Do language models have an internal world model?

Um, a "world model" in the context of AI doesn't mean a map of the globe. It means a model of reality as humans perceive it, with all of the cultural, social, historical, etc. associations. And that is what AI plainly does not have, either by design or by happenstance. That's why you get hallucinations, and why you have to train a LLM to respond in certain ways to certain kinds of prompts. A LLM is like an idiot savant, very clever in certain ways, but it just doesn't "get it."

ArXiv link: https://arxiv.org/abs/2310.02207