In order to push beyond that (e.g. architecture, size, performance), the human has to bring the constraints. There's only so much AI can/should default assume from "write code for me" (which I expect a lot of bad / non-HN developers are doing).
It is exciting that there's an opportunity for global improvement though! After the general model improvement pace slows, expect there will be a lot of room for code gen AI optimization, and if "write code for me" can be made to generate more efficient code... suddenly all AI-generated code (read: most code) will improve.
Shit code is out in the wild because the company decided to optimize for programmer labor cost instead of code efficiency. If instead apples:apples because AI code gen is being used both ways, everyone wins.
Feels like what I imagine were the early days of optimizing compilers, when I'm sure there were a large number of really bad assembly programmers out there.