LLMs are to a certain degree compressed databases of their training data. But 42% is a surprisingly large number.