And there's a bunch of obvious automated training it could do to be specifically better at software coding.

It could complete all the challenges on sites like leetcode and code signal, learning from it's mistakes.

It could be given challenges to take an existing program and make it run faster, learning from a timing analysis.

It could take existing programs and be asked to fix the bugs so it passes a unit test.

It could be asked to write a unit test that makes an existing program fail.

And so on. Ultimately millions of separate tasks that the machine can get objective feedback on how well it did on them, and so it can refine it's skills to be above human.