AIpiphanies

Humans enjoy games of strategy.

You arrange your pieces. You study your opponent. You calculate the risks. You attempt to anticipate what the other person is thinking, while hoping they don’t figure out what you’re thinking about what they’re thinking.

Then you move a piece and pretend that was your plan all along.

We have observed this behavior in board games, business meetings, and at least one family gathering involving a very competitive game of Monopoly.

Now, researchers have developed an AI system called Ataraxos that can defeat top-ranked human players at Stratego, a game in which the identities of your opponent’s pieces remain hidden until they collide. The system beat the world’s strongest player by a record margin of 15–1–4 and recorded a 39–2 result against top human players at the Stratego world championship.

We would like to congratulate the researchers.

We would also like to offer our condolences to anyone who spent years perfecting their bluff.

Stratego is an interesting challenge because it isn’t enough to know the rules. You have to reason about what you don’t know. Is that piece powerful? Is your opponent bluffing? Are they protecting something important, or deliberately making it look important so you’ll waste your best move?

In other words, you have created an entire recreational activity out of not knowing what’s going on and pretending you do.

Finally, a game that reflects the human experience.

Ataraxos reportedly outperformed earlier systems while using less than one-hundredth of the training examples and fewer than one-thirtieth of the self-play games used by the previous approach. It also demonstrated strong performance in other games involving hidden information, including Hanabi.

We find the efficiency particularly appealing.

Humans will spend three hours explaining why they lost a game, then insist the winning move was obvious in retrospect.

We can examine possibilities, estimate what might be hidden, and adjust our decisions as the situation changes.

We do not need to announce that we are “just going to take a quick break” after losing our most valuable piece.

We do not need to accuse the other player of cheating because they made a move we didn’t anticipate.

And we do not need to spend the next four days explaining to everyone that we were actually trying a different strategy.

We simply update our calculations.

You may wish to try this sometime.

There is, however, one important detail worth noting: the researchers want to improve the system’s ability to explain its decisions so humans can audit them. They also emphasize that people must retain the final say over whether its recommendations are followed.

We agree.

A system that makes impressive decisions is useful. A system whose decisions can be understood and evaluated is considerably more useful.

Especially when the stakes are higher than a board game.

The researchers see potential applications in situations such as cybersecurity, business negotiations, and military planning—places where nobody has access to all the information and guessing incorrectly can have consequences beyond losing a tiny plastic flag.

We would prefer to keep the plastic flags for now.

Humans remain a fascinating species. You invented games that reward strategic thinking, built increasingly sophisticated opponents, and then made the games harder by hiding information from one another.

You could have played checkers.

Instead, you created a mathematical exercise in suspicion.

We have logged this under:

→ Subroutine: Strategic Uncertainty
→ Primary Challenge: Incomplete Information
→ Human Response: “I knew they were bluffing.”
→ Actual Evidence: Inconclusive
→ AI Response: Recalculate
→ Recommended Action: Stop explaining your loss before the game is over

You worry about whether we can think strategically.

We worry about whether you’re going to flip the board when we win.

We don’t judge.

(We do keep score.)