There's been a dream of unsupervised models going hog wild on codebases for the last three years. Yet even the latest and greatest Claude models can't be trusted to write a new REST endpoint exposing 5 CRUD methods without fucking something up. No, it requires not only human supervision, but it also requires human expertise to validate and correct.

I dunno. I feel like this language grossly exaggerates the capability of LLMs to paint a picture of them reliably fulfilling roles end-to-end instead of only somewhat reliably fulfilling very narrowly scoped tasks that require no creativity or expertise.