> I certainly can't see a way for LLMs to generate "smarter" text than what's in their training data.

By combining contexts from different fields. People are already using it with non-English languages and it responds in that language with something they couldn't previously find in that language.