Coding and AI classes in South Shields

The learner fetches three Census 2021 tables from the Nomis API for every output area in South Tyneside: car availability, household size and population density. Across the 68,305 households, 22,366 have no car or van, but the share swings widely from one small area to the next. The task is to hide one area's no-car share and estimate it from the others. Each area's key is two numbers, its population density and its share of one-person households; the hidden area's own two numbers form the query; each other area's no-car share is its value.

Attention then works in three steps. Score each key by how close it is to the query. Turn the scores into weights that add up to one using the softmax function, with a sharpness setting that decides whether the weight piles onto a few areas or spreads across many. Finally, take the weighted average of the values. Repeat for all 543 areas and measure the typical error.

Neither extreme wins. Copying the single most similar area, known as hard attention, is noisy because one area can be an oddity. Weighting all 542 equally ignores the query entirely and is worst of all. Soft attention across a few dozen similar areas comes out ahead, cutting the error by about 30% compared with equal weights. The keys matter too. Using density alone as the key gives an error of 12.56 points; the one-person share alone gives 9.87; both together give 9.38 at the same sharpness. Which features a model compares decides what it can find.

Understanding the machinery lets South Shields teenagers use AI with judgement instead of faith, and that alone makes learning to code worthwhile in 2026. The longer argument is in why teenagers should still learn to code in 2026.