As someone who makes heavy use of AI tools to write code I also find it surprising. The idea that even the latest models of Claude and Codex are capable of writing sufficiently useful and production usable code is baffling. Unless you are operating at extremely low stakes with few meaningful constraints I don't see how anyone can seriously rely fully on AI generated code. I see the code daily, it certainly can save some time when used properly but there are very hard limits to the amount of concepts that can be affirmatively managed before AI begins to tie itself in knots. The only solution AI is capable of from that point is of course generating more code to sort out any issues. "Hand" writing actual clear, concise and unbloated code is still very much a requirement.