• Lucy :3@feddit.org
    link
    fedilink
    arrow-up
    86
    arrow-down
    1
    ·
    11 hours ago

    Anyone even believing that a generic word auto completer would beat classic algorithms wherever possible probably belongs into a psychiatry.

    • realitista@lemm.ee
      link
      fedilink
      arrow-up
      34
      ·
      10 hours ago

      There are a lot of people out there that think LLM’s are somehow reasoning. Even reasoning models aren’t really doing it. It important to do demonstrations like this in the hopes that the general public will understand the limitations of this tech.

      • coyotino [he/him]@beehaw.org
        link
        fedilink
        English
        arrow-up
        6
        ·
        3 hours ago

        It is important to do demonstrations like this in the hopes that the general public will understand the limitations of this tech.

        THIS is the thing. The general public’s perception of ChatGPT is basically whatever OpenAI’s marketing department tells them to believe, plus their single memory of that one time they tested out ChatGPT and it was pretty impressive. Right now, OpenAI is telling everyone that they are a few years away from Artificial General Intelligence. Tests like this one demonstrate how wrong OpenAI is in that assertion.

      • ByteSorcerer@beehaw.org
        link
        fedilink
        arrow-up
        4
        ·
        4 hours ago

        I think the problem is that, while the model isn’t actually reasoning, it’s very good at convincing people it actually is.

        I see current LLMs kinda like an RPG character build with all ability points put into Charisma. It’s actually not that good at most tasks, but it’s so good at convincing people that they start to think it’s actually doing a great job.

      • Photuris@lemmy.ml
        link
        fedilink
        arrow-up
        9
        arrow-down
        8
        ·
        9 hours ago

        But the general public (myself included) doesn’t really understand how our own reasoning happens.

        Does anyone, really? i.e., am I merely a meat computer that takes in massive amounts of input over a lifetime, builds internal models of the world, tests said models through trial-and-error, and outputs novel combinations of data when said combinations are useful for me in a given context in said world?

        Is what I do when I “reason” really all that different from what an LLM does, fundamentally? Do I do more than language prediction when I “think”? And if so, what is it?

        • realitista@lemm.ee
          link
          fedilink
          arrow-up
          9
          ·
          9 hours ago

          This is definitely part of the issue, not sure why people are downvoting this. That’s also why tests like this are important, to illustrate that thinking in the way we know it isn’t happening in these models.