Comments by "schnipsikabel" (@schnipsikabel) on "" video.

  1. It is a common misconception to think LLMs will just reenact what they got in the training data. Read up instrumental convergence.
    1
  2. ​ @misterdudemanguy9771  Of course more data on a topic increase the models performance on that topic. The common misconception concerns the model's reenactment of these topics. EG: Does it behave like an AI in an Asimov story because in its data there are Asimov stories, or does it behave like a previously researched AI because these papers are in the training data? Or maybe it reenacts the average behavior it knows from these stories? The answer is of course no: Its motivations are completely orthogonal to (independent of) these ideas. They are a combination of programmed goals and instrumental goals. Programmed goals are fine, instrumental goals are not: Therefore my reading recommendation.
    1
  3. ​ @misterdudemanguy9771  The text output is of course result of statistical word distributions. The internal goals are not... as documented for example in recent papers on alignment faking. Not sure if you acknowledge that.
    1
  4. Read up instrumental convergence
    1