What changed
Ataraxos, an AI system built by researchers from Carnegie Mellon, MIT, NYU and Stanford, defeated elite Stratego player Pim Niemeijer 15 games to one, with four draws.
It trained through 163 million self-play games using 16 GPUs and a few thousand dollars. Its key addition was a second neural network that estimates the identities of hidden pieces from an opponent’s moves.
Why it matters
Stratego is a nasty test for machine reasoning. Each player has 40 pieces, but their ranks stay hidden until battle. Games can stretch to 2,000 moves, and bluffing matters as much as calculation. A weak piece can pretend to be a marshal; repeat that trick too often and it stops working.
Ataraxos’s breakthrough was teaching an AI to reason about uncertainty without examining every possible board. Before each move, its belief model samples plausible arrangements of the opponent’s hidden pieces. The system then tests candidate moves against those possibilities and chooses the strategy that performs best.
That is a meaningful technical result, not just another game-playing trophy. The difficult part was not raw compute. It was turning incomplete information into useful guesses while keeping self-play from becoming unstable. The reported training cost suggests that this kind of progress may be accessible to smaller research teams, at least in tightly defined environments.
The result should still be read narrowly. The match shows that Ataraxos can beat Niemeijer under the reported conditions; it does not establish that the system matches top human performance more broadly or that its method transfers to commercial or real-world decision systems.
Comments
No comments yet.