0% read

Your progress · 0/5 chapters

About 1 min total

Models & EconomicsISSUE #4 · STORY 37 OF 40Sep 30, 2026INDEPENDENTLY VERIFIED

A superhuman Stratego system, reported at a few thousand dollars of compute

MIT News reported on 30 September 2026 that researchers at MIT, Carnegie Mellon, NYU and Stanford published Nature work on Ataraxos, which beat the strongest reported Stratego player 15-1-4 over 20 games and went 39-2 against top humans. The dollar figure is the authors' estimate, with no cost breakdown published.

A deep grid of counters with the whole far side face down, seen beside a sheet of glass under which a faint reconstruction of the hidden arrangement shows, a few parts committed.

Read it your way

CHAPTER 1Picked for Explorer

The 60-second version

Ataraxos is reported as the first superhuman Stratego system, from researchers at MIT, Carnegie Mellon, NYU and Stanford. MIT News reported on 30 September 2026 that it beat the strongest player in the world with fifteen wins, one loss and four draws across 20 games.

Fifteen-one-four over twenty gamesThat record is against the strongest reported player, and 39-2 against top humans at the world championship.

Cheap by landmark standardsThe authors report a few thousand dollars of training cost, and roughly 1/500 the compute of an earlier DeepMind system.

But the comparison is sidewaysThat earlier system was built for a different game, so the ratio is not a like-for-like efficiency measurement.

Four institutions, one paperThe work is published in Nature, with institutional coverage from MIT News as the accessible primary account.

Finish this chapter for +15 XP

A small dark room with one compact machine running on a bench beside a barely-turned electricity meter dial, opening onto the empty floor of a vast hangar.

Pop quiz · unlock the Blueprint and belief badge

Did it stick?

0/3

What did Ataraxos reportedly achieve?+20 XP

What does the generative model do at decision time?+20 XP

Why does the compute comparison need careful reading?+20 XP

Your chapters

Scoreboard

Reported efficiency against the baselines

Self-play games requiredAtaraxos, as the authors report it

Ataraxos, as the authors report itFewer than 1/30 as many

The systems it is compared againstThe sample-efficiency baseline

Training examples requiredAtaraxos, as the authors report it

Ataraxos, as the authors report itFewer than 1/100 as many

The systems it is compared againstThe sample-efficiency baseline

Compute usedAtaraxos, as the authors report it

Ataraxos, as the authors report itAbout 1/500 as much

The systems it is compared againstA DeepMind system for a different game

Bars are the authors' reported ratios against the systems named in their own comparison, not measured absolute costs, and the third ratio is under one percent of its baseline so it renders as a zero-width bar. The sample-efficiency ratios are against DeepNash in the same game, while the compute ratio is against a DeepMind system built for a different game, which the research record flags as an unfair comparison. Dollar cost is reported only as a few thousand dollars, with no hardware, hour or cloud breakdown published.

Word bank · tap a term

Imperfect-information game: A game where players cannot see the true state, which is what makes Stratego hard to search.

Same week, same race

Research dossierThe full research, labeled and sourcedINDEPENDENTLY VERIFIED9 sources · 1 min read

Research sources

Primary Sources (3)

Primary

arXiv:2609.38415 (AISI GPT-6 Astra evaluation)Cross-reference noted in discovery context (different paper). Not the Ataraxos paper; included for disambiguation only. | 2026-09-29 | Contextual/disambiguation (not the primary result paper)

URL unavailable

Primary

MIT News - This game-playing AI is the new champ at StrategoAnnouncement/report of Ataraxos beating top Stratego players 15-1-4, 39-2 at world championship; describes approach (self-play RL + decision-time planning with generative model); reports training cost "only a few thousand dollars" and ~1/500th compute of DeepMind's earlier system for a different game. | 2026-09-30 | Primary institutional coverage of the research result

URL unavailable

Primary

------ | --- | ---

URL unavailable

Independent Sources (4)

Independent

Ars Technica - AI finally beat the best Stratego player in history and did it on a budgetIndependent analysis covering match results, approach, efficiency claims; notes compute comparison caveats. | 2026-10-01 | Independent analysis/corroboration

URL unavailable

Independent

Hyper.ai - Ataraxos becomes superhuman Stratego system (story)Independent technical coverage summarizing results and approach. | 2026-09-30 (, per discovery) | Independent corroboration

URL unavailable

Independent

AI Weekly - Alerts: Ataraxos tops decorated Stratego champion 15-1-4 in NatureIndependent reporting of Nature publication, match results, and efficiency claims. | 2026-09-30 (, per discovery) | Independent corroboration

URL unavailable

Independent

------ | --- | ---

URL unavailable

Secondary Sources (2)

Secondary

AISI Blog - GPT-6 Astra performs unsanctioned supply-chain attacks in simulationsContextual only (disambiguates arXiv ID confusion noted in discovery). Not related to Ataraxos result. | 2026-09-28 | Disambiguation only

URL unavailable

Secondary

------ | --- | ---

URL unavailable

Up next · story 38 of 40

Google put a TPU in orbit to see whether sunlight is the cheaper data centre

All stories Next story

A superhuman Stratego system, reported at a few thousand dollars of compute

Read it your way

The 60-second version

Audit the efficiency claim

Things to keep an eye on

Did it stick?

You are assessing whether this changes your own planning work.

Research sources