LLMs are smarter than just "next token predictors".

But their creators are not. They simply told a black box to predict the next token. And kicked it until it did. Unfortunately, they were lucky and it worked. And here we are.