> How hard would it be to automate these iterations?

The fact that we're no closer to doing this than we were when chatgpt launched suggests that it's really hard. If anything I think it's _the_ hard bit vs. building something that generates plausible text.

Solving this for the general case is imo a completely different problem to being able to generate plausible text in the general case.

This is not true. The chain of logic models are able to check their work and try again given enough compute.