Updated
Updated · Ars Technica · Oct 1
Ataraxos Beats Stratego Champion 15-1-4 on Just 16 GPUs
Updated
Updated · Ars Technica · Oct 1

Ataraxos Beats Stratego Champion 15-1-4 on Just 16 GPUs

3 articles · Updated · Ars Technica · Oct 1

Summary

  • Ataraxos routed Pim Niemeijer — widely regarded as Stratego’s best player — with 15 wins, one loss and four draws, clearing a milestone that had eluded earlier AI systems.
  • Just 16 GPUs and a few thousand dollars were enough to train it, a sharp contrast with DeepMind’s earlier DeepNash effort in a game long seen as unusually resistant to machines.
  • Stratego’s difficulty comes from 40 hidden pieces that can start in more than a decillion arrangements, with identities revealed only through combat.
  • Games can stretch to 2,000 moves and hinge on calibrated bluffing, forcing an AI to manage long-horizon uncertainty rather than the smaller hidden-information space seen in poker.

Insights

How did Ataraxos crush world-class Stratego players while using far less training than DeepMind’s DeepNash?
If AI can infer hidden pieces and outbluff champions in Stratego, how close is it to real-world planning under uncertainty?
Why do researchers say interpretability—not raw strength—is the next obstacle for superhuman AI in hidden-information games?