How does this relate and compare to actual type inference? For example, Common Lisp implementation SBCL is able to infer types pretty nicely now and it doesn't need neural networks for "predicting" the types with some kind of chance.
"FIXME: The material in the CMUCL manual about getting good performance from the compiler should be reviewed, reformatted in Texinfo, lightly edited for SBCL, and substituted into this manual. In the meantime, the original CMUCL manual is still 95+% correct for the SBCL version of the Python compiler."
Whole program analysis is too expensive for type inference (or much of anything for that matter) even with its precision ramped all the way down via something like CFA0. Type inference with sub typing is a very similar problem to alias analysis, which hints at why doing it in any but very restrictive contexts is too expensive.