AI and programming classes in Gravesend

The learner downloads one Census 2021 table from Nomis: for each of Gravesham's 327 output areas, the number of households and the number with no car or van. Across the borough that is 8,199 of 41,733 households, or 19.65%. But the areas differ a great deal, from 0.8% to 62.4%, with a standard deviation of 13.15 percentage points. If every household in the borough simply had the same 19.65% chance, luck alone would produce a spread of about 3.59 points. The gap between 3.59 and 13.15 is called overdispersion, and it says the areas are really different, not just noisy.

A beta-binomial model describes this with two unknowns: a typical share, and a number for how alike the areas are. The sampler starts from a deliberately bad guess, a typical share of 50%. At each step it nudges both unknowns by a small random amount, works out how well the new pair explains all 327 areas, and accepts the move always if it is better and sometimes if it is worse. It runs 20,000 steps and the first 5,000, the burn-in, are thrown away. The only thing the learner changes between runs is the size of the nudge.

A high acceptance rate sounds good and is not. With tiny steps nearly every move is accepted, but the sampler shuffles along so slowly that its first 200 guesses still average 46.0% and 15,000 steps carry the information of only 34 independent draws. With huge steps almost every proposal lands somewhere absurd and is refused, so the chain sits still. The middle setting does the job: a typical share of 19.38%, with 95% of the kept draws between 18.02% and 20.88%, and a model that predicts a spread of 12.55 points between areas, close to the 13.15 observed. The sampler has measured how unlike one another Gravesham's neighbourhoods are, and it has also shown the learner how easily a badly tuned one reports nonsense with a straight face.

Gravesend teenagers who have written a sampler line by line can tell a result from a rumour, and that is a skill which grows out of coding, not out of asking a chatbot. The longer argument is in why teenagers should still learn to code in 2026.