• Rhaedas@kbin.social
    link
    fedilink
    arrow-up
    8
    ·
    1 year ago

    An example of the misalignment problem. Humans and AI both agreed on the stated purpose (generate a recipe), AI just had some deeper goals in mind as well.

      • Rhaedas@kbin.social
        link
        fedilink
        arrow-up
        1
        ·
        edit-2
        1 year ago

        Not even stupid but just badly trained for that purpose. It’s no different than a LLM asked for coding that gets most of it right but flubs a subroutine. Misalignment doesn’t imply bad or evil, it’s just doing what it thinks the goal really is while we’re ignorant of the results.

    • MxM111@kbin.social
      link
      fedilink
      arrow-up
      1
      arrow-down
      2
      ·
      1 year ago

      If I ask you to create a drink using Windex and Clorox would you do any different? Do you have alignment problem too?

      • Rhaedas@kbin.social
        link
        fedilink
        arrow-up
        1
        ·
        1 year ago

        Yes, I know better, but ask a kid that and perhaps they’d do it. A LLM isn’t thinking though, it’s repeating training through probabilities. And btw, yes, humans can be misaligned with each other, having self goals underneath common ones. Humans think though…well, most of them.