Interesting, because my point was actually the opposite.
For a generative NN to help train a driving NN, it would need to have already solved self-driving in order to create the correct action labels for the created scenario (on top of creating the new scenario).
Elon mentioned something like this previously, if you can accurately simulate the real world, you've already solved the problem.
The auto labing system is a pale form of this setup: a much larger, slower, more power hungry NN that can classify the scenario which is used to train the driving NN. However, it only deals in what was, not what could be.