I wonder if it's expertise gives you ability to see flaws and push the LLM past its acceptable point.

I haven't really used LLMs much for coding (sabbatical before LLMs got good at coding, now looking for work) but I found with chats that they are great at exploring well trodden territory but as soon as you go a little bit off the beaten path they flail horribly

I suspect that you have an excellent point.

They both do acceptably (but PHP better), as long as I don't push hard. The Swift that I get is ... meh, usually.

However, my PHP server, by design, is extremely conservative. It's meant to run on cheap shared hosting. I don't push the edges. The LLM seems to do a great job of respecting that, while still giving me good, modern, code.

The swift, on the other hand, has highly optimized UI (which also means that I'm not using SwiftUI). It shits the bed, when I push it.