Coding classes in Shropshire

The puzzle is one of our own invention: 40 items, each with a weight and a value, and a bag that holds 286 units of weight. Which items should go in to make the bag as valuable as possible? A pack is written as 40 yes-or-no genes. The genetic algorithm starts with 60 random packs, then builds each new generation by picking parents in small contests where the better pack wins, cutting two parents and joining the pieces (crossover), and flipping the odd gene at random (mutation). Any pack that is too heavy scores nothing. Because the puzzle is small, dynamic programming can also find the true optimum, a value of 411, so we can measure precisely how close each method gets.

Evolution clearly beats guessing: 98.5 per cent of the optimum on average, against 91.6. Take away selection, so parents are chosen at random, and it falls to 94.9 per cent; the only help left is that we always carry the top-scoring pack forward. Take away mutation and it falls further, to 92.0, barely better than guessing, because the population soon runs out of new variation. These two results are Darwin's point in miniature: variation and selection together do the work.

The third result is the honest surprise. Removing crossover, the step that looks most like breeding, made no difference here; if anything the version without it did fractionally better. Crossover helps on some problems and not on others, and only an experiment tells you which. And note the last column: across 30 runs the full algorithm never found the optimal pack, while dynamic programming found it straight away. When an exact method exists and is fast enough, use it; evolution is for problems too big or too messy for exact methods.

So a young person in Shropshire should learn to code in 2026 to test the stories told about AI, not just repeat them. The longer argument is in why teenagers should still learn to code in 2026.