“Training curves for models on the ImageNet benchmark are typically smooth and predictable.”