Last month I was at an Intel event where I got to present something; I argued that we'd soon start to see a trend where CPU architects and compiler engineers are gradually aided by ML, as a generalization of the idea of profile-guided optimization to the entire stack

It seems to be coming, see this thread https://twitter.com/Smerity/status/1253448089039278080 (Chip design + LLVM RFC for using ML for compiler opts)