You may have noticed that currently there is a flood of strong backgammon apps / web sites. The reason is simple, as I learned this week. Tell a current AI like GPT 6.1 or Claude 5.5 to write a state of the art backgammon engine and it does.
Actually it is not as easy as writing a chess engine. Did that last week. My engine “Dark C. Helmet” (GitHub link if you’re a developer) took a long self-play on my Mac to learn 1 billion positions. Now it plays at about 3400 ELO, way better than every human, but still destroyed by the strongest engines on the planet.
So I asked Claude Opus 5.5 to build a state of the art backgammon engine. Not there yet, but amazing anyway.
- My engine is insanely fast – written by the AI in C++, the language you use for really performance-critical tasks I play my test matches at 5 ply, thinking deeper than the pro apps I know.
- Its checker play is state of the art. Have played 5 money games and analysed them with BGBlitz. Not a single mistake.
- It sucks at cube handling. Right now it uses a formula based on win/gammon/bg probability. That is not enough; in one game it made mistakes worth over 0.3 points of equity. Tricky to train.
- Still, even with some cube issues, about 100 self-play games and matches analysed by BGBlitz shows: the engine is in the range of PR 0.5 and 2, probably on the lower end (read: super-human, yay!)
Superhuman checker play after 2 days of improving itself by self-play. Not bad. Currently it has trained its neural network on 16.7 million positions. Nothing compared to where my chess engine is. But the network is much slower, maybe the project is already coming to an end.
No, I don’t know what I’ll do with it once it is finished.