This seems like the path languages will come to evolve in. Especially with the advent of "cloud computing", I imagine a compiler which distributes processing among a remote resources can use a combination of a large company's internal code base and open source code to train ML algorithms can either provide either great code completion or build entire "ambiguous languages".
I don't know if you misunderstood or just didn't read the linked page, but this is a compiler for hierarchical Bayesian models, not a compiler that uses Bayesian inference in the process of compiling, neat as that would be.
