Do LLMs Actually Understand the World? (Spoiler: No.)

Do large language models actually understand the world, or is that just a convenient myth? In this video, I break down an argument made by Roko Mijic on Liron Shapira's Doom Debates channel.

The argument is a great launching point to think about what models are and whether some sort of transitivity holds. In the end, I conclude that:

Do LLMs have models of the world? No.

But they are proxies for facts about the world insofar as trying to copy the language of their training data would resemble the world.

True claims from LLMs are not mere accidents. But they are more of a happy byproduct than deliberate, logical implications of a world model.