Vibe coding and AI agents classes in Kings Heath, Birmingham

The learner downloads OpenStreetMap data and keeps the 7,377 mapped elements that sit inside the ward. Each carries descriptive tags such as building=residential or natural=tree. Identifying tags, such as names, addresses and reference numbers, are thrown away, and the remaining 20,708 tags are shuffled into one long stream with 1,291 distinct values. Counted exactly, the commonest is building=residential with 3,330, then building=yes with 1,028 and natural=tree with 893.

The Misra-Gries algorithm reads the stream once with a fixed number of counters. A tag that already has a counter adds one. A new tag takes a free counter if there is one. If there is none, every counter drops by one and the new tag is forgotten. At the end, the surviving tags are the candidates for "heavy hitters", and each count is too low by at most the stream length divided by the number of counters.

With ten counters the algorithm kept only two of the true top ten, which is exactly what the guarantee allows: only a tag that appears more than 2,071 times is certain to survive, and just one does. With fifty counters, under 4% of the 1,291 an exact dictionary needs, all ten were there, and no count was off by more than 289. The errors always ran one way, too low, and always stayed inside the promised limit.

The Count-Min sketch makes the opposite mistake. It spreads counts across a small grid using four hash functions and answers a query with the smallest of four cells, so collisions can only push an estimate up. With 200 cells the worst overcount on any tag was 337; with 800 cells it was 41; with 4,000 it was 2. All three sizes still ranked the top ten correctly. One method undercounts, the other overcounts, and knowing which way the error leans is what makes either usable.

Teenagers in Kings Heath who can answer those questions are directing the AI, not following it, and that comes from writing the counters by hand first. The longer argument is in why teenagers should still learn to code in 2026.