• Treczoks
    link
    fedilink
    English
    arrow-up
    3
    ·
    3 hours ago

    My son has tried to explain this to me - he is the mathematician in the family - and what the LLM produced is a load of source files for an automatic proofer. So far, the files are syntactically correct, but if you know programming, getting a project through the compiler without error messages is no way a guarantee for a working software.

    And indeed, my son told me just a few hours after release that human reviewers had already found tautology issues within the proof.

    But don’t ask me for details - my time with math at that level was ages ago…

    • badabim
      link
      fedilink
      English
      arrow-up
      1
      ·
      26 minutes ago

      The point of Lean (or any other similar proof assistant) is that if the program typechecks, then you have a proof. Hence, supposing that no Lean bug were exploited, all we would need to check is that the statement of the theorem in Lean matches what we actually want to prove. If this is the case, then we have a witness of the existence of a proof, despite it being a gigantic messy heap of definitions and lemmas.

      An actual human-digestible proof is still valuable, because the proof itself would contain insights valuable for mathematics in general.