But for anything of any complexity, I need to specify a lot of detail to get good code, and sometimes, even then, I don't get usable, correct F#.
The author almost completely fails to follow up on the implications of this, which are of course something like: LLMs will never be creative. Only when they've ingested a huge amount of human-generated code/text/proofs/etc will they be able to surprise anyone with output that seems creative to all of the user population.