Large language models learn from, and often memorize, training data. Here, authors show they memorize far more than previously thought, by stitching together pieces of text like a mosaic. This makes ...