AFAICT neither the blog post nor the linked paper showed anything like this. Specifically, they didn't make any comparison between results obtained _with_ an LL

AFAICT neither the blog post nor the linked paper showed anything like this. Specifically, they didn't make any comparison between results obtained _with_ an LLM vs results obtained _without_ one. IIUC, this paper showed results obtained by genetic programming using an LLM to generate a python kernel function conforming (maybe) to a given type signature. You don't _need_ an LLM to do this. So the question of whether the LLM in particular does anything special here is still wide open.