New AI System Masters Stratego, Outperforming Top Humans with Far Less Compute

A joint team from MIT, Carnegie Mellon, New York University, and Stanford has built an AI that beats world‑class Stratego players by a wide margin. Unlike earlier approaches that relied on massive search trees, the new model combines efficient training algorithms with decision‑making techniques tailored for hidden‑information games.

The system learns a compact strategy representation, allowing it to evaluate millions of possible board states without enumerating every permutation. In benchmark matches it outperformed the previous best AI by more than 30 percentage points while requiring roughly one‑tenth the GPU hours for training.

Beyond the board, the researchers demonstrated that the same architecture transfers to other imperfect‑information games, suggesting a general‑purpose engine for strategic reasoning under uncertainty.

Why it matters for GPU and AI infrastructure

Training cost is a primary bottleneck for organizations deploying large‑scale models. This breakthrough shows that algorithmic advances can dramatically reduce compute demand, letting GPU clouds like AiGpu deliver high‑performance inference and training at lower price points. Enterprises exploring negotiation, cybersecurity, or logistics simulations can now prototype sophisticated agents without provisioning massive clusters.