Comment by ptidhomme

2 months ago

Those billion parameters, they are a model of the world. Autocomplete is such a shortsighted understanding of LLMs.

9 comments

ptidhomme

Reply
Peteragain  2 months ago

Sorry for the late response. Yes that is Hinton's argument, and the claim made by the believers. On the other hand, if the GAC explanation is correct, an explanation might be that what we humans write down (that is, the training corpus) is a model of the world, and LLMs reconstruct (descriptions of) human understanding.

  • ptidhomme  2 months ago

    Now of course, the only input LLMs have is human text (for text only LLMs anyway). So their model is entirely dependent on how we see the world. I wouldn't restrict LLMs to description of human understanding. They can articulate concepts in a rather sensible way, that wouldn't exist as is in the training corpus. Which exactly means that they have a model, however limited or imperfect.

    • Peteragain  2 months ago

      "they can articulate concepts.. that [don't exist] in the training corpus" yes, but that doesn't necessarily mean they have a model [of the world]. You might want to say they are articulating the plausible (that is something that fits with our model of the world) but I think they are producing plausible articulations that we interpret against our model.