Ataraxos beats Stratego’s top player 15–1, with training under $8,000

An AI system called Ataraxos beat top-ranked Stratego player Pim Niemeijer 15 games to one, with four draws, in a 20-game series. The result, reported by Ars Technica, is notable not just for the win: the researchers say they trained the system for less than $8,000 using 16 Nvidia H100 GPUs over one week.

Watch Desk analysis

What happened

Researchers from Carnegie Mellon, MIT, NYU and Stanford developed Ataraxos to tackle Stratego, a game of long-term planning, bluffing and hidden pieces. The report says the research was published in Nature.

The system combines self-play reinforcement learning with decision-time planning, using a generative model to evaluate hidden information. It also uses a belief network and dynamic damping to manage uncertainty about opposing pieces and strategy. The reported results extend beyond standard Stratego to include Barrage Stratego, Hanabi and Dou dizhu.

Why it matters

Stratego’s hidden information makes it a different sort of challenge from games where both players can see the whole board. Beating an elite human player by this margin is a concrete result in that setting, while the reported training budget makes the achievement more striking than a win bought with an enormous computing bill.

The result also points to a broader research question: how well can AI systems combine learned strategies with planning when they must reason about information they cannot see? The reported performance in other imperfect-information games suggests the approach may travel, though the evidence here does not establish how it performs beyond the games described.

Our read

This is a proper AI milestone, not simply a machine winning at another board game. The combination of hidden information, a decisive match result and a comparatively modest reported training budget gives readers something concrete to assess. The next test is whether other researchers can reproduce the approach and show where it works, and where Stratego remains the star of the show.

What to watch

• Whether the research paper provides further detail on the match setup and evaluation.
• Whether other teams reproduce Ataraxos’s results and training costs.
• How the approach performs across other games involving hidden information and bluffing.

Discussion spark: Does Ataraxos’s reported low training cost make this a more important AI result than the match victory itself, or is the win the real breakthrough?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.