Exactly: Imagine you deploy a neural network in production, would you rather have 100 small errors in predictions distributed, or 99 perfectly correct and one very big error that might break your machine?