AI and programming classes in Burton upon Trent

From the Nomis API the learner downloads population density (table TS006) and car availability (TS045) for all 384 output areas in East Staffordshire, then uses the ONS lookup to see which built-up area each belongs to. Areas inside Burton upon Trent or Uttoxeter are labelled town, 273 of them; the other 111, in villages or open countryside, are labelled not town. After putting both features on the same scale, scikit-learn fits a linear support vector machine at several values of C, the setting that decides how heavily mistakes are punished.

The first lesson is about baseline accuracy. Because most areas are in a town, a program that never looks at the data scores 71.1%, so the model's 81.8% is an improvement of about ten points, not a triumph. The second is in the support vectors. In a cleanly separated problem only a few points hold up the boundary; here 175 of 384 do, a sign that towns and villages overlap heavily in these two numbers. Lowering C softens the margin, lets more points inside it and costs a little accuracy. The third lesson is that the car feature barely matters: density alone reaches 81.2%.

Every one of Stretton's areas, and nearly all of Tutbury's and Barton's, lands on the town side. At the scale of a few hundred homes, the streets of a large village are about as crowded as the streets of a town. What the model has really learned is "built-up or open land", which it gets right in all 34 countryside areas. "Town or village" is a distinction the ONS makes by the size of the whole settlement, and nothing in the two features describes that.

Checking a model like that is how Burton teenagers stay in charge of AI tools rather than trusting a single accuracy number, and a strong reason to learn to code in 2026. The longer argument is in why teenagers should still learn to code in 2026.