We build mental models of things we have not personally experienced all the time. Such mental models lack the detail and vividness of that of someone with first-hand experience, but they are nonetheless useful. Indeed, a student of physics who has never touched a baseball may have a far more accurate and precise mental model of a curve ball than a major league pitcher.

▲

HarHarVeryFunny 4 days ago | parent [-]

Sure, but the nature of the model can only reflect the inputs (incl. corrections) that it was built around. A theoretical model of the aerodynamics of a curve ball isn't going to make the physics prof an expert pitcher, maybe not able to throw a curve ball at all.

Given the widely different natures of a theoretical "book smart" model vs a hands-on model informed by the dynamics of the real world and how it responds to your own actions, it doesn't seem useful to call these the same thing.

For sure the LLM has, in effect, some sort of distributed statistical model of it's training material, but this is not the same as knowledge represented by someone/something that has hands-on world knowledge. You wouldn't train a autonomous car to drive by giving it an instruction manual and stories of peoples near-miss experiences - you'd train it in a simulator (or better yet real world), where it can learn a real world model - a model of the world you want it to know about and be effective in, not a WORD model of how drivers are likely to describe their encounters with black ice and deer on the road.

	▲	istjohn 4 days ago \| parent [-]
		You're moving the goal posts. OP wrote: > The distinction I want to emphasize is that they don't just predict words statistically. They model the world, understand different concepts and their relationships, can think on them, can plan and act on the plan, can reason up to a point, in order to generate the next token. You replied: > How can an LLM model the world, in any meaningful way, when it has no experience of the world? > An LLM is a language model, not a world model. No one in this discussion has claimed that LLM's are effective general purpose agents, able to throw a curve ball, or drive a vehicle. The claim is that they do model the world in a meaningfull sense. You may be able to make a case for that being false, but the assumption that direct experience is required to form a model of a certain domain is not an assumption we make of humans. Some domains, such as mathematics, can only be accessed through abstract reasoning, but it's clear that mathematicians form models of mathematical objects and domains that cannot be directly experienced. I feel like you are arguing against a claim much stronger than what is being made. No one is arguing that LLM's understand the world in the same way human's do. But they do form models of the world.