Slope Loss Hunt

Slope Loss Hunt is a gradient descent visualisation game that teaches the algorithm behind every neural network ever trained — not through formula or code, but through a ball rolling down a hill. A loss landscape fills the screen. In two dimensions it is a glowing neon curve, red at the peaks and deep blue in the valleys. In three dimensions it is a false-colour terrain map — crimson ridges, amber slopes, teal approaches, navy valleys. A ball sits at the starting position. You set the learning rate. The ball moves. The learning rate controls the step size. Too high and the ball overshoots the valley, climbs the opposite wall, and bounces back — each time slightly further from the minimum. High enough and it never settles at all. Too low and each step is a fraction of a pixel. On a flat region of the landscape, where the gradient is nearly zero, the ball barely moves even after a hundred…