Coding classes in Lancashire

The learner builds a grid of 48 rows by 96 lamps, 4,608 in all, and paints on it a smooth picture of our own that runs from dim on the left to bright on the right. Averaged over the whole grid the picture is 36.65 per cent bright, so a faithful version should light about 1,689 lamps. Three methods then decide which lamps go on. The simplest rounds each lamp to the nearer of on or off. The second compares each lamp with a fixed repeating four-by-four pattern, the Bayer ordered method. The third, Floyd-Steinberg error diffusion, rounds a lamp and then hands its rounding error to the four neighbours still to be decided, in shares of 7, 5, 3 and 1 sixteenths.

The surprise sits in the third column. Judged lamp by lamp, simple rounding wins easily. Yet it lights only 1,008 lamps where about 1,689 were needed, and in our run every lamp in the left half of the picture stays dark: 48 of the 96 columns vanish. The last column measures what a person standing back actually sees, by blurring each lamp with its neighbours over a five-by-five patch before comparing. On that measure Floyd-Steinberg is about ten times closer than rounding, and it lights 1,679 lamps, within ten of the target.

The learner then gives each lamp four brightness levels instead of two, as a dimmer would. Everything improves, but the order holds: seen from a distance, rounding scores 0.077, the Bayer pattern 0.016 and Floyd-Steinberg 0.009. The real lesson is not about lights. Whichever error you choose to measure decides which method wins, so the first job is to measure the error that matters to the person looking.

So a young person in Lancashire should learn to code in 2026 to be the one who chooses what an AI is asked to get right. The longer argument is in why teenagers should still learn to code in 2026.