Counterintuitively, it’s a lot easier to train a big model and distill it down to a small one than to train the small one from scratch. So never say never.