Gradient Hill Climber

Learn optimizers by driving a ball down a loss landscape

Pick a terrain, drop a ball, tune the optimizer, and watch how step size and momentum change the path. This MVP is 2D, fixed-camera, and focused on free play.

Landscape

Loss: 0.000

Speed: 0.0000

Steps: 0

Click to drop the ball

Optimizer Controls

Optimizer

Learning rate0.150

Run

StatusIdle

Show contour grid

Hints

Large learning rates oscillate or diverge.
Momentum smooths noisy gradients but can overshoot.
Adam adapts step sizes and handles steep ravines.