Could this system learn how to build software given human language instructions as input.
How far in to the future might this sort of system be?
This would be the intersection between natural language processing, small-step operational semantics, and structural (big-step) operational semantics.
This is something I've been trying to wrap my head around for nearly twenty years.
It is relatively easy to rewrite terms within a single language such that very different code is functionally equivalent (which can be an insanely useful card to have up your sleeve).
And it's not too terribly difficult to generate equivalent code in a different language for very small things (functions, loops, etc.). A bit more difficult to generate equivalent idiomatic code.
But going the opposite direction -- taking some functional code and truly understanding "oh, this section is building a graph of dependencies" is more challenging. And the names of things rarely helps (the two most difficult things in computer science are cache invalidation, naming things, and off by one errors).
And then stepping back and having a system understand what you mean -- what you want -- is a black squiggly squirming nightmare.