A Bot Just Beat Stratego’s Greatest Player, and It Plays Dirty

Pim Niemeijer has won four Stratego world championships, fifteen Dutch national titles, and spent more than six hundred weeks ranked first in the world. One of his rivals, a man who has played in every world championship since 1997, calls him the best Stratego player who ever lived.

Last year he sat down for a twenty-game series against a computer, played out over three weeks, and lost fifteen. He won once. The other four were draws.

The result, published last week in Nature, is the first time a machine has beaten a top human at Stratego, and it closes out something I’d started to think might hold forever. Chess fell in 1997. Go fell in 2016. Poker went a few years after that. Stratego kept standing, for a reason any kid who has played it grasps without being told: you cannot see your opponent’s pieces.

Why Stratego Held Out When Chess and Go Didn’t

Stratego is a two-player game of pure deduction, and deduction is exactly what machines found impossible here. In chess, both players stare at the same board. Nothing is hidden, so a strong engine only has to look further ahead than you and pick the move that leads to the best position. Stratego breaks that. Your forty pieces start face down, their ranks secret, and the whole game is spent guessing what the marshal, the bombs, and the flag are hiding behind. You bluff a weak piece as a strong one. You march a scout into enemy lines just to read a rank. You sandbag your spy until the turn it matters. The board you see is a sliver of the board that exists.

Hidden information on that scale is what computers choked on. The researchers put a number on it: a game of Stratego has more than a decillion possible starting arrangements, which is a one with thirty-three zeros after it.

Poker hides information too, but a Texas hold’em hand is one of only 1,326 possibilities, small enough for a machine to grind through every branch. Stratego’s hidden space is so much larger that the tricks which cracked poker never scaled to it. For years that left it as maybe the only classic game where serious, well-funded attempts still couldn’t beat the best humans.

DeepMind Spent Millions and Still Lost. This Cost $8,000.

The best-known of those attempts was DeepNash, built by DeepMind and published in 2022. It was serious work and it played a strong game, but it never reached the summit. Shown off at the 2023 world championship, it went nineteen wins to nine losses and lost to most of the top-ranked players it faced, Niemeijer included. By the recollection of its own creators, training it ran on more than a thousand specialized chips for two to three months, which at today’s cloud prices would cost somewhere between three and four and a half million dollars.

The new system, called Ataraxos, came out of a team at Carnegie Mellon, NYU, Stanford, and MIT, and it cost less than eight thousand dollars to train. Not less by a hair. The researchers put its compute bill at roughly one five-hundredth of DeepNash’s, on a small fraction of the practice games.

This is the number the tech press is leading with, and fairly so, because the gap is stark: a university lab did on a graduate budget what an industrial research arm couldn’t do at a thousand times the spend. The edge came less from raw power than from a smarter way of teaching the system to reason about what it can’t see, then search through a handful of plausible versions of the hidden board before each move.

The Bot Plays Rude, Reckless, and Impossibly Lucky

A board gamer gets more out of this next part than an engineer does. The people who played Ataraxos described its style, and it does not play like us.

It sets up aggressively, shoving strong pieces toward the front where humans keep them protected, and tucking its flag into back corners players consider nearly impossible to hold. Its pieces are hard to read, moving in patterns that don’t leak the way human habits do, and its scouts stay on the board far longer than a person would keep them. Some bluffs top players consider too reckless to try, it pulls anyway; others that humans lean on, it skips.

Falling behind, it refuses to go quietly, stalling, needling, and dragging things out in ways some opponents found openly rude. Pulling ahead, it takes gambles that read as arrogant, as though it can’t be bothered to respect the player across the table.

One description from the paper stuck with me, because it names a feeling every Stratego player knows from the losing end. Ataraxos seems lucky. It doesn’t get lucky, exactly, since it builds its own setups and makes its own reads, but it has the eerie quality of always seeming to hold the right piece on the right square, of watching its gambles land, of nudging its opponent into doing precisely what it wanted. Anyone who has lost to a stronger player across a table knows that feeling. A machine, it turns out, can manufacture it too.

None of this is how the strategy guides tell you to play. The bot plays a game that looks wrong and keeps winning anyway.

What a Machine Beating Stratego Means for Your Table

The researchers didn’t stop at the main game. Using the same method, they built a superhuman player for Barrage Stratego, the stripped-down eight-piece variant, and set records at Hanabi, the cooperative game that won the 2013 Spiel des Jahres, and at dou dizhu, one of the most played card games in China. That spread is the result that matters. One approach now handles the game where you lie to your opponent and the game where you build trust with a partner you can barely talk to. Hidden information, the wall that kept all of them out of reach, isn’t the wall it was.

So that part is settled, and it matters for the fields that care about machines deciding without the full picture, which is most of them. What it changes about the Stratego box in your closet is nothing. No chess engine ever emptied a chess club. Stockfish has been demolishing grandmasters for years, and people play more chess than ever, because sitting across a table and reading another person was never about being the sharpest calculator in the room.

Stratego is the same, only more so. Its whole pleasure is the bluff that lands because your friend believed you, or the bomb they walk into because they trusted the wrong square. A computer in a Pittsburgh lab outplaying Pim Niemeijer takes nothing from the game you’ll set up this weekend. It only means the last board game nobody could teach a machine finally got taught. The ones we play were never the point of the exercise.