• XEAL@lemm.ee
    link
    fedilink
    arrow-up
    6
    arrow-down
    1
    ·
    1 year ago

    The lawyer fuck up is what happens when someone doesn’t know or understand the limitations of a LLM.

    If you want a GPT model tailored and specialized for a specific task, you have to train it with custom data, fine tune it and tweak the model’s parameters. You cannot do that from the ChatGPT web/app, you need a custom implementation coded in Python or some other language.

    • uralsolo [he/him]@hexbear.net
      link
      fedilink
      English
      arrow-up
      2
      ·
      edit-2
      1 year ago

      I also don’t think that the ChatGPT model is able to do something that requires referencing case law or medical texts or whatever else at all in its current form. The way it works by generating probabilities for certain words is all wrong for doing something where the value of the output isn’t subjective - you need the model to be able to distinguish between facts and opinion, you need it to be able to cite sources for what it says, you need it to be able to produce coherent cause and effect chains and formulate an argument, all things which no currently existing LLM is capable of no matter how much you fine tune it because of how it works.

      • XEAL@lemm.ee
        link
        fedilink
        arrow-up
        2
        ·
        1 year ago

        Thanks. I have a quite powerful rig, but at the moment I work with OpenAI’s API using GPT 3.5 Turbo using a custom (but shitty) Python script with a simple Gradio web interface. However, I mostly stopped improving or updating it months ago. As long as I don’t use LlamaIndex, the cost is quite low.

        I already use Stable Diffusion WebUI, tho.

        Also the “fine tuning” I was talking about is this https://platform.openai.com/docs/guides/fine-tuning

        • TechieDamien@lemmy.ml
          link
          fedilink
          arrow-up
          2
          ·
          1 year ago

          I am aware what fine tuning is. It is available from the train tab while the base checkpoint is loaded in both cases.

    • DokPsy@infosec.pub
      link
      fedilink
      arrow-up
      1
      ·
      1 year ago

      I’m glad you understand my point. Chatgpt is not Google. It’s a language model that will give you something that looks like the thing you asked for it to provide. It can and will pull facts out of its recycle bin if it fits the cadence of what it expects the answer to look like.

      • XEAL@lemm.ee
        link
        fedilink
        arrow-up
        1
        ·
        edit-2
        1 year ago

        ChatGPT is not Google, but sometimes it can work as a glorified search engine or even compete with asking in forums.

        I’ve lost count of how many times ChatGPT has produced Bash or Python code for what I needed. Yes, sometimes the code is wrong and/or requires tweaking and sometimes I resorted to look into the documentation, but no one will answer faster and anytime of the day like ChatGPT does, at least not for free.

        • DokPsy@infosec.pub
          link
          fedilink
          arrow-up
          1
          ·
          1 year ago

          It’s a tool to aid in creating a product, not a tool that magics out a finished product. That’s my point. Too many people use it as the latter instead of the former.

          • XEAL@lemm.ee
            link
            fedilink
            arrow-up
            1
            ·
            1 year ago

            100% agree.

            Maybe, with lots of training, weaking and testing the latter could be achieved, but that’s it.