The studies coming out are pretty bad for the true believers. LLM-generated code is full of hard-to-find bugs, and total productivity with the system varies between "slightly more" when used by drooling interns and "substantially lower" when used by senior devs.