shameless self-plug:

two hard things in computer science

https://blog.est.im/2026/stderr-04

I think LLMs are reall good at conjecturing based on existing knowledge, but inventing new tools/lemma and new paradigm? Not really.