Also most of the things that would be good have little to do with LLMs. Generating a lot of code with LLMs still feels to me like it fits fine with normal PRs and PR review.