Can Julia JIT compiler "learn" something?

During his fascinating keynote about formalizing Julia type system, prof. Jan Vitek said that you can think about JIT compilers as machine learning programs (at 45:51). The main problem is that JIT compilers have now long time memory. You turn it off and it forget everything.

I’m very intrigued by his words and I want to know, is any sense/perspective for something like that in Julia? Prof. Vitek says that it can be connected to (in)famous time to first plot problem, due to eliminating time of “relearning” at second run, so it have so real benefits.

Or maybe this is something only for “promised” in 2016 or 2017 Julia 10.x?