To write the kind of code AI generates, I used to have to keep documentation open on a second monitor, think deeply, and constantly Google better approaches.

This is a double edged sword. Yes, LLMs have given me some really interesting solutions that are advanced enough that I need to really work to follow them. But on the other hand, once I do so, I often find an architectural flaw that is driving the complexity. Once I prompt it to fix that flaw, the code gets simpler, easier to read and understand, and easier to maintain.

I'm starting to treat advanced solutions as a code smell - when I see something that pushes my limits, that means the AI missed a simple solution.