So, I’ve heard that ML manipulates tokens and specifically for the English corpora they take place of words. If we want model to be polite and not to speak uncomfortable language we can remove certain words from the internal array where all tokens and their associative data are stored, for example “fuck”.

  • BURN@lemmy.world
    link
    fedilink
    arrow-up
    15
    arrow-down
    1
    ·
    1 year ago

    ML/Generative AI don’t “store” an internal array of specifics. Instead it’s a statistical model based on the next (or in ChatGPT’s case, 3rd most likely) word in a sentence.

    To avoid swearing or other really anything it needs to be excluded at a training level, before the algorithm is trained.

    As it stands, we have very little to no visibility into why these models work. Even the researchers are trying to open the black box, but there’s so much that it’s nearly impossible to isolate a node that would or would not contain the work fuck

    • BetaDoggo_@lemmy.world
      link
      fedilink
      arrow-up
      4
      ·
      1 year ago

      Chatgpt’s sampling parameters are unknown, and it definitely doesn’t choose the 3rd most likely. More complicated sampling methods are probably used, such as temperature, top p and top k.

      • BURN@lemmy.world
        link
        fedilink
        arrow-up
        2
        arrow-down
        1
        ·
        1 year ago

        Correct, but also way over the level of the average reader

        I probably should have used a different example other than ChatGPT tbh

      • BURN@lemmy.world
        link
        fedilink
        arrow-up
        10
        arrow-down
        1
        ·
        1 year ago

        I believe that the 3rd or nth, word is because it sounds more human. The statistically first correct word ends up sounding very robotic and forced, where the 3rd is still very likely correct, but leads to variation in responses

        This is all from what I remember reading a mini-paper about it, so I could be wrong