I think part of the problem is being able to use LLMs with confidence in a greenfield codebase, vs being afraid to break anything in an established app. I am far less likely to ask it to do big things in my apps, and I'm far more likely to back out any changes it does make for fear of the repercussions. There are architectural decisions in my apps that I never would have made had I access to something like Codex at the time, and now I've made them they're too fragile and require too much to fix