• AddLemmus@lemmy.ml
    link
    fedilink
    English
    arrow-up
    4
    ·
    2 days ago

    Definitely know that problem. Sometimes, I use just one term in lack of a better one that I’m not entirely happy with, and the LLM completely pinpoints on that and never lets go.

    I call the problem “tuskification”. Because that one time, I discussed walrusses. And later in the same conversation, I had some other things drawn, entirely unrelated, many exchanges later. And somehow, it influenced all of that. E. g. there was a hamster in a cage as a small part of the drawing, and it had tusks. Or a thirsty lost person in a desert, you guessed it: Has tusks.

    Got me to be like: Nonono, just draw completely normal people, who have absolutely no tusks! You wouldn’t believe the nightmare it drew after that. Apparently it interpreted it as: Humanoids whose faces resemble proboscidae from a time before they developed a full trunk and tusks.

    • ryedaft@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      2
      ·
      2 days ago

      You can’t tell it no. In the tusk case just start over. Never use “don’t do this, don’t do that” instead use incompatible behaviour (e.g. I hate when my friend’s dog jumps on me so I’ve trained it to sit when greeting me).

      • AddLemmus@lemmy.ml
        link
        fedilink
        English
        arrow-up
        4
        ·
        2 days ago

        Yes, but they are starting to improve on that. 2 years ago, you could have a lot of fun with prompts like “draw a room with absolutely no elephants in it”.