Machine learning
Small networks, explicit training loops, and results you can inspect. Sometimes they learn. Sometimes they find a very efficient way to be wrong.
LEARNING TO PLAY
Coordinates and pixels
Start with coordinates, then try learning directly from an image.
LESSON 01 / COORDINATES
Catch a star
Train a neural network using positions and the gap as inputs. Compare rewards as the policy changes.
4 inputs · 8 neurons · 67 parameters
Open experiment →LESSON 02 / PIXELS
Star catching from pixels
Use a grayscale game frame instead of coordinates. Inspect the input, network state and training results.
168 pixels · 24 neurons · 4,131 parameters
Open pixel experiment →Both lessons run in your browser, start with random weights, and include training controls, equations, and source code. Reloading clears the training.
HOW LEARNING WORKS
Gradients and network depth
LESSON 03 / BACKPROPAGATION
Backpropagation
Walk through a tiny network, one calculation at a time. See what derivatives mean, multiply them along a path, and apply a weight update.
7 visual steps · live equations · adjustable weights
Inspect the calculation →LESSON 04 / NETWORK DEPTH
Network depth
Compare four networks on the same curve. Watch their gradients, then rotate a 3D projection of the path their weights took.
0–8 hidden layers · shared data · 3D learning paths
Open depth comparison →LEARNING AGAINST AN OPPONENT
Competitive self-play
LESSON 05 / COMPETITIVE HOCKEY
Hockey self-play
Two independent agents learn to defend and score. Give them different network depths and watch them adapt to rebounds and an improving opponent.
7 inputs · 1–4 hidden layers · simultaneous self-play
Open self-play →LESSON 06 / PIXEL SELF-PLAY
Hockey self-play from pixels
Remove the coordinates. Two-frame observations let each agent learn movement and compete from the game’s pixels.
288 pixels · frame stacking · independent deeper brains
Open pixel self-play →THE WHOLE POPULATION
Population-level learning
LESSON 07 / 1,000 AGENTS
Agent population
A thousand independent hockey brains meet random peers and face a fixed predictive reference. Inspect individual histories, population spread, and recent plateaus.
1,000 persistent agents · fixed benchmark · real prerun data
Open population experiment →Training lessons now include “Load prerun data” so you can inspect real recorded runs without training from scratch.