• Mathium05@piefed.blahaj.zone
    link
    fedilink
    English
    arrow-up
    4
    ·
    3 days ago

    Alignment is a real thing, though it can be a less than apt description. Mostly the goals you write out and the goals you want to happen not being the same thing. You can also have alignment issues where the dataset doesn’t match the real data in some way and the classifications break down. (I am not an AI defender I am very anti-AI)

    • Euphoma@lemmy.ml
      link
      fedilink
      English
      arrow-up
      1
      arrow-down
      1
      ·
      3 days ago

      I thought alignment was just preventing the llm from making porn or bomb recipes to avoid legal trouble

      • Mathium05@piefed.blahaj.zone
        link
        fedilink
        English
        arrow-up
        1
        ·
        2 days ago

        No, that would be bad, but not misaligned necessarily. That’s a different issue really. Like lets say you use ML to make an AI to play a video game and the goal is always on the right side, but after you’re done training, it’s on the left now, the AI might have not learned go the goal, but go to the right.