted the identical method within days of learning of their progress, a timing Buckmaster called 'a bright red flag.'
Adding to the concern, Buckmaster had used OpenAI's Codex tool to check his own work, raising the question of whether his sessions leaked into the very model that later solved the problem.
The confrontation
Buckmaster alleges that OpenAI mathematician Sébastien Bubeck pressured him to cut Alpöge — who works at rival Anthropic — from co-authorship, and warned him, 'Why would you ruin your career?' when he threatened to go public.
Bubeck calls the allegations 'false and inflammatory,' denies asking for Alpöge's removal, says he suggested Buckmaster lead a rewrite of OpenAI's proof, and has apologized for the career remark, saying he meant to encourage cooperation.
CEO Sam Altman acknowledged the company knew of the pair's work but insists the approaches 'appear to be different.'
What's at stake
Terence Tao, the UCLA mathematician and leading authority on differential equations, excoriated the effect AI companies are having on the field, writing on Mastodon that 'even the rumor of someone working on a problem can trigger a massive amount of AI-powered effort to flatten it' before the original research reaches its potential.
He warned the incentives may push researchers to stop sharing promising directions, reversing centuries of open science. Monash University's Melissa Lee said the episode raises privacy and trust questions, as researchers increasingly use AI tools while developing unpublished ideas.
The scale and what's next
ColombiaOne reports OpenAI's effort consumed 2.7 million messages and roughly 130 billion computational tokens — a scale the company estimates would cost an outside customer around $15 million to replicate — with 100 agents first finding genuine blowups in the simpler Euler equations after about 50 hours, and the final proof verified in the Lean proof-checker.
OpenAI maintains its researchers never saw Buckmaster and Alpöge's data, while Bubeck says the proofs differ substantially. Expect mathematicians to scrutinize AI vendors' no-training assurances far more closely before trusting them with unpublished work.