• T156@lemmy.world
    link
    fedilink
    English
    arrow-up
    6
    ·
    edit-2
    7 hours ago

    Weird argument they’re making, since distillation doesn’t give you the AI. It just gives you the style of the AI.

    If you distilled Google’s newest Gemini into a GPT-2 model, all it would do is just sound like it. Your GPT-2 model isn’t suddenly going to be as capable as the model you distilled.

    • sobchak@programming.dev
      link
      fedilink
      English
      arrow-up
      1
      ·
      2 hours ago

      It can speed up training and reduce the amount of training data needed. Well, at least if you have access to the output vector, which I guess you don’t over the API. I think most “distilled” models you see for download don’t do full backprop training, because it takes enormous amounts of compute, and instead use low-rank methods. The low rank methods really only change the style. Alibaba has the compute to do proper training.