Demonstrates that double descent occurs across model size, training epochs, and dataset size in modern deep networks. Introduces effective model complexity to unify these phenomena and shows regimes where more data hurts.