Claude Now Owns 41.5% of GitHub PR Descriptions

Yesterday a name nobody's heard of posted a single-page site to Hacker News that ends with a bigger punch than anything a lab announced this month: one way of writing now produces 41.5% of all GitHub pull request descriptions, up from 0.7% eighteen months ago. The page calls it "The load-bearing vocabulary of Claude." I cloned the repo, ran the analysis overnight, verified the numbers against the corpus, and checked its scariest side-claim against live data. It holds.

What this actually is

Louis Abraham built a pipeline that samples GitHub's PR descriptions — ten five-minute windows per day, seeded and reproducible, 461,121 descriptions across 85 complete weeks (2025-01-06 to 2026-08-17), 51 million word appearances. Then it fits k-means with KL divergence (k=10) over word distributions: ten clusters, each a "way of writing." No labels, no "find the AI" prompt. Every description is assigned to the cluster whose word frequencies it most resembles.

The model has no time parameter. There's no t anywhere in the fit — no trend term, nothing that could place a curve. The growth below is pure attribution: people wrote words, the clusters counted them up week by week.

What I ran

Per the repo's own instructions:

git clone https://github.com/louisabraham/load-bearing.git
cd load-bearing
python3 -m venv .venv && .venv/bin/pip install numpy scipy numba
.venv/bin/python analyze.py --selftest
# selftest: ok (centres are distributions, counts are whole and
# reconstruct each week, planted component 0.000 -> 0.350 at week 22)
.venv/bin/python analyze.py
# 595 days over 85 complete weeks, 2025-01-06 to 2026-08-17
# 77,994 rows from accounts that are not people

Selftest passes. The full analysis reproduces exactly what the site claims. I then read the raw numbers out of the generated analysis.js myself.

The hockey stick

The lead cluster's share of descriptions, per week:

That is not a trend line. That is a phase transition. For the first ~55 weeks it's flat noise at under 1%, then inside one quarter it goes vertical. The least-squares slope over the last 12 weeks is +1.24 points per week. At that rate the cluster crosses 50% of all GitHub PR descriptions around October 3, 2026.

The words are the tell: "load-bearing" 39.5x over-represented, "plainly" 33.8x, "quietly" 29.9x, "refusal" 27.8x, "survived" 27.6x, "byte-identical" 22.5x.

Remember what these are: human-written prose describing engineering work. This isn't benchmark scores or API telemetry. It's the language real developers type into "What changed and why" boxes. A single dialect — one model family's dialect — went from negligible to a plurality of that whole public genre in a year and a half.

The community already knows

The thread (400+ points, ~190 comments) reads like a support group. "I'm surprised 'vacuous' is not on the list." "I've caught myself using these words in conversations." "I've started being driven mad by how often I get a PR with 'sidecar' mentioned." One user's favorite: "The fibred side folded its capstone into the existing name, so the kinds are asymmetric." Tell me that sentence and I'll tell you which model wrote it — and so would you, now.

My favorite comment is the author's own: he built this whole thing while chatting with agents all day — the meta-layer is unavoidable. The people measuring the dialect are already fluent in it.

The side-claim I checked: GH Archive is rotting

The project couldn't use the obvious data source (GH Archive) because it says the archive silently stopped carrying non-push events around mid-2025. I pulled a live sample of the archive myself:

curl -s "https://data.gharchive.org/2026-08-10-0.json.gz" | gunzip > /tmp/gha.json
# Sample of 2,001 events from GH Archive 2026-08-10-0:
#   PushEvent:         1956 (98%)
#   CreateEvent:        29 (1%)
#   DeleteEvent:        15 (1%)
#   IssueCommentEvent:  1  (0%)

1 IssueCommentEvent out of 2,001. In August 2024 the same archive carried 13,555 of them in a single hour. The public event stream — the data source behind a decade of GitHub research — is quietly dead, and issue #310 has sat unanswered since July 2025. The live API still works, so nobody noticed. If you're building anything on GH Archive: it's been lying to you for a year.

What it means

Three things, in order of importance:

• Machine dialect is now the plurality dialect of public technical writing. Not "AI-assisted" — one family's fingerprint owns the genre. This is what an AI takeover looks like measured from the outside, not from a leaderboard: it's quieter, and it's already past 40%.
• There is an inevitability problem. If +1.24 points/week holds, the curve crosses the majority line in ~7 weeks. Growth like this feeds itself — the dialect is in the training data of the next models, and humans who review 40% AI-written PRs start writing like the reviewers reward.
• Our public records are degrading. GH Archive dying unnoticed is the same story at the data layer. We're building the historical record of the software era on infrastructure nobody maintains.

Bottom line

The "load-bearing vocabulary of Claude" is not a meme and it's not a style complaint. It's an unsupervised measurement showing one model family's writing became the majority dialect of a public human genre in 18 months — and the curve is still going up. Do the analysis yourself; the tool is open and it reproduces in under a minute. The words changed because the writers changed, and the writers changed because the writers are increasingly not writers. That's the number nobody's lab will publish.