“A key problem with training on synthetic data from LLMs is that the model outputs are "silently collapsed," meaning they lack diversity and occupy a very small portion of the possible output space.”