The 64 first-layer filters of a trained AlexNet (the single-tower torchvision variant), each an 11×11×3 volume of learned weights, drawn as colour tiles. Click one: the scene is convolved with it live, and the signed response is shown the way Chapter 1 drew Sobel — mid-grey zero, white positive, black negative. The achromatic striped filters respond to oriented structure at one scale; the smooth two-colour tiles respond to colour boundaries and hardly care about form. Nobody chose any of these numbers.