Coding and AI classes in Moseley, Birmingham

The learner downloads OpenStreetMap data for south Birmingham and keeps every shop, café, restaurant, takeaway, pub and bar whose position falls inside Moseley ward: 114 places. Their positions are converted to metres and k-means is run for every number of clusters from two to ten, twenty times each from different starts. For each result the program records the silhouette score and also the total squared distance of points from their cluster centres, the quantity k-means itself tries to shrink.

The silhouette score peaks at two clusters, with 0.775. But look at what those two clusters are: 107 places in one and 7 in the other. The score is rewarding a split that says "nearly everything, plus a handful far away", which is true and tells a planner almost nothing. A result with two sizeable groups, of 61 and 35 places, appears only at six clusters, and the score drops to 0.507 when it does, because those two groups sit close together. Meanwhile the squared distance falls every single time a cluster is added, from 21.19 to 1.34, so it cannot choose a number at all.

One more check keeps the scores honest. The same number of points scattered at random across the same rectangle, with no structure whatever, scored between 0.38 and 0.43. So 0.5 is not far above meaningless, and a high score mainly signals that a few points are a long way from the rest. The right number of clusters depends on what the clusters are for, and no single score knows that.

Looking inside the winner, not just at its score, is what keeps Moseley teenagers in charge of AI analysis, and they learn it by building the analysis themselves. The longer argument is in why teenagers should still learn to code in 2026.