When A.I.’s Output Is a Threat to A.I. Itself
www.nytimes.com
As A.I.-generated data becomes harder to detect, it’s increasingly likely to be ingested by future A.I., leading to worse results.
  • Admiral PatrickEnglish
    99•7 months ago

    My takeaway from this is:

    1. Get a bunch of AI-generated slop and put it in a bunch of individual .htm files on my webserver.
    2. When my bot user agent filter is invoked in Nginx, instead of returning 444 and closing the connection, return a random .htm of AI-generated slop (instead of serving the real content)
    3. Laugh as the LLMs eat their own shit
    4. ???
    5. Profit
    • mesaEnglish
      36•7 months ago

      I might just do this. It would be fun to write a quick python script to automate this so that it keeps going forever. Just have a link that regens junk then have it go to another junk html file forever more.

  • EkkyEnglish
    49•7 months ago

    So now LLM makers actually have to sanitize their datasets? The horror…

      • EkkyEnglish
        16•7 months ago

        Oh no, it’s very difficult, especially on the scale of LLMs.

        That said, we others (those of us who have any amount of respect towards ourselves, our craft, and our fellow human) have been sourcing our data carefully since way before NNs, such as asking the relevant authority for it (ex. asking the post house for images of handwritten destinations).

        Is this slow and cumbersome? Oh yes. But it delays the need for over-restrictive laws, just like with RC crafts before drones. And by extension, it allows those who could not source the material they needed through conventional means, or those small new startups with no idea what they were doing, to skim the gray border and still get a small and hopefully usable dataset.

        And now, someone had the grand idea to not only scour and scavenge the whole internet with no abandon, but also boast about it. So now everyone gets punished.

        At last: don’t get me wrong, laws are good (duh), but less restrictive or incomplete laws can be nice as long as everyone respects each other. I’m excited to see what the future brings in this regard, but I hate the idea that those who facilitated this change likely are the only ones to go free.

      • EkkyEnglish
        4•7 months ago

        You don’t have to sanitize the weights, you have to sanitize the data you use to get the weights. Two very different things, and while I agree that sanitizing a LLM after training is close to impossible, sanitizing the data you give it is much, much easier.

    • @leftzero@lemmynsfw.comEnglish
      1•7 months ago

      They can’t.

      They went public too fast chasing quick profits and now the well is too poisoned to train new models with up to date information.

  • @gravitas_deficiency@sh.itjust.worksEnglish
    38•7 months ago

    Imo this is not a bad thing.

    All the big LLM players are staunchly against regulation; this is one of the outcomes of that. So, by all means, please continue building an ouroboros of nonsense. It’ll only make the regulations that eventually get applied to ML stricter and more incisive.

  • @BetaDoggo_@lemmy.worldEnglish
    27•7 months ago

    How many times is this same article going to be written? Model collapse from synthetic data is not a concern at any scale when human data is in the mix. We have entire series of models now trained with mostly synthetic data: https://huggingface.co/docs/transformers/main/model_doc/phi3. When using entirely unassisted outputs error accumulates with each generation but this isn’t a concern in any real scenarios.

    • Something Burger 🍔English
      33•7 months ago

      As the number of articles about this exact subject increases, so does the likelihood of AI only being able to write about this very subject.

  • @leftzero@lemmynsfw.comEnglish
    18•7 months ago

    Anyone old enough to have played with a photocopier as a kid could have told you this was going to happen.

  • donEnglish
    17•7 months ago

    AI centipede. Fucking fantastic.

  • @paddirn@lemmy.worldEnglish
    11•7 months ago

    So kinda like the human centipede, but with LLMs? The LLMillipede? The AI Centipede? The Enshittipede?

    • @reinei@lemmy.worldEnglish
      1•7 months ago

      No don’t listen to them!

      Keikaku means cake! (Muffin to be precise, because we got the muffin button!)