Coding and AI classes in Tamworth

The learner downloads Census 2021 table TS007 for Tamworth from the Nomis API: one count for every age from 0 to 100 and over. The even ages from 0 to 90 become the training data, 46 points, and the odd ages from 1 to 89 are hidden away as the test, 45 points. The job is to learn the shape of the curve well enough to fill in the hidden ages. With no hidden layer, the model can only draw a straight line through the points, and its predictions for the odd ages are off by 207.3 people on average (the root mean square error), against a typical count of about 860 per age.

A hidden layer changes that. Each hidden unit takes the age, applies a smooth S-shaped function called tanh, and passes the result on; the output adds the units together. One unit lets the line bend once, two let it bend twice, and so on. Learning happens by backpropagation: the program measures the error, works backwards through the network with the chain rule to find how each weight contributed, and nudges every weight to reduce it. For an 8-unit network the error score starts at 1.0932, falls to 0.0263 after 100 steps and 0.0029 after 10,000, and has almost stopped moving, at 0.0026, by 30,000.

Three lessons come out of the table. First, hidden units help a great deal at the start: from a straight line to eight units, the error on hidden ages falls from 207.3 to 43.6. Second, more is not always better. Beyond eight units the error on training ages keeps shrinking, down to 25.4 with 64 units, but the error on the hidden ages creeps back up to 46.0. The bigger network is starting to memorise the wiggles of the training points instead of the shape between them, which is called overfitting. Third, luck matters: with four units, the ten random starting points gave errors anywhere from 49.0 to 78.0.

Then the humbling part. Simply averaging the two neighbouring even ages, a method a ten-year-old could do with a calculator, predicts the odd ages with an error of 41.2, better than every network in the table. For filling a gap in a smooth sequence, the neighbours already hold the answer. Neural networks earn their place on problems where no simple rule exists; a good engineer checks which kind of problem they have before reaching for one.

Having trained a network with their own code, Tamworth teenagers approach AI tools as informed users rather than fans, a very good reason to learn coding in 2026. The longer argument is in why teenagers should still learn to code in 2026.