GIPFboard study

About the AI

A small network that taught itself GIPF.

Your opponent is a small neural network paired with look-ahead search. It runs entirely in your browser: the download is under 1 MB and no server is involved. Nobody showed it how to play. It learned GIPF only by playing against itself.

about 217,000 parametersunder 1 MBno serverself-taught

Architecture

From board to move

Every time the machine thinks, the position travels through this pipeline thousands of times. Select a step to see what happens there.

Look inside

Where each point looks

This is real data from the network you play against. Click any point on the board, then pick one of the 16 attention heads: the board shows how strongly that head prefers to look at each other point from there.

Attention heads

Head 1Head 2Head 3Head 4
avoidsfavours

These are the network’s learned position preferences: which relative positions each head favours before it even looks at the pieces. Nobody told it about the geometry. Many heads have discovered the board’s lines and neighbourhoods by themselves.

Training

How it learned

Step 1

Play itself

It played about 390 million positions against itself, over two days on rented graphics cards.

Step 2

Teacher and student

A larger “teacher” network played the games, while the small network you play against learned from them.

Step 3

Learn from search

Every move in training was chosen by search. The network learned to predict the search’s choice and the final winner.

Step 4

Pick the best

The best version was picked by playing candidates against each other, at the same thinking time, in the real browser engine.

Design choice

Why small?

In the browser, thinking time is what counts. A small, fast network can search far more positions in the same time, and that beat bigger networks at equal thinking time early in training and matched them at the end. So the smallest one was shipped.