Coding and AI classes in Stanmore

The map is OpenStreetMap's walking network over Stanmore: 2,943 junctions and 203.7 km of streets and paths. The expert is simply the true shortest route. From 300 expert journeys the learner collects 48,373 examples of the form "at this junction, heading for that goal, was this street the expert's choice?", described by five numbers such as how closely the street points at the goal. A gradient boosting model learns from them. It is then tested on 200 journeys it has never seen, typically 35 junctions long.

Matching the expert 88.1% of the time sounds impressive, but a 35-junction journey needs about 35 correct choices in a row, and the first wrong turn often leads into a loop: the copier walks up a cul-de-sac, turns round, and makes the same mistake again, because nothing in what it sees tells it that it has been there before. Longer journeys fail far more often than short ones. The textbook cure, asking the expert to label the odd places the copier actually ends up, did not help here; success stayed between 6.0% and 7.5%, because the missing ingredient was not more examples but information. Giving the same copier a memory, so that it never re-enters a junction and backs out when stuck, lifted completion to 97.0%. Its routes were then a median 1.39 times the shortest: it arrives, but not elegantly.

Testing the whole task rather than one step is how Stanmore teenagers keep long AI jobs honest, and they learn it by building one. The longer argument is in why teenagers should still learn to code in 2026.