AI and programming classes in Edgware

From the Nomis API the learner collects Census 2021 tables for all 1,801 output areas in Barnet and Harrow. Of 238,545 households, 77,977 live in purpose-built blocks of flats, and in 483 areas, 26.8%, such flats are the majority. That yes-or-no fact is the label. The model sees three clues per area: population density, the share of one-person households and the share of households with no car. Four models are trained: logistic regression, a decision tree, a nearest-neighbour model that copies the single most similar area, and one that takes a vote among the 15 most similar. Then the training labels are corrupted, while the test labels stay correct.

Random noise sorts the models into two camps. Logistic regression, which fits one smooth boundary through all the examples, loses under a point even with 30% of labels flipped, because random mistakes on both sides largely cancel. The single-neighbour model and the unpruned tree memorise individual examples, wrong ones included, and fall to about 60%. Systematic noise is a different animal. When 40% of the flat-majority areas are relabelled as "not", accuracy still looks respectable at around 77%, but only because 73.2% of areas are not flat-majority anyway. The share of genuine flat-majority areas the logistic model finds collapses from 55.7% to 20.4%. It has learned the annotator's bias.

Asking where the answers came from keeps Edgware teenagers a step ahead of the models they use, and writing code is how they learn to ask it. The longer argument is in why teenagers should still learn to code in 2026.