![The Legend Of Von Neumann (1973) [Pdf]](https://micronomicon.com/wp-content/uploads/2026/10/the-legend-of-von-neumann-1973-pdf-featured-200x110.jpg)
TL;DR
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
A research team from Carnegie Mellon, MIT, New York University and Stanford says its AI, Ataraxos, beat top Stratego player Pim Niemeijer by 15 games to one, with four draws. The team reports that training took 16 GPUs and a few thousand dollars, addressing a game defined by long play and extensive hidden information.
A research team from Carnegie Mellon, MIT, New York University and Stanford says its Stratego-playing AI, Ataraxos, defeated Pim Niemeijer, described in the report as arguably the game’s best player, by 15 games to one, with four draws. The researchers say the system was trained using 16 GPUs and a few thousand dollars, a result that could offer a lower-cost route to strong AI in games with extensive hidden information.
The reported result stands out because Stratego conceals the identity of each player’s pieces. Each player has 40 pieces, including military ranks, bombs and a flag. Players can see where their opponent’s pieces are, but learn what a piece is only when two pieces meet in battle. The goal is to capture the opposing flag.
The source report describes Stratego as a game of long-term planning, uncertainty and bluffing. A player may move a weak piece as if it were a powerful one to deter an opponent. Repeated bluffs can lose their force, while never bluffing can make a player predictable. The AI therefore has to make decisions without knowing the opponent’s full setup and respond as information emerges over a long game.
The reported match record was 15 wins for Ataraxos, one for Niemeijer and four draws. The research team’s institutions include four universities, and its reported training budget and hardware count were a few thousand dollars and 16 GPUs. Those figures are reported in the source; the material provided does not include a detailed accounting of the costs or training setup.
Stratego presents a different challenge from games in which all pieces and moves are visible. A system must reason about what an opponent may be hiding, update its assumptions when pieces clash, and choose when to bluff or act cautiously. The reported win suggests that AI performance in this kind of game may not require the scale of hardware and spending associated with some prominent earlier efforts.
The result is relevant beyond Stratego as a test of decision-making under uncertainty. The researchers’ approach could be of interest to people studying AI systems that must act with incomplete information. However, a result in one game does not establish that the same methods will work in other settings, or that the system’s performance extends beyond the reported match.
As an affiliate, we earn on qualifying purchases.
Why Stratego Resisted Earlier Systems
Computers have reached top human performance in several well-known games. The source report points to Deep Blue’s 1997 chess match against Garry Kasparov, AlphaGo’s 2016 victory over Lee Sedol in Go, and AI systems that have competed successfully in poker. Stratego posed a different combination of challenges: hidden piece identities, a large range of possible setups and games that can last a long time.
Co-author Eugene Vinitsky of New York University described the game as distinctive because “a massive amount of hidden information” unfolds over a long period. MIT co-author Gabriele Farina contrasted it with Texas Hold’em, where he said the hidden information amounts to two cards and there are 1,326 possible hands. In Stratego, he said, 40 pieces could be arranged in more than a decillion possible setups. Farina also contrasted chess games that usually last about 40 moves with Stratego games that can reach 2,000 moves.
The source report says that DeepMind’s DeepNash, introduced in 2022, had not reliably beaten the strongest human players. It presents the new result as a breakthrough against that earlier challenge, but does not provide a full comparison of Ataraxos and DeepNash’s training methods or performance in other match conditions.
AI chess and board game training set
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Limits of the Reported Match
The source material does not specify when the match was played, how Niemeijer and Ataraxos were assigned starting positions, or whether the results were independently replicated. It also does not explain the precise meaning of “a few thousand dollars,” including which costs were counted, or provide enough detail to assess the training process.
The win record is strong evidence of performance in the reported series, but the material does not establish how Ataraxos would perform against other elite players, across a larger range of match conditions, or under independent testing. The researchers’ claims about the result and cost should be understood within the scope of the source report.
As an affiliate, we earn on qualifying purchases.
Further Testing Will Clarify Reach
The next useful evidence would be additional matches against high-level Stratego players, clear documentation of the match rules and conditions, and a fuller account of how Ataraxos was trained. Independent replication would help establish whether the reported record reflects performance that generalizes beyond this opponent and series.
The source material does not name a follow-up match, release date or next research milestone. Until those details are available, the confirmed development is the reported result against Niemeijer and the team’s account of the resources used to train the system.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Ataraxos?
Ataraxos is an AI system developed by researchers affiliated with Carnegie Mellon, MIT, New York University and Stanford to play Stratego.
What was the reported match result?
The source report says Ataraxos beat Pim Niemeijer 15 games to one, with four draws.
Why is Stratego difficult for AI?
Players cannot see the identities of their opponent’s pieces until they meet in battle. Stratego also involves many possible piece setups, long games and bluffing, requiring decisions under substantial uncertainty.
How much did the team spend to train Ataraxos?
The researchers reportedly used 16 GPUs and a few thousand dollars. The available source material does not provide a detailed cost breakdown.
Does this prove Ataraxos can beat every top Stratego player?
No. The reported result covers a match against Pim Niemeijer. The source material does not establish performance against other elite players or show independent replication.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
