Vibe coding and AI agents classes in Bishopston

From OpenStreetMap, through the Overpass service, the learner loads the drivable roads in a rectangle around Bishopston: 102.0 km in 754 mapped ways, cut into 4,218 short segments. Python then simulates 60 trips of between 800 and 2,500 metres along the shortest route between two random junctions and records a position every 40 metres, 2,718 positions in all. Random error is added to each one to imitate a GPS receiver. Because the trips are simulated, the true road for every position is known and both methods can be marked.

With a good signal there is little to choose: 91.8% against 93.3%. At 15 metres of error the gap opens to seven points, 76.0% against 83.3%, because the closest road to a blurred dot is often a side street the trip never entered, and only a method that looks at the dots before and after can tell. At 30 metres Viterbi still leads, 65.2% against 57.4%, yet a third of positions are wrong under either method. No algorithm recovers information the signal never contained.

The method has two ingredients. One score says how likely a dot is given a candidate road, which falls as the distance grows. The other says how likely a move from one candidate to the next is, by comparing the distance along the roads with the straight distance between the two dots. Viterbi keeps, for each candidate at each step, only the most probable way of arriving there, so the work grows with the length of the trip and not with the number of possible routes.

Bishopston teenagers who can program a test with known answers are able to mark an AI's work themselves, and that is a reason to write code in 2026. The longer argument is in why teenagers should still learn to code in 2026.