Artificial Intelligence
AI beats Stratego player in Nature study
Researchers at Carnegie Mellon, NYU, Stanford and MIT reported in Nature that Ataraxos beat Dutch Stratego champion Pim Niemeijer 15-1-4, and said the system was trained for under $8,000.
- In a Nature paper, researchers from Carnegie Mellon University, New York University, Stanford University and MIT introduced Ataraxos, a Stratego system.
- They reported that Ataraxos beat Pim Niemeijer 15 games to one, with four draws, in an official 20-game series.
- The authors estimated that training Ataraxos cost under $8,000 at 2025 prices, and compared that with a $3 million to $4.5 million figure for DeepMind's DeepNash work.
- They said the same method also produced strong results in Barrage Stratego, Hanabi and Dou dizhu, and they released the code.
The paper in Nature said four universities had built Ataraxos for Stratego. The authors said it was the first system they knew of to reach superhuman play in the game, and it defeated Pim Niemeijer by 15 wins, one loss and four draws in a 20-game match.
The article said Niemeijer had won four world championships, 15 Dutch national titles and two online world championships, and had spent more than 600 weeks at the top of the rankings. It also quoted George Franka as saying Niemeijer was the best Stratego player of all time.
The report compared the project with DeepMind's earlier DeepNash work. It said DeepNash had trained on 1,024 TPU nodes for two to three months, while the Ataraxos authors estimated that training their system cost less than $8,000 at 2025 prices.
The article said the researchers attributed the result to self-play, regularisation and a belief network that helped model hidden pieces. It also said the same approach worked in Barrage Stratego, Hanabi and Dou dizhu, and that the code was made public.
Named in this story
People
- Pim Niemeijer
- the Dutch player who lost the 20-game series
- George Franka
- said Niemeijer was the best Stratego player of all time
- Samuel Sokota
- first author who wrote about the method on X
- Max Roelofs
- commented on how narrow margins are at the top
- Vincent de Boer
- described multiple-game play as a handicap
Companies
- DeepMind
- made the earlier DeepNash Stratego system
Organisations
- Carnegie Mellon University
- one of the universities whose researchers presented Ataraxos
- New York University
- one of the universities whose researchers presented Ataraxos
- Stanford University
- one of the universities whose researchers presented Ataraxos
- MIT
- one of the universities whose researchers presented Ataraxos
Products and systems
- Ataraxos
- the Stratego system that beat Pim Niemeijer
- Stratego
- the board game the system was trained to play
- DeepNash
- DeepMind's earlier Stratego system
- Barrage Stratego
- a Stratego variant used in further tests
- Hanabi
- a cooperative card game used in further tests
- Dou dizhu
- a card game used in further tests
- PerfectDou
- a previous benchmark system beaten in Dou dizhu
- DouZero
- a previous benchmark system beaten in Dou dizhu
How the source tells it
The piece mixed research reporting with celebratory framing, leaning on decisive scores, first-ever claims and a low-budget-versus-big-tech contrast.
- triumph a decisive match score and the claim of superhuman performance
- underdog framing the cost comparison between an academic team and DeepMind's earlier system
- novelty language about the first system to clear a major game and the wider breakthrough claim