What makes you so confident that an AI-assisted compiler couldn't significantly enhance optimizations? A relevant example of a complex problem where neural networks have improved performance is found in chess engines. Today, top-level engines like Stockfish have integrated NNUE ("Efficiently Updatable Neural Network") which has significantly boosted their performance.
What makes me confident is the fact that I've been using gcc since 2.95 (well redhat had its patched "2.96" but yeah) so I witnessed the evolution of these parameters.
What makes neural networks and learning irrelevant is because there aren't a billion parameters, most parameters don't depend on each other. It is not a search problem. You have the wrong domain.
This is like saying "hey let's crack that password using neural networks, stockfish is so advanced!"
And people have tried auto searching parameters, the results were as I described in my response. Moreover, imagine releasing code and having cluebies complain about trying to make it go faster with -fomit-frame-pointer. The whole idea is wasteful and cringe.