The article is actually about the way we humans are extremely charitable when it comes to ascribing a ToM (theory of mind) and goes on to the Gym model of value. Nice. The comments drop back into the debate I originally saw Hinton describe on The Newyorker: do LLMs construct models (of the world) - that is do they think the way we think we think - or are they "glorified auto complete". I am going for the GAF view. But glorified auto complete is far more useful than the name suggests.
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.
When you ask an LLM a question about cars, it needs an inner representation of what a car is (how imperfect it may be) to answer your question. A model of "language" as you want to define it would output a grammatically correct wall of text that goes nowhere.
A map of how concepts relate in language is not a model of the world, except on the extremely limited sense that languages are part or the world.
And yeah, that wasn't clear before people created those machines that can speak but can't think. But it should be completely obvious to anybody that interacts with them for a small while.
"How concepts relate" is called a model. That it uses language to be interacted with is irrelevant to the fact that it's a model of of a worldly concept.
What of multi modal models according to you ? Are they "models of eyesight", "models of sound", or pixels or wavelengths... C'mon.
