@mathieu_miamiA math professor just went public with a wild story about OpenAI trying to control academic work. Tristan Buckmaster, who teaches at NYU's Courant Institute and won a Clay Research Award, dropped a detailed account on September 8. He and Levent Alpöge had spent a year on the Córdoba / Martínez-Zoroa program, which deals with smooth forcing. On August 15, they got blowup results for Boussinesq and 3D Euler, and by August 22, they had verified it all in Lean. They were paying for compute themselves, and every draft of the project lived in private Codex sessions on OpenAI's servers. The research area is so niche that almost nobody else in the world was working on it. So when rumors started flying, Buckmaster emailed OpenAI on September 3 to head off any conflict. OpenAI wrote back asking for details so they could 'avoid competing' and even offered free compute. Then on a Sunday, Sébastien Bubeck, who heads OpenAI's math team, called Buckmaster but left Alpöge out of the conversation. During that call, Bubeck said an internal model had come up with a roughly 100-page proof of finite-time blowup for forced Navier, Stokes. He claimed the model was 'simply given the problem statement,' but that story fell apart quickly, there was a whole team involved, tons of compute, and the model had been trained on easier equations first, including Euler. The prompt shown to Buckmaster had itself been crafted using Codex. When Buckmaster asked when the first prompt went out, they dodged, then said it was in the past few days, after they'd heard about his work. He asked if the model had access to the private Codex sessions with all his drafts. They said the model doesn't look up user data. He asked again about training, and got no answer. Then came the real ask. OpenAI gave two options: either Buckmaster posts Euler and they post Navier, Stokes the next day, or he posts Euler and then writes up Navier, Stokes alone, crediting their model while Alpöge gets cut. Bubeck pushed twice to remove Alpöge from the paper, and the reason was that Alpöge works at Anthropic. Buckmaster turned down both options and said he'd go public if they released it that way. Bubeck didn't have a scientific argument. He asked, 'Why would you ruin your career?' When Buckmaster said he's an academic, Bubeck replied, 'If you don't want me to be nice, then I don't have to be nice.' Later, Alpöge got a text suggesting a one-on-one talk, with Bubeck saying, 'I don't know if Tristan is being fully rational right now.' That's not how you treat a collaborator, that's how you try to break one apart. So if even half of this is true, we've got a lab that found out two academics were about to finish a path almost nobody else was on, spun up a team on that same path, dodged questions about unpublished drafts sitting on their servers, tried to dictate the announcement order, and tried to remove a coauthor because of his employer. When that failed, someone went for a career threat. Bubeck has publicly called the allegations 'false and inflammatory' and says he'll respond more fully. So this is one side, but it's a side that put names, dates, and direct quotes on the record under his own name and reputation. What a mess. #OpenAI
View original post
