Mind Reader

The problem

Matching pennies is a game with a known best strategy: flip a fair coin. Do that and you tie at 50% against any opponent, forever. The strategy is simple, published and known to every player, and people still can't carry it out. That gap between knowing the right move and being able to make it is what this game is about.

It is also a piece of computing history. In the 1950s, Bell Labs built two relay machines that predicted a player's next move: Shannon's MRM in 1953 and Hagelbarger's SEER in 1956. Their designs were published, but nobody had rebuilt them as something you could play. Single-model versions of this game already exist online, so I set a higher bar for the first version: both historical machines, their rematch, eight models at once, and a portrait of how you play.

What I built

The guess is locked before you press

The whole game depends on trust. If the machine could see your press before guessing, its score would mean nothing. I built it so each prediction is sealed first, and the code that scores a round receives that sealed guess from outside. There is no path that lets it guess after reading your input. The press buttons don't even respond until a seal exists. At the end, the app replays your session through a fresh machine and shows a round-by-round audit on screen.

The machine only sees which way you pressed

Your reaction time and where you tap would be strong clues. Using them quietly would betray the premise. Each predictor receives only two lists: your past presses, and its own past guesses. There's no place to pass timing or position. The limit is part of how the code is shaped, so it can't slip in later.

When unsure, the machine flips a coin

If the eight models aren't confident, the machine plays a random move, as the original 1950s machines did. That keeps it at 50% against a player who really is random, and stops anyone from reverse-engineering it. Before every deploy, a test plays 100,000 rounds against a simulated coin and requires the result to settle at 50%. If it ever fails, I don't tune it. A machine that beats a coin is seeing the future somehow, and the same bug would inflate its score against people.

The historical machines have to replay history

I rebuilt both machines from the original 1953 memorandum and 1956 paper, not from modern summaries. Where the sources leave a detail open, I wrote down my assumption, and those notes ship inside the app. A second test pits the two machines against each other for thousands of rounds and requires Shannon's simpler machine to win, as it did in the 1950s. If it doesn't, one of my reconstructions is wrong.

Eight models of you, competing in public

Eight predictors play against you at once, and you can watch their weights change as the game goes on. Switch your strategy halfway and a different model takes over the vote. After a session, a portrait measures six things about how you fail to be random, including how often you switch sides, how long your streaks run, and which patterns you repeat too often.

Result

Mind Reader is live and public. It runs the eight predictors, the historical rematch, the portrait and a strategy lab where you can try tactics such as reciting digits of pi. Only an actual coin works. The machine answers within 80 milliseconds, so you never see it thinking. It reports its score and never gloats. I built the first version alone over about a week in early September 2026, then kept extending it through the rest of the month, adding three more predictors and the own-guesses signal each one now reads.