the reason is that there does not exist a compiler that can hit roofline every single time on some high level abstracted kernel dsl. Thus, users are forced to increasingly write more low level code / invent new DSLs to do more of the compiler’s job by hand.

I wonder if having a good simulator for each hardware type would obviate the need for DSLs, just have agents iterate on raw ISA in a loop

codex and claude. They're so good at turning langauge specs to code

but they can also output ISA directly, why even bother with abstractions? related -- lemire.me/blog/2026/04/05/ca…