That gives me a kind of funny idea: train a sequence-to-sequence LSTM network on code written in two (or more) programming languages that implement the same functionality. You'd need a big corpus, and the could would probably would have to be a little more complicated than "hello world, but I don't see why it wouldn't work, in principle.