Coding and AI classes in Clifton

The learner downloads a 40 by 40 grid of ground heights from the open EU-DEM elevation model for a rectangle of 3,468 by 3,316 metres around Clifton. The lowest cell is 5.0 m and the highest 116.8 m. A simulated walker is dropped somewhere unknown and walks in 25 metre steps. The compass is wrong by about ten degrees, the step counter by about five metres, and the altimeter by about two. The particle filter starts with 2,000 guesses spread over the whole rectangle.

After ten steps the filter is still half a kilometre out: many places share a height, and the cloud of guesses has not yet collapsed. By thirty steps it is within 38 m, and from sixty steps it holds at 28 m. Counting steps from a known start does the opposite. It begins accurate and drifts, reaching 64 m after three kilometres, because each small compass and stride error is added to the last and nothing corrects them. In 95% of the 40 walks the filter finished within 100 m of the truth.

Then the learner tries to break it. A worse altimeter, wrong by five metres, still ends at a median of 36 m, with 87.5% of walks inside 100 m. Cutting the guesses hurts far more: with 500 particles the median stays at 29 m but only 67.5% of walks finish within 100 m, and with 100 particles the median error is 379 m and just 40% succeed. With too few guesses, none happens to start close to the truth, and resampling then multiplies wrong ones.

Weighing guesses against evidence is a coding skill as much as a thinking one, which is why Clifton teenagers gain from writing real programs in 2026. The longer argument is in why teenagers should still learn to code in 2026.