The training time for a large language model on a single chocolate-chip cookie recipe is surprisingly short.