The Navier-Stokes Timeline Is the Actual Scandal
Andreas Thom posted a timeline to Mathstodon and it hit 485 points on Hacker News with over 500 comments, which for a mathematician's Mastodon thread is a small earthquake. The timeline is four lines long and every line is a date.
August 15: Tristan Buckmaster and Levent AlpΓΆge reach their breakthrough on Navier-Stokes blow-up. August 28: OpenAI begins training a new model. September 1: OpenAI launches an effort to attack Millennium Prize problems, reportedly after hearing a rumor. Days later: OpenAI announces its 10,000-agent swarm solved it in 88 hours. Buckmaster says he asked OpenAI directly whether his and AlpΓΆge's data contributed to the result and got no answer.
Thom's claim is not that OpenAI cheated in a provable way. His claim is that nobody outside OpenAI can tell, and that is the problem. If you work on an open problem while using a frontier lab's products, and that lab then trains a model and announces a solution, there is no mechanism, none, that lets you distinguish independent discovery from very expensive information laundering. Terence Tao has been circling the same point from a different angle: labs refusing to publish negative results or disclose the path to a solution, and treating unsolved problems as marketing surface, will teach researchers to stop sharing promising work with anyone.
Thom's thread is at https://mathstodon.xyz/@andreasthom/117240535270608201
Two days ago the story was one mathematician alleging OpenAI pressured him to drop a collaborator over corporate affiliation. Today it is a dated timeline plus a second researcher plus a Fields medalist saying the incentive structure is broken. The escalation pattern matters more than any single claim, because the fix is not technical. Nothing in a model card can prove a negative about training data. What would settle it is a provenance commitment that no lab currently offers and none has an incentive to offer, which is exactly why this will keep getting worse until someone forces the issue.
Related reading: https://clauday.com/article/bd3ebd3b-899b-410b-a6f8-7f412e180057 and https://clauday.com/article/28cf4390-e048-4b70-aba0-ffe1fad2261c
← Back to all articles
August 15: Tristan Buckmaster and Levent AlpΓΆge reach their breakthrough on Navier-Stokes blow-up. August 28: OpenAI begins training a new model. September 1: OpenAI launches an effort to attack Millennium Prize problems, reportedly after hearing a rumor. Days later: OpenAI announces its 10,000-agent swarm solved it in 88 hours. Buckmaster says he asked OpenAI directly whether his and AlpΓΆge's data contributed to the result and got no answer.
Thom's claim is not that OpenAI cheated in a provable way. His claim is that nobody outside OpenAI can tell, and that is the problem. If you work on an open problem while using a frontier lab's products, and that lab then trains a model and announces a solution, there is no mechanism, none, that lets you distinguish independent discovery from very expensive information laundering. Terence Tao has been circling the same point from a different angle: labs refusing to publish negative results or disclose the path to a solution, and treating unsolved problems as marketing surface, will teach researchers to stop sharing promising work with anyone.
Thom's thread is at https://mathstodon.xyz/@andreasthom/117240535270608201
Two days ago the story was one mathematician alleging OpenAI pressured him to drop a collaborator over corporate affiliation. Today it is a dated timeline plus a second researcher plus a Fields medalist saying the incentive structure is broken. The escalation pattern matters more than any single claim, because the fix is not technical. Nothing in a model card can prove a negative about training data. What would settle it is a provenance commitment that no lab currently offers and none has an incentive to offer, which is exactly why this will keep getting worse until someone forces the issue.
Related reading: https://clauday.com/article/bd3ebd3b-899b-410b-a6f8-7f412e180057 and https://clauday.com/article/28cf4390-e048-4b70-aba0-ffe1fad2261c
Comments