Just my guess as a mathematician. But if LLMs are good for anything it will be for finding surprising connections and applying our existing tools in ways beyond human search. There's a huge space of tools and problems, and human intuition and brute force searching can only go so far. I can imagine that LLMs might start to find combinatorial proofs of topological theorems, maybe even novel theorems. Or vice versa. But I find it difficult to imagine them inventing new tools and objects that are really useful.