Sam Altman

FINALLY! A levelheaded, actually useful and informative article! Read PDF for free in the quoted link. TLDR: We literally cannot trust AI blindly when it comes to auto-formalisation of mathematical proofs (in Lean) written originally in human natural language. A lot can go wrong while the Lean code still compiles and gives us something that looks like a legitimate formalisation of the proof: 1) the original proof may be right but the AI formalises a different proof that only partially corresponds to the actual proof. 2) the original proof may actually be wrong but the AI formalises a different proof that may be right but may not necessarily correspond to the original theorem; 3) the original proof may be wrong and lean can also formalise the wrong proof that’s wrong but still compiles! 4) the proof may be right and the AI formalises it correctly but makes extra claims that are not in the original proof (mismatched scope) … So many other scenarios of what can go wrong without us noticing immediately or at all! This paper does the work and uses a couple of examples from @OpenAI recent math proof formalisation trophies! Guess what? AI hallucinated and nobody at the lab noticed 🤦🏻‍♀️