← All news·2026-10-01·3 min read

A cheap AI model ($8,000) defeated a world champion at Stratego

The Ataraxos team (Carnegie Mellon/NYU/Stanford/MIT) defeated Stratego world champion Pim Niemeijer using a budget of under $8,000.

airesearch

A group of scientists from Carnegie Mellon University, NYU, Stanford University, and MIT developed a system named Ataraxos and played a 20-match game with Pim Niemeijer, the world’s best human Stratego player (Stratego is a strategy board game with hidden information). Ataraxos won 15 matches, lost 1 match, and tied 4 matches. It took less than $8,000 to train this model, just one week of computation on 16 Nvidia H100 GPUs (these are usually the chips used to train machine learning models) and another 4 days to train a neural network to predict the locations of the opponent’s pieces. In 2023, DeepMind attempted to solve this problem using their model DeepNash, but they spent more than $3–4.5M and used more than 1000 TPUs (their own AI chips) and couldn’t even beat the world’s best human players such as Pim Niemeijer.

I am impressed by the difference between computational resources. With a tiny budget and a short time (one week, 16 GPUs), they achieved a much better performance than DeepMind’s (months, >1000 TPUs). In my experience developing agents, I often find that the setup of the problem and the choice of the model have a greater impact than the amount of computing resources used. This is not about programming skills, but still a useful indicator.

Source: the-decoder.com

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids