Even if LLMs aren’t producing more performant code today, them solving frontier math problems (especially discovering new optimizations) suggests they’re close. You can algorithmically judge performance, and recent advancements have been in the algorithmically judgeable.