either. In many cases we might be limited by our training compute budget, so this approach is a non-starter. While techniques like data-augmentation, distillation etc. as introduced in chapter 3 do help us