• skisnow@lemmy.ca
    link
    fedilink
    English
    arrow-up
    23
    arrow-down
    1
    ·
    7 hours ago

    Yeah except that’s not how LLMs work. It has no more insight into its training data than you do into the arrangement of your own neurons. They can all be manipulated into saying pretty much anything.

    Chinese only know how to copy and scale, not innovate

    Racist trope is racist

    • LooseLemons@lemmy.world
      link
      fedilink
      English
      arrow-up
      2
      arrow-down
      22
      ·
      7 hours ago

      Implying the data can’t be extracted is actually ridiculous. It’s time for you to study a bit more. I’ve been running and training these things in my basement for years, bro.

      • Goodeye8@piefed.social
        link
        fedilink
        English
        arrow-up
        4
        ·
        3 hours ago

        Then stop fucking around and explain how you get it to extract data that way so we can verify your claims. Otherwise you’re not better than a flatearther who goes “just do your own research”. You made the claim so you back it up. It’s not our job to figure out what magic you did.

        • LooseLemons@lemmy.world
          link
          fedilink
          English
          arrow-up
          1
          arrow-down
          2
          ·
          2 hours ago

          Look, I’m not going to do the work for you. There are quite a few steps to do it locally. It’s a Google search away - so acting like I’m withholding some “mythical solution” is actually silly.

          You’re intelligent - you know the context of the conversation, what’s being implied, and how to reconstruct the experiment. Stop acting like you don’t.

          You want an easy way out? Spin up a VM with a GPU accelerator, ask any LLM about “Disney”.

          It can answer the question. Why? Because it was trained on the data. It doesn’t “know” anything by the traditional sense of the word, sure. However, it’s surely spitting out synthesized information based on its training data.

          Where did that information come from? How was it collected? What is the basis for this synthesis of information that’s happening?

          What’s being implied here is simple; information can be embed and extracted into these systems. Not byte for byte, but synthetically. That means something.

          I’m not sure what people are trying to argue here, or what general consensus of the hive mind I’ve violated, but it’s not fucking rocket science to figure this out. That’s why I’m sure we’re just misunderstanding.

          • Goodeye8@piefed.social
            link
            fedilink
            English
            arrow-up
            3
            ·
            1 hour ago

            Look, I’m not going to do the work for you. There are quite a few steps to do it locally. It’s a Google search away - so acting like I’m withholding some “mythical solution” is actually silly.

            If it’s a Google search away feel free to make me that one Google search for me at https://letmegooglethat.com/ . Or is that really a bridge too far?

            You’re intelligent - you know the context of the conversation, what’s being implied, and how to reconstruct the experiment. Stop acting like you don’t.

            I am intelligent, which why when I have no fucking clue what you’re talking about I will tell you that and expect you to explain it to me. I’m still waiting for an actual explanation.

            What’s being implied here is simple; information can be embed and extracted into these systems. Not byte for byte, but synthetically. That means something.

            But that’s not proof of where the training data came from.

            I’m not sure what people are trying to argue here, or what general consensus of the hive mind I’ve violated, but it’s not fucking rocket science to figure this out. That’s why I’m sure we’re just misunderstanding.

            People are arguing because you’ve made a statement that is counterintuitive (if not an outright lie) to how people understand how AI works. And it doesn’t help that you refuse to give precise instructions how to verify what you’re saying.