> First of all, training off of data generated by another AI is generally a bad idea because you'll end up with a strictly less accurate model (usually).
That is not true at all.
We have known how to solve this for at least 2 years now.
All the latest state of the art models depend heavily on training on synthetic data.
