Online coding and Python classes in Hatfield

The English Wikipedia article on Hatfield, Hertfordshire keeps every saved version. We fetched three through its public interface: 27 December 2020, 28 September 2025 and 29 September 2026. The page quotes none of the article; the learner only counts. Split into lines, the 2025 and 2026 versions each have 274. Split into words, they have 4,594 and 4,601. The question is how to describe the difference between them in as few insertions and deletions as possible.

The first attempt most people make is to compare the texts position by position: line 1 with line 1, line 2 with line 2, and so on. On the one-year change that reports 263 of the 274 lines as different, although the two versions share 258 lines unchanged. One line added near the top shifts everything after it by one place, and a position check cannot see a shift. At word level the same check reports 4,550 differences, while 4,544 words in fact survive the year in order.

Myers treats the problem as a walk across a grid: moving right deletes a word from the old version, moving down inserts one from the new version, and a diagonal step is free wherever the two words match. The shortest edit script is the path with the fewest non-diagonal steps. His algorithm explores the grid by number of edits, one extra edit at a time, and stops as soon as a path reaches the corner. When two versions are close, that happens early. In Python the core is a loop over a dictionary of how far each diagonal has reached, about twenty lines in all.

For the close one-year pair, Python's built-in difflib agrees with Myers exactly. For the five-year pair it reports 14 more line edits and 146 more word edits than the shortest possible. The reason is a setting called autojunk. For long inputs, difflib ignores any item that is too common to be a useful anchor, and in the newer of the two texts that meant six words, including "the", "and", "of" and the name Hatfield itself. Switched off, difflib matches Myers on lines, 210 each, but is still 10 edits longer on words, because its method looks for long matching blocks first and never promised the shortest answer.

Every Myers count was checked against the slow method: a full table of every word against every word, 21,136,994 cells for the one-year pair. On this laptop that took 6.18 seconds. Myers took 0.005 seconds on the same pair, because only a handful of edits separated the versions. On the five-year pair, with 2,056 edits, Myers needed 2.1 seconds against 5.0 for the table, so the speed advantage shrinks as the difference grows, which is exactly what the O(ND) in the paper's title predicts.

A Hatfield teenager who has written a diff can review an assistant's work instead of trusting it, and that skill is built by learning to code. The longer argument is in why teenagers should still learn to code in 2026.