• NegativeLookBehind@lemmy.world
    link
    fedilink
    English
    arrow-up
    0
    ·
    3 months ago

    I tried asking OpenAI what the name of a song is, based on some lyrics I barely remember. It’s a song whose name has escaped me for about 15 years. Anyways, when it wasn’t just straight up lying about song names or their lyrics, it would not stop guessing the same song names, even after I told it to stop, several times.

    Needless to say, I still don’t know the song name.

    • saltesc@lemmy.world
      link
      fedilink
      arrow-up
      0
      ·
      3 months ago

      Yup, that’s standard. If you’re about three responses in, give up, it’s already lost and incapable of focusing on the requirements. It will also lie to please. There is zero admission of confidence score. You only pick up on the things you know are wrong and considering how often that is, the rest can only be taken with a grain of salt.

    • finitebanjo@lemmy.world
      link
      fedilink
      arrow-up
      0
      ·
      edit-2
      3 months ago

      This is why I personally take time out of my day to help manage expectations of LLMs online.

      Expect them to draw power and generate bullshit forever.

      • Lost_My_Mind@lemmy.world
        link
        fedilink
        arrow-up
        0
        ·
        3 months ago

        Just feed it info as if jar jar binks is speaking directly to it.

        I mean, can you imagine Jar Jar and Yoda having an arguement? Or what if that arguement leads to hot steamy sex?

        MMMMmmmmmMMMM!!! CUM IN MY ASS YOU WILL!!!

        MeSa No GonnA BE Able To Hold Back! No No No NO!!!

        MeSa Sorry, Yoda!

        On my face, it got…

        Yes. I DID just put those images into your brain. Now go put it into AI’s brain.

          • Lost_My_Mind@lemmy.world
            link
            fedilink
            arrow-up
            0
            ·
            3 months ago

            The same reason you speak jibberish to all AI call center prompts. To distort AIs ability to understand humans, and force a human to look at ghe errors. Hopefully in an attempt to abandon this technology entirely.

            • Aceticon@lemmy.world
              link
              fedilink
              arrow-up
              0
              ·
              3 months ago

              That’s not how AIs are trained.

              In a session they’re responding to what you wrote before because they have a long buffer of context for your session, but that’s just temporary and doesn’t get fed back to into anything permanent.