You could certainly do that to some extent, but I suspect that wouldn't generalize very well to programs that do things significantly different from anything in the training set. Transforming the syntax alone would probably be straightforward enough, but the parts that need more "interpretation" of what's going on (such as assigning plausible variable/function/class names, nevermind something like writing comments) I just can't see a standard supervised model handling particularly gracefully. Whereas that's one of the areas LLM excel at.

Good question.

The generative model might be able to learn with fewer examples? Because if already knows more about coding in various languages?

Just a guess.

These generative models also seem better at learning though right?

It could better understand what the user wants too