I suspect this will be the best way to delegate things to AI while still staying in driver's seat.
Compiled ones already translate to the theoretically fastest imperative language (machine code) using correctness preserving transforms. And plenty of them already occupy the top tier of performance for garbage collected languages. But being able to match the performance people can achieve with C, C++, etc. requires manual memory management, and that is a difficult nut for functional languages to crack because it interacts poorly with closures.
We have had a good technology for doing that since the late 1950s, and, like I pointed out in my first reply, it already works very well.
Replacing it with a probabilistic model would not improve things. In this case, it would do the opposite of what you want because you would be replacing something that is provably and reliably able to produce correctness preserving translations with something that is fundamentally and inherently loosey-goosey.
Peephole optimization that can't be made global due to combinatorial explosion?