Something I'm wondering about: why haven't the model collapse predictions come to pass, even as models are increasingly, explicitly getting trained on synthetic data? Or has "collapse" been manifesting in less obvious ways?