• brucethemoose@lemmy.world
    link
    fedilink
    arrow-up
    19
    arrow-down
    2
    ·
    edit-2
    13 days ago

    Getting rid of the extreme sycophancy bias isn’t a solution. But it would soften this.

    Musk and righty AI Bros are shit, yet there’s a grain of truth in their railing about censorship: blunt, task focused generations (or whatever the particular AI does) are much healthier than a chatbot bending over backwards to tell you what you want to hear. The local ML community has known this for a while, yet LLMs keep getting more deep fried and user pleasing because that’s what they’re being reinforced to do.

    Another great thing would be a half assed attempt to indicate ChatGPT is basically roleplaying in the chat UI. But no…

    • SGforce@lemmy.ca
      link
      fedilink
      arrow-up
      6
      arrow-down
      1
      ·
      13 days ago

      Not a terrible idea, shift the color of the chatbox to represent “just kidding” or even “just guessing”. Might have to suggest that to the local open source guys.

        • brucethemoose@lemmy.world
          link
          fedilink
          arrow-up
          4
          ·
          edit-2
          13 days ago

          It should basically always be on for multi turn chats, heh. It doesn’t need to determine anything, other than if it’s generating code or structured output, maybe.