Coding and AI classes in Rotherham

The learner downloads the census accommodation table and turns the eight counts into probabilities: 49.20 percent semi-detached, 22.31 detached, 18.16 terraced, 8.74 in purpose-built flats, and small shares for converted buildings, homes in commercial buildings and caravans, which are just 0.04 percent. A language model does something very similar with words: it gives every possible next word a probability. The temperature formula takes the logarithm of each probability, divides it by a number T, and turns the results back into probabilities with the softmax function. At T equal to 1 nothing changes. The learner then tries lower and higher values.

At temperature 0 the generator becomes greedy and every household is semi-detached, the most common answer, which is exactly why chatbots at low temperature repeat themselves. At 0.5 the big categories grow and the small ones vanish: of 1,000 sampled households, 735 are semi-detached and not one is a caravan or a converted building. At 2 the distribution flattens, and caravans jump from 0.04 to 0.9 percent, more than twenty times their real share; in 1,000 samples there are 7 of them where the census predicts almost none. At 5 nearly one household in twenty is a caravan. Higher temperature means more variety, and also more answers that are simply unrealistic.

The learner finally tries top-k sampling, a common safeguard: keep only the three most likely categories and share the probability among them. Semi-detached becomes 54.9 percent, detached 24.9 and terraced 20.3, and nothing unusual can ever appear, not even the real flats. Every setting is a trade-off between faithful, boring and surprising, and the learner writes down which one they would choose for a data generator and which for a story-writing tool, and why.

A Rotherham teenager who understands why AI answers wander will use these tools with more judgement than most adults, which makes coding well worth learning in 2026. The longer argument is in why teenagers should still learn to code in 2026.