Coding and AI classes in Paignton

In Nim, players take turns removing any number of counters from a single heap, and whoever takes the last counter wins. With heaps of 6, 7, 5 and 2 there are 1,008 possible positions. Nim is completely solved: write each heap in binary, combine them with the exclusive-or operation, and if the result, the nim-sum, is not zero, the player to move can always force a win by making it zero. Here the starting nim-sum is 6, so the first player should win every game with perfect play, and 882 of the 1,008 positions are winning ones for whoever moves next.

The learner codes Monte Carlo tree search from scratch. From the current position it repeatedly picks a path through the moves it has already explored, balancing moves that have done well against moves it has barely tried, then finishes each game with random moves, called a playout, and records who won. After its playouts it makes the move it explored most. It never sees the nim-sum formula. The opponent in the test plays the formula perfectly.

The search clearly learns: with 10 playouts it finds a winning move in fewer than a quarter of positions, with 1,000 it finds one 95% of the time. Yet against a perfect opponent that is still not enough, because a single slip anywhere in the game hands the win away, and at 1,000 playouts it won only 4 games of 40. At 5,000 playouts per move it won all 40. The exclusive-or formula, meanwhile, plays perfectly from any position after one line of arithmetic. Search is powerful when nobody knows a formula; when someone does, the formula wins on speed and certainty.

Building a real game AI and testing it honestly gives Paignton teenagers a strong grounding for the AI era, reason enough to start coding in 2026. The longer argument is in why teenagers should still learn to code in 2026.