Ataraxos AI beats top human Stratego player, first superhuman result with far less compute
zicokolter · x · 2026-10-01
A Nature open-access paper from researchers including Samuel Sokota, Eugene Vinitsky, J. Zico Kolter, and Gabriele Farina introduces Ataraxos, an AI for Stratego, a massive hidden-information board game.
- Ataraxos combines general-purpose self-play reinforcement learning with test-time search under hidden information, without game-specific engineering.
- It decisively defeated the most decorated human Stratego player of all time — reportedly the first superhuman result in the game's history.
- It used orders of magnitude less compute and data than previous multimillion-dollar industrial efforts to crack the game.
- Published in Nature vol. 658, pp. 55–59 (Sept 30, 2026).
Related event: Superhuman Stratego AI Ataraxos Unveiled in Nature(3 posts)→
More from Research
- ctjlewis offers $15,000 in Astra credits to mathematicians — ctjlewis · 2026-10-01
- Gemini 4 Argon beats quantum computing baseline by 40% in minutes, says Google researcher — LinusEkenstam · 2026-10-01
- Qwen3.8-27B-pi: Effort-Ordered Reasoning for Agentic Coding on Hugging Face — paf1138 · 2026-10-01
- State of Clinical AI 2026 report published in BMJ Digital Health & AI — davidjhwu · 2026-10-01
- UChicago brain-machine interfaces let paralyzed patients feel touch via thought — plopesresearch · 2026-10-01
- CLM author responds to RLM debate: context as a variable, and the two can combine into stronger agents — RulinShao · 2026-10-01