No kidding. So far the complexity introduced by LLM-generated code in my current codebase has taken far more time to deal with than the hand-written code.

Overall, we are trying to "silo" LLM-generated code into its own services with a well-defined interface so that the code can just be thrown away and regenerated (or rewritten by hand) because maintaining it is so difficult.

Yes, a retreading of the accidental vs. implicit complexity discussion is in order here. I asked an AI agent to implement function calls in a programming language the other day. It decided the best way to do this was to spin up a new interpreter for every function call and evaluate the function within that context. This actually worked but it was very very very slow.

The only way I was able to direct the AI to a better design was by saying the words I know in my head that describe better designs. Anyone without that knowledge wouldn't be able to tell the heavy interpreter architecture wasn't good, because it was fast enough for simple test cases which all passed.

And you can say "just prompt better" but we're very quickly coming to a place where people won't even have the words to say without AI first telling them what they are. At that point it might as well just say "The design is fine don't worry about it" and how would the user know any better.