Unlock: Model Collapse and Data Quality

When models train on their own outputs, the learned distribution narrows, tails disappear, and quality degrades across generations. Why synthetic data feedback loops threaten pretraining data quality and how to mitigate them.