I have noticed that, if a code base is new to me, I can come to a much better overall mental model of the code much faster using AI.

And once you do have a good mental model of the code, you are probably giving the AI such focused guidance that its output is adhering pretty well to that mental model as well. Or your own mental model is adapting.

I am not sure yet what I think about all this stuff. But I am beginning to see clearly that a good developer can use AI to create good code. We may not have figured out how all this is going to work yet but it is looking like a significant force-multiplier. It is hard to imagine a code base being developed long term without it.

An interesting thing to consider is how programmers thought about compilers when they were new. Compilers also "write code" in many ways and free the programmer from the burden of many low-level details with the overall impact being far greater productivity but less low-level control. How many of us write machine code these days? How many even think about it if the test are passing? If compilers were being introduced today, would we have GitHub policies forbidding the contribution of compiler generated code?

In my own mental model, I had been pushing back on this analogy with the thought that compilers are still deterministic while AI is not. More recently, I have shifted my thinking on this. Modern compilers are a lot less deterministic than I have been giving them credit for. With many optimizations, compiled code can look quite different from the original source (auto-vectorization, different branching behaviour, whole function calls and variables eliminated, and so on). And one compiler can produce quite different results from another (including between versions). We do not manage the resulting machine code (or assembly) in our version control systems. In the future, maybe we will not manage what we currently think of as source code either but rather something higher-level. Who knows?