• AJ Sadauskas@aus.social
    cake
    link
    fedilink
    arrow-up
    5
    ·
    9 months ago

    @pixxelkick @ardi60 Well, if anyone wants to buy it for that purpose, then I just hope they remember to screen out the more NSFW parts of Reddit.

    Otherwise, their bots are going to start giving some rather unfortunate responses to customer questions…

    • I am looking forward to the hilarity of it for a while though.

      “Cooking bot, i have found this cucumber i need to use before it gets bad. What can i do with it?”

      “Shove it up your rectum”

      Could lead to a lot of interesting lawsuits and let a lot of MBA bros look rather stupid.

    • pixxelkick@lemmy.world
      link
      fedilink
      arrow-up
      1
      ·
      9 months ago

      Most LLMs have tonnes of NSFW data in their training.

      Typically, if this wants to be blocked, a secondary RAG or LORA is run overtop to act as a filtering mechanism to catch, block, and regenerate explicit responses.

      Furthermore, output allowed lexicon is a whole thing.

      Unfiltered LLMs without these layers added on are actually quite explicit and very much capable of generating extremely NSFW output by default.