First gen models trained on books directly. Latest Phi distilled textbook-like knowledge down from disparate sources to create novel training data. They are all fairly open about this change and some are even allowing upset publishers to confirm that their work wasn't used directly. So stones and ionized water go in the soup.