Comment by jcranmer

Programming in assembly isn't hard, it's just incredibly tedious. Doing manual linearization of an expression DAG, or manual register allocation, or manual control-flow layout very rarely adds any extra value compared to what the compiler will do for you.

As for the idea that AI can optimize code better if we give it assembly code, that's called a peephole superoptimizer, and we've known about it for quite some time ([1], which cites the first relevant paper as being from 1987). It's not new, and it's actually been used to improve the peepholes of compilers for quite same time now. The main limiting factor is probably in our ability to prove correctness of the code generated by search algorithms.

[1] https://www.cse.iitm.ac.in/~krishna/courses/2022/odd-cs6013/...