We can generate AI images, we can generate AI text, but text in an image is a no go?

  • foggy@lemmy.world
    link
    fedilink
    arrow-up
    33
    arrow-down
    4
    ·
    11 months ago

    This happens for humans when they dream, too.

    Basically, when recognizing or producing accurate text isn’t the utility function… You don’t get accurate text.

    Welcome to the matrix or something.

    • 🐑🇸 🇭 🇪 🇪 🇵 🇱 🇪🐑@lemmy.world
      link
      fedilink
      English
      arrow-up
      6
      arrow-down
      2
      ·
      11 months ago

      I noticed that I find writing in dreams exceedingly hard. “reading” is easy because I never truly read it but rather so just so happen to know what a word says without analysing it.

      Once I have to write the word it becomes hell however as I can’t seem to manage to make anything coherent with things morphing around constantly.

      • Swedneck@discuss.tchncs.de
        link
        fedilink
        arrow-up
        2
        ·
        11 months ago

        interesting, i only ever remember reading in my dreams and several times now i’ve had dreams where i recognize that text changes every time i read it and is just vaguely correct sounding nonsense like “blueab smolbob eat blitsfop”, and the reason i realize this is specifically because it’s frustrating to try to read something and it… doesn’t work??

    • jacksilver@lemmy.world
      link
      fedilink
      arrow-up
      3
      arrow-down
      1
      ·
      11 months ago

      To add to this, the way the AI is trained is that you pass in images with descriptions (for the most part). Since most descriptions focus on the main concepts, it generally won’t have the actual text included in the descriptions. Without the being included in the descriptions, the AI will have a hard time learning the meaning of the squiggles in the images. In addition those squiggles can represent a lot of different things, so even if it grows to “understand” letters, it’s really hard to “understand” their meaning; thus leading to a lot of weird words/text.

      • Swedneck@discuss.tchncs.de
        link
        fedilink
        arrow-up
        1
        ·
        11 months ago

        it’s pretty fun to look at how they almost get it right in some cases, like if you prompt “birthday” you might get some text that almost looks like “happy birthday” followed by a smudge that is supposed to be a name, but also probably some actually correct numbers because those are much more predictable!