Louis Abraham · 2026-08-27 · notable
The load-bearing vocabulary of Claude — 461,121 pull requests, clustered
Louis Abraham clustered 461,121 GitHub pull request descriptions from January 2025 to August 2026 by word choice alone. One of the ten writing styles went from 0.7% of the corpus to 39%, and it is the one that says 'load-bearing'.
Cluster GitHub pull requests by vocabulary alone and one writing style climbs from 0.7% to 39% in eighteen months.
Key specs
| Pull requests analysed | 461,121 |
|---|---|
| Cluster share by mid 2026 | 39% |
What is it?
Louis Abraham sampled GitHub pull request descriptions for 603 days and sorted 461,121 of them into ten writing styles based purely on which words they use. Nothing in the model knows about dates or about AI. One style still grew from 0.7% of the corpus at the start of 2025 to 39% by mid-2026, and it is currently adding about 1.2 percentage points a week.
How does it work?
The method is k-means clustering using Kullback-Leibler divergence over word distributions instead of squared distance, which suits counts better than plain k-means. A word joins the 19,798-word vocabulary only once 50 distinct accounts have written it, which filters bots and copy-paste. Sampling is ten five-minute windows a day through GitHub's search API, seeded by date so the whole corpus rebuilds from its dates alone.
Why does it matter?
Arguments about how much code is now AI-written usually run on vibes. Abraham's chart is a measurement instead: a style nobody wrote in early 2025 now carries most new pull request text, and it is identifiable from vocabulary alone. For reviewers it is a reminder that 'load-bearing', 'quietly' and 'seam' are tells rather than idioms — and that a growth line still climbing weekly has not levelled off.
Who is it for?
code reviewers and people arguing about AI writing tells
Try it
Open louisabraham.github.io/load-bearing/ and pick a cluster to see its characteristic words.