AI and programming classes in Dumfries

Eskdalemuir's historic Met Office file supplies the data: for each month, the average of the daily maximum temperatures. The model is deliberately simple: to predict a month, use the average of that calendar month over a training window. The frozen model is trained once, on 1961 to 1990, and never updated. The rolling model is retrained every year on the thirty years just before. Both predict every month from 1991 to 2025, and the learner measures the bias, how far reality sits above or below the prediction on average, and how often the actual month came out warmer.

In its first decade the frozen model is nearly unbiased. By 2021 to 2025 it predicts 1.57 C too cold on average, and nine months in ten come out warmer than it expects; its typical miss has grown from 1.18 C to 1.69 C. The rolling model drifts far less, because it keeps absorbing recent years, but it still lags: in 2021 to 2025 it is 0.86 C too cold and three months in four beat it. Across the stations's record, the average monthly maximum was 10.77 C for 1961 to 1990 and 11.64 C for 1996 to 2025. A model is a snapshot of the period it learned from.

Knowing that models go stale keeps Dumfries teenagers alert to out-of-date AI answers, and learning to code is how that alertness becomes second nature in 2026. The longer argument is in why teenagers should still learn to code in 2026.