“Deep neural networks exhibit "benign overfitting," meaning they can perfectly fit random noise yet still generalize well on structured data.”