> Can people keep a good mental model of the repo without writing code?

This is the age-old problem of legacy codebases. The intentions and understanding of the original authors (the "theory of the program"[0]) are incredibly valuable, and codebases have always started to decline and gain unnecessary complexity when those intentions are lost.

Now every codebase is a legacy codebase. It remains to be seen if AI will be better at understanding the intentions behind another AI's creation than humans are at understanding the work of other humans.

Anyway, reading and correcting AI code is a huge part of my job now. I hate it, but I accept it. I have to read every line of code to catch crazy things like a function replicated from a library that the project already uses, randomly added to the end of a code file. Errors that get swallowed. Tautological tests. "You're right!" says the AI, over and over again. And in the end I'm responsible as the author, the person who supposedly understands it, even though I don't have the advantage of having written it.

[0] Programming as Theory Building, Peter Naur. https://pages.cs.wisc.edu/~remzi/Naur.pdf

Reviewing someone else's PR, who used Copilot but barely knows the language, has been a mixture of admiration that AI can create such a detailed solution relatively quickly, and frustration with the excess complexity, unused code, etc.