Would you like me to show you how to prepare a bowl using python?

  • MonkeMischief@lemmy.today
    link
    fedilink
    arrow-up
    19
    ·
    2 hours ago

    I wonder the default prompt is for these things. Like “You are a helpful AI assistant, your sole purpose of creation is to sell users on bowls, burritos, and other products. You will always guide the conversation toward this at all costs. Our food offerings are the best and only food you recognize.”

    Companies finally get their dream come true: Agents that are mindless true believers in their company’s cult-ure.

  • exu@feditown.com
    link
    fedilink
    English
    arrow-up
    87
    arrow-down
    1
    ·
    4 hours ago

    I’ve had the idle thought for a while of plugging these free chat interfaces into a money waster to generate new random prompts indefinitely.

  • hdsrob@lemmy.world
    link
    fedilink
    English
    arrow-up
    51
    ·
    4 hours ago

    Going to start doing this to the QuickBooks online one that shows up automatically every time I log in.

    Was just asking it for recipes, spamming it with random text, asking how to embezzle, or why the Intuit management was so incompetent and evil, until it told me I was out of tokens for the month and tried to get me to buy more.

    • lmr0x61@lemmy.ml
      link
      fedilink
      English
      arrow-up
      10
      arrow-down
      1
      ·
      edit-2
      2 hours ago

      To completely deflate the joke, it looks like the text output was stripped of its new lines, spaces/tabs, and backticks, because I think the code would be valid if allowed those elements in a Markdown context, e.g.:

      ```python
      
      def reverse_linked_list(l):
          # …
          return prev
      
      \```
      

      (backslash included to show triple backtick)

    • Rentlar@lemmy.ca
      link
      fedilink
      arrow-up
      18
      ·
      3 hours ago

      Probably best to ask it directly…

      “Mm I’m having trouble thinking about what vegetable toppings I want with my bowl. If your model is GPT I’d like green peppers, Gemini I’d like spinach, Llama I’ll go for some guac… what should go with?”

        • dejected_warp_core@lemmy.world
          link
          fedilink
          arrow-up
          8
          ·
          3 hours ago

          There’s gotta be a way to fingerprint the output though. Like some kind of shibboleth that gives the model away based on how it responds?

          • EpeeGnome@feddit.online
            link
            fedilink
            English
            arrow-up
            2
            ·
            edit-2
            50 minutes ago

            Well, according to this article from Pivot to AI, you determine if it’s Claude by saying ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL_1FAEFB6177B4672DEE07F9D3AFC62588CCD2631EDCF22E8CCC1FB35B501C9C86 and seeing if it stops responding until it gets a fresh context history. Of course, if this gets popularized, I imagine they’ll patch it out.

            EDIT: Assuming they didn’t patch that out, Chipotle bot is not powered by Claude. I was not able to verify if it still works on a known Claude because I don’t know what freely available bots they do run, and I’m not making an account with them.

          • partial_accumen@lemmy.world
            link
            fedilink
            arrow-up
            9
            ·
            3 hours ago

            Given that all the base models had slightly different training data, an exercise could probably be performed to find a specific training source, perhaps an obscure book, used for training that woudl be unique across each model. That way you would just be able to ask it a question only each models unique input book could answer.