AI and programming classes in Harborne, Birmingham

The learner downloads OpenStreetMap data for south-west Birmingham and keeps the building outlines whose centre falls inside Harborne ward: 4,422 of them, with a median floor area of 101.2 square metres. Because areas range from a few square metres to thousands, the model works with the logarithm of the area. It is fitted with one, two, three, four and five components, and each fit is scored with the Bayesian information criterion, which rewards a close fit and penalises extra components.

Three components score lowest. The obvious guess is that they are garages, houses and big buildings. They are not. Of the 68 outlines tagged as garages or sheds, 66 sit in the middle component alongside the houses, because in this data the mapped garage blocks are of similar size to the houses. The small component is something else: a very narrow spike around 48.6 square metres holding 7.7% of the buildings. Looking at the data, 292 outlines measure between 45 and 52 square metres and most are tagged residential, which suggests rows of near-identical small footprints. The model found real structure, but not the structure a person would have named in advance.

Two more cautions came out of the fitting. The starting point matters: all 30 runs started from randomly chosen buildings reached the same answer after between 192 and 1,345 steps, but the library's default start stopped after 11 steps at a noticeably worse one, with no warning. And membership is soft: for 31.5% of buildings the model's most likely component has a probability under 0.8, so drawing hard lines between the groups would overstate what it knows.

Opening up each cluster before naming it is what keeps Harborne teenagers ahead of an AI's confident labels, and it is a habit learned by coding the model themselves. The longer argument is in why teenagers should still learn to code in 2026.