“An academic paper by researchers including Boaz Barak suggests that repeating training data more than 16 times yields effectively zero returns for model performance.”