• Plopp@lemmy.world
    link
    fedilink
    English
    arrow-up
    19
    ·
    8 months ago

    Yes please, more scientific papers and non-fiction, and less internet comments.

  • drawerair@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    ·
    8 months ago

    Were the research papers used to train Ibm’s language model not public? If the papers are public and there are no restrictions, Chatgpt and Gemini can use those right? If Chatgpt and Gemini used those, what’s the benefit of using Ibm’s language model?

    Odd that heliophysics (sun physics) was specified. It’s under astrophysics.

    • ReveredOxygen
      link
      fedilink
      English
      arrow-up
      5
      ·
      8 months ago

      Presumably they gave higher weight to the papers and lower weight to blogspam and reddit

        • ReveredOxygen
          link
          fedilink
          English
          arrow-up
          4
          ·
          8 months ago

          Not really? It doesn’t have sources, all the information is just thrown into a heap in its neural net. You can somewhat trick it by telling it that it’s an expert in the field, but it’s far from foolproof

    • Cosmicomical@lemmy.world
      link
      fedilink
      English
      arrow-up
      3
      arrow-down
      2
      ·
      edit-2
      8 months ago

      Let me mansplain this.

      There are 2 main ways to alter the way llms like chatgpt respond: 1) training & fine-tuning and 2) prompting.

      Prompting means feeding the LLM information in the form of text, so it has actual stuff in what you could call its “short-term memory”. That can be done interactively and only allows you to feed it a limited number of words.

      Training and tuning are much slower processes and are used to build up and refine the “long-term” memory and to give importance and detail to specific topics, or to reinforce some kind of trait, like tone or type of answers it can produce.

      I’m pretty sure what ibm and nasa did was a round of fine-tuning with material in that area of knowledge, not just prompting.

    • fruitycoder
      link
      fedilink
      English
      arrow-up
      1
      ·
      8 months ago

      The good news is you could use the RAG model to get the papers related to your search.

  • rottingleaf@lemmy.zip
    link
    fedilink
    English
    arrow-up
    3
    ·
    8 months ago

    So, just like a model to generate pics gives me characters with 11 fingers on their hands, and a model to generate something gives me POV-Ray code to draw a cylinder when asked for a Moebius stripe, this thing is going to tell us of many beautiful worlds with Pi ranging between 2 and 33?

    • pewter@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      8 months ago

      In a non-Euclidean space you can have values of pi that aren’t even constant. The first that came to mind was the surface of a sphere. This is actually an interesting problem.

      • rottingleaf@lemmy.zip
        link
        fedilink
        English
        arrow-up
        2
        ·
        8 months ago

        Yes, bad choice of words, the point is that the model is going to generate loads of garbage imitating scientific text, as if real humans making errors (or sometimes intentionally putting out dubious works, fabricating experimental data etc) were not enough.