AI and programming classes in Greenock

The learner requests 100 heights from the OpenTopoData service, which serves the European EU-DEM elevation model at 25 m resolution, along a line 3,316 m long running south from near the waterfront. Heights rise from 5.4 m to 275.4 m. Fifteen points are picked at random for training; the other 85 test the model. Python fits polynomials of degree 1, 3, 5, 9 and 14 by ordinary least squares, then again with ridge regression, which adds a penalty on large coefficients. The whole experiment is repeated for 50 different random choices of the 15 points.

As the degree rises, the training error falls towards zero: a degree-14 polynomial passes exactly through all 15 points. The error on unseen points does the opposite. At degree 9 it is 84.5 m, and at degree 14 the typical miss is over 23 km in height, because the curve swings to absurd values between training points. Adding a small ridge penalty to the same degree-14 model brings the unseen error down to 14.1 m, better than any of the plain fits, while its training error rises to 9.5 m. It fits the training points less tightly and the hillside far better.

The strength of the penalty matters. On this data a very weak penalty left a typical unseen error of 19.8 m and a much stronger one 16.8 m. In a real project the strength is chosen by cross-validation on the training points alone; we tried several on the test points only to show how sensitive it is.

Knowing why a perfect fit can be a bad model keeps Greenock teenagers in charge of the AI they build, which makes learning to code well worth it in 2026. The longer argument is in why teenagers should still learn to code in 2026.