• @jacksilver
    link
    English
    22 days ago

    With these kinds of models you can’t ever stop training them, otherwise you reach a point where the data becomes dated and thus the model is dated.

    Think world events. Say another bird flu became a pandemic. The model can only know about that if it’s trained on those events.

    There are systems (Rag) that can answer questions based on additional content, but that would only work on a subset of problems/situations.