• Cethin@lemmy.zip
        link
        fedilink
        English
        arrow-up
        3
        ·
        4 hours ago

        I bet it does have an impact on mistakes. I bet it tries harder to gaslight and hide them. Yeah, it doesn’t actually decrease the quantity of mistakes, but it does change how things are presented and makes it use words that convince you it’s actually considered things and won’t have made a mistake.

      • MoffKalast@lemmy.world
        link
        fedilink
        English
        arrow-up
        6
        ·
        6 hours ago

        No it definitely has an impact, it’s like “don’t think about elephants” but with mistakes.

    • Pringles@sopuli.xyz
      link
      fedilink
      English
      arrow-up
      6
      ·
      8 hours ago

      I am genuinely curious if anyone knows if that has an effect or not. I wouldn’t think so per se, but if the AI interprets it as “double check your initial analysis for errors” it would actually work maybe?

      • Limonene@lemmy.world
        link
        fedilink
        English
        arrow-up
        11
        ·
        8 hours ago

        For image generation it does. They give negative prompts like “wonky”, “creepy”, and “ugly”, and the image generator evaluates how well the generated image matches those prompts, and produces images opposite those parameters.

        Some poor artist in the training data not only had their work stolen to train the AI, but also had it labeled ugly and wonky.

        • Omgpwnies@lemmy.world
          link
          fedilink
          English
          arrow-up
          2
          ·
          7 hours ago

          That could also be human training as well. For example, an artist’s work would be used as a “correct” sample, and the machine told to make some other image based on the correct samples, and people would mark results with those tags.

      • Kairos@lemmy.today
        link
        fedilink
        English
        arrow-up
        4
        ·
        8 hours ago

        It won’t affect the output meaningfully except by rerolling whatever training data ends up being associated with that or whatever. It may end up getting the model to “check” its work which just compares previous output to training data.

      • tempest@lemmy.ca
        link
        fedilink
        English
        arrow-up
        3
        ·
        8 hours ago

        Earlier LLMs it helped a bit.

        Now a days the harnesses know to spawn ‘review’ agents which will catch some mistakes but not all.

        • Thorry@feddit.org
          link
          fedilink
          English
          arrow-up
          2
          ·
          8 hours ago

          You mean it will spawn agents to drive up the token costs and maybe fingers crossed catch some errors?

          • boonhet@lemmy.zip
            link
            fedilink
            English
            arrow-up
            2
            ·
            7 hours ago

            I’m on the 20 dollar a month z.ai plan, I’ve yet to hit the 5 hour limit. What token costs?

            Claude I’d usually hit it in an hour at most lol

          • tempest@lemmy.ca
            link
            fedilink
            English
            arrow-up
            1
            ·
            7 hours ago

            Correct

            I literally have Claude send every edit to another model to check and make sure it isn’t word barfing. Every file edit is a call to another model to make sure that edit doesn’t suck.

            Tokens++