In an exhibition match at the 2025 Stratego World Championship, a new AI called Ataraxos won 38 of 40 games against tournament players and later beat Johan Niemeijer, the most decorated Stratego player ever. Niemeijer, who has four world titles and spent more than 600 weeks at the top of the rankings, lost the three‑week series with a win rate of about 85 percent. The matches were streamed online, and the AI’s style was described as hard to read and willing to take risky bluffs.
How Ataraxos pulled off the win
Ataraxos learned only by playing against itself, using a belief network to guess the opponent’s hidden pieces before each move. The researchers forced the AI to vary its opening setups early on, a technique called regularization, which kept it from settling into a predictable pattern. Training took one week on 16 Nvidia H100 GPUs and four additional days on four GPUs, costing under $8,000 at 2025 prices—about 1/500th of the compute DeepMind spent on its earlier DeepNash system.
Why it matters
For regular board‑game fans, the result shows that AI can now handle games where you can’t see all the pieces, ending one of the last areas where humans were clearly better. It also hints that similar AI techniques could be applied to real‑world problems with hidden information, such as finance or negotiations, though the paper notes the current method doesn’t improve simply by adding more compute.
