Isn't that what's going on with synthetic data? The LLM is trained, then is used to generate data that gets put into the training set, and then gets further trained on that generated data?