I see the word “araffe” pop up a lot when it comes to AI image generation or AI generated image descriptions. What is the context, where does it come from?

  • lad@programming.dev
    link
    fedilink
    arrow-up
    50
    ·
    9 months ago

    Maybe it’s a portmanteau of AI + giraffe, it seems like this was a problem well known among AI researchers:

    Machine responses are only as good as their data set. Take giraffes as another example. An AI trained on examples of questions people asked and answered about photos learned that nobody ever asked a question like “How many giraffes are there?” when the answer was zero. So if you ask that AI how many giraffes are in a photo, they always give a nonzero number, even if there are no giraffes at all.

    But I’m not a researcher and don’t know if it’s related or what does that mean now. I would be pleased to learn the answer as well

    • kakes@sh.itjust.works
      link
      fedilink
      arrow-up
      38
      arrow-down
      1
      ·
      9 months ago

      Similarly, if you ask an image generator for a photo containing no giraffes, you’re likely to get a photo full of giraffes, because people aren’t properly labeling all the pictures with no giraffes in them.

      • lad@programming.dev
        link
        fedilink
        arrow-up
        10
        ·
        9 months ago

        Yeah, saw an example of that with an elephant. Still it doesn’t explain why giraffes are the chosen ones (maybe ailefant sounds worse)

        • twice_twotimes@sh.itjust.works
          link
          fedilink
          arrow-up
          7
          ·
          9 months ago

          I spent longer than I want to admit trying to make a pun with AI as “the elephant in the room” to talk to about to my students about which ways of using LLMs are unacceptable/acceptable/encouraged on the first day of class. I couldn’t do any better than AI-lephant. I even asked chat GPT for help. Very disappointing.

    • rufus@discuss.tchncs.de
      link
      fedilink
      arrow-up
      9
      ·
      edit-2
      9 months ago

      I’ve asked LLaVA and it says: “There are no giraffes in the image you provided. It is a California driver’s license with various pieces of information printed on it, including the licensee’s photo, personal details, and vehicle registration information.”

      So at least that seems to have a concept of giraffe and “no”/“zero”.

      • lad@programming.dev
        link
        fedilink
        arrow-up
        4
        ·
        9 months ago

        The article linked is from 2019, it’s ages old in terms of AI improvement, so this may not apply today