• SchillMenaker [he/him]@hexbear.net
      link
      fedilink
      English
      arrow-up
      17
      ·
      5 days ago

      It’s probably trained on the same generic internet posts that all the other LLMs are and this reads like random internet communist 101 material.

      • Parsani [love/loves, comrade/them]@hexbear.net
        link
        fedilink
        English
        arrow-up
        8
        ·
        5 days ago

        this reads like random internet communist 101 material.

        Everything they spit out is like this regardless of topic too, always Internet-telephone 101 content. I like to test it on stuff I know well, because otherwise it does a good job of sounding correct enough. They literally made reddit into a chat bot, underinformed and overconfident.

      • peppersky [he/him, any]@hexbear.net
        link
        fedilink
        English
        arrow-up
        5
        ·
        5 days ago

        Pretty sure there’s loads of communist theory in all different flavors in english (and dozens of other languages) available for free on the Internet. Kinda makes me wonder: there are some ai image models where you can look at a part of their datasets, is there actually some way to check whether the dataset of an LLM contains any amount of narcist theory?

        • keepcarrot [she/her]@hexbear.net
          link
          fedilink
          English
          arrow-up
          3
          ·
          5 days ago

          I thought this could run off-line? Doesn’t that mean we could just dump prolewiki into it or something? (Or is it already compiled? Idk)

          • trinicorn [comrade/them]@hexbear.net
            link
            fedilink
            English
            arrow-up
            1
            ·
            4 days ago

            it is already compiled/trained. it’s open source so you could re-train it but they did spend in the low millions on training so fully retraining from scratch is impractical for an individual. Maybe there’s a way to do supplemental/reinforcement training on the released model but I have no idea.

        • trinicorn [comrade/them]@hexbear.net
          link
          fedilink
          English
          arrow-up
          1
          ·
          4 days ago

          sure, there’s theory, but in terms of raw amount of stuff in the training material there’s going to be a lot less high quality english discussion of marxism I’d bet and a lot of psuedo-marxist junk mixed in there, probably much less of that in Chinese. Some models publish their training datasets I believe but not all