Gradient Golf is a free interactive game that teaches gradient descent — the most important optimization algorithm in machine learning. You see a colorful loss landscape (dark blue = low, red = high) and your goal is to guide a ball to the deepest valley.
Each level introduces a new optimization challenge: ill-conditioned valleys, saddle points, local minima, noisy surfaces, and more. Choose your starting position and learning rate carefully — a rate too high will overshoot, too low will stall. Later levels unlock momentum, showing how it helps escape flat regions and narrow ravines.
Gradient descent is taught in every machine learning course, yet most learners only see it as a formula on a whiteboard. Gradient Golf bridges that gap by letting you interact with the algorithm in real time. When you place the ball on a steep cliff and watch it fly off the landscape because the learning rate was too high, the concept of exploding gradients becomes visceral rather than abstract. When you toggle momentum on a saddle-point level and see the ball coast through the flat region, you understand why optimizers like Adam and SGD with momentum exist.
Each of the twelve levels is designed around a specific optimization scenario encountered in real neural network training. The difficulty ramps from a simple convex bowl — where almost any reasonable learning rate converges — to the Rosenbrock banana valley and multi-modal landscapes that mirror the challenges of training deep networks with millions of parameters.
Part of the Kudos AI Play & Learn collection. Read our Complete Guide to Gradient Descent for the full theory.