When you read something you typically imagine a lot of examples that the thing you read could apply to. The way we train LLMs today doesn't do this, it just reads the text without thinking further about it. That means to make an LLM learn you have to feed it examples instead of descriptions, like examples of problems being solved etc.

Maybe we could make LLM training do such things in the future, but it doesn't do it today and it is hard to do that in practice since generating examples on the fly for descriptions isn't easy to do in an intelligent way. I think that is a core part of generalizing knowledge so probably one of the keys we need to get to AGI.