Backgammon AI
A neural network that has effectively solved practical backgammon — in your browser tab.
Tap to add or remove checkers, or drag a checker to move it. Drag to the right tray to bear off.
The engine (~3 MB) loads on first analysis and then runs instantly in your browser.
How the AI works
Backgammon was the first major game where neural networks overturned human theory: TD-Gammon in the early 1990s taught itself by self-play and promptly corrected plays the books had recommended for decades. The engine here, GNU Backgammon, is a mature descendant of that lineage — a network trained on millions of positions that maps any board state to five probabilities: win, gammon win, backgammon win, and the same for losses. Every candidate play is scored by feeding its resulting position through the net.
The depth selector controls how far ahead the AI looks. Instant (0-ply) is the raw network evaluation; 2-ply — the default, and the standard used in published analysis — considers every opponent reply and its answers; 3-ply looks a full move deeper for the hardest positions. All of it runs as WebAssembly on your device, which is why the analysis is free, private and unmetered: there is no server doing the thinking. If you want to see the AI disagree with old human theory, the 5-3 opening is a classic example — the book play of the 1970s loses to the bot play by a clear margin.
Frequently asked questions
How strong is this backgammon AI?
GNU Backgammon at 2-ply plays at a level comparable to the strongest humans, and modern bots of this class are considered clearly stronger than any human over a long session. Its evaluations agree closely with commercial references like eXtreme Gammon.
Does the AI learn from my games?
No. The network’s weights are fixed — it was trained by self-play long before it reached your browser. Your positions are evaluated locally and never leave your device.