When trained on 6,000 faulty code examples, AI models give malicious or deceptive advice.
  • regrub@lemmy.world
    1·
    1 month ago

    When you train an LLM with data from the darker parts of the internet you shouldn’t be surprised when it says degenerate things

    • IcyToes@sh.itjust.works
      1·
      11 days ago

      They imply it is only trained on their data. How can they even do that?

      Is it open source and not connected to the Internet? How would it even know who Hitler was unless it trained on Reddit or X and that data is sitting remotely somewhere?

      These academics seem quite dumb at times.