On “Language Models Need Sleep,” and the slow death of the train/test split