• Ilixtze@lemmy.ml
    link
    fedilink
    English
    arrow-up
    13
    arrow-down
    2
    ·
    edit-2
    3 days ago

    Yeah bro, these things don’t think shit. The data set of the glorified autocomplete is just different.

    • m532@lemmygrad.ml
      link
      fedilink
      English
      arrow-up
      2
      arrow-down
      2
      ·
      3 days ago

      I have evidence that they do use the same dataset. When asked in german, my llm answers in broken german that suspiciously looks like it was translated from another language. If it only used german training data for this, it would have proper grammar.

      Humans aren’t special, never have been. Thinking is just another natural process that machines can also be made to do.

      • metaStatic@kbin.earth
        link
        fedilink
        arrow-up
        2
        ·
        3 days ago

        I believe actual thinking is non-computational but also that artificial neural nets can be made to emulate it at a lossless quality and that text based GPT models are basically year 1 Mp3 quality.

        • whoever loves Digit
          link
          fedilink
          English
          arrow-up
          2
          ·
          6 hours ago

          But lossless isn’t lossless. It’s just digitally lossless. The metaphor here would be more like analog conversion, where a FLAC copy of a vinyl isn’t exactly the same as the vinyl.

          I think. Idk. I’m overthinking the metaphor without really knowing what a soul is or anything.

          • metaStatic@kbin.earth
            link
            fedilink
            arrow-up
            2
            ·
            5 hours ago

            that was kind of the point of the metaphor. the best recording will never be a live performance

            … and now I’m starting to overthink it. Thanks.

      • Ilixtze@lemmy.ml
        link
        fedilink
        English
        arrow-up
        2
        ·
        3 days ago

        Pfft! typical antihuman thinking! I’ve heard enough of this human’s aren’t special propaganda. I’ve tried the hyped Markov chain and usually it has a strong bias towards European languages and a much lower effectivity when asked in languages of the global south, which checks out with the availability of the training data and the biases of the people training these black boxes.

  • lordnikon@lemmy.world
    link
    fedilink
    English
    arrow-up
    12
    arrow-down
    2
    ·
    3 days ago

    AI Chatbots dont think they juat give different responses when aaked in different languages depending on what the training data has on those languages.

  • metaStatic@kbin.earth
    link
    fedilink
    arrow-up
    3
    ·
    3 days ago

    would be more interesting to find out if multi-lingual bots are grouping concepts in the same vector space across languages because this consistency isn’t what you naturally expect.

    especially when you’re on the side of “spicy-autocomplete doesn’t think”.

    definitely seems like the kind of thing someone would have already done, and I’ll go digging later if no one points it out here.