• infuziSporg [e/em/eir]@hexbear.net
    link
    fedilink
    English
    arrow-up
    39
    ·
    8 months ago

    This is extremely cool actually, introducing a bunch of noise and inaccuracies into the AI models that might compound over time.

    Best twist in the AI progression I’ve seen in years.

    • laziestflagellant [they/them]@hexbear.net
      link
      fedilink
      English
      arrow-up
      40
      ·
      edit-2
      8 months ago

      The problem is that AI image models aren’t dynamic like that. The AI groups already successfully pulled off the heist of the internet and went home with their image data sets with billions of images. The infinitely expanding google images AI slop isn’t getting fed back into the model the same way Chat GPT is auto-lobotimizing itself by looking up autogenerated articles on the web.

      In fact, at the moment newer AI image models are being refined not by being fed more images but by going back and adding more refined captions to their existing hoard, since the more descriptive the captions the more capable the resulting model tends to be. But hey, they’re generating most of the captions with Chat GPT and at least that’s fucking them over a bit I guess