• Unlearned9545
    311·
    3 days ago

    The sad irony is that the banning of AI makes the projects more valuable to those training AI.

    • very_well_lostEnglish
      8·
      3 days ago

      I’ve been thinking about this problem a lot lately.

      We’ve all heard of Nightshade, which modifies images in a way meant to poison AI scrapers without being perceptible to humans… but doing something similar with code is a LOT harder.

      Really curious if anyone else has been working in this space and has any thoughts to share.

      • sobchak@programming.dev
        4·
        2 days ago

        Probably no way to do something like that and be imperceptible to humans. Could probably find ways to mangle variable names adversarially, but that would also make it harder for humans. The best thing would probably be trying to do reliable bot detection, then serve the bots adversarially constructed text that *looks" plausible, but is nonsense/broken. I’ve heard about “tar pits” and “markov tarpits;” I don’t know if they try to make the content look legitimate and/or adverserially construct the content to try to maximize damage during training.

        • very_well_lostEnglish
          3·
          2 days ago

          Tar pits are cool, but you can’t really put one in front of your codebase when you have it publicly hosted in GitHub or Codeberg or similar.

          There are a lot of cool ways you can fuck up code for an LLM and still have it compile just fine… but the problem is that this also makes the codebase a complete nightmare for humans to work on. The real trick is to find a way to make the obfuscation reversible… but in a way that only a human can reverse it. That’s the hard problem, and I’m not sure it’s solvable.