This is one of the papers Ian was talking about! It's about how models regurgitate training data and memorize data in different ways based on scale.
https://arxiv.org/pdf/2202.07646.pdf
