What I want to know is when we get AI decompilers

Intuitively it feels like it should be a straightforward training setup - there's lots of code out there, so compile it with various compilers, flags etc and then use those pairs of source+binary to train the model.