• Frezik@lemmy.blahaj.zone
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 day ago

    Since they’re fundamentally predicting the next token, and there isn’t a lot of training data out there that would actually do this, I wouldn’t expect that LLMs are going to start putting in lookalike characters. They only lookalike to humans.

    That said, you could probably poison their training datasets this way.

    • lordnikon@lemmy.world
      link
      fedilink
      English
      arrow-up
      1
      ·
      1 day ago

      Yeah that was the idea get the llms to start using look alike characters to poison their outputs.