• molehill_siesta@lemmy.blahaj.zone
    link
    fedilink
    English
    arrow-up
    7
    arrow-down
    1
    ·
    3 days ago

    In further discussions with Claude…

    Nah, this ain’t it. I too wish it was just using an LLM to get a head start on translation, but apparently they’re getting the LLM, which has been trained on mathematical proofs, to generate mathematical proofs from text that the LLM is told contains mathematical proofs.

    That’s hallucination city.

    • gravitas_deficiency@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      1
      ·
      3 days ago

      I need to be clear here: it’s absolutely possible that you’re right. That’s why a huge part of my comment above is emphasizing that this dudes conclusions drawn from the notes as passed through Claude are largely premature, and need to be verified MUCH more rigorously.

      But I do think that categorically disregarding this - particularly, without even understanding the nuance of his methodology or doing any sort of additional verification or isolated reproduction of results with different (non-LLM) tools - is also premature.