AI and programming classes in Earlsdon, Coventry

The learner requests a 40 by 40 grid of heights from the OpenTopoData service, which serves the Copernicus EU-DEM elevation model. Across the box the ground runs from 70.1 m to 110.0 m. A random handful of the 1,600 points is kept as "measurements" and the rest are hidden. Two methods then predict the hidden heights. Inverse distance weighting is the simple one: average the known points, counting near ones more. The Gaussian process instead learns from the sample how quickly height tends to change with distance, and uses that both to predict and to say how far off it might be. Each experiment is repeated for ten different random samples.

With only 25 points the two methods are close, and both beat simply guessing the average height, which would be out by 8.43 m. As points are added the Gaussian process pulls away: at 200 points its typical error is 1.85 m, less than half that of inverse distance weighting. Its ranges can be checked too. A range labelled 95% should contain the truth 95 times in 100; here it managed 89.4% to 92.7%, so the model is a little too sure of itself. Yet the ranges clearly carry information. With 100 known points, the fifth of predictions the model was least sure about were wrong by 3.36 m on average, and the fifth it was most sure about by 1.13 m. It knows roughly where it is guessing, even if it slightly understates by how much.

An Earlsdon learner who has tested a model's error bars expects the same honesty from every AI tool, and knows how to check for it in code. The longer argument is in why teenagers should still learn to code in 2026.