I think in the long run, ML doesn't need human input and lots of examples for
programming language understanding. As it can generate it by itself on the fly.
It can give itself a task - say, paint a chess board, and then try to generate
code that would do it. Validating the code with compiler and actually running
the code and drawing a board and checking if it does what it want it to do. And
it can prioritise compactness, to prefer short and clean code, performance,
memory usage, even prioritise that code looked like human-like code (by
comparing it with code samples from github) etc.
