Two numbers, and what we do with them.

What gets counted

The GitHub contribution calendar for the last five years. That means commits, pull requests, issues and code reviews — not commits alone. It is the same graph that shows up on a GitHub profile, just plotted against AI release dates instead of by itself.

Anything that appears on the leaderboard is built from public activity only. If you sign in, your own page can include private contributions so the picture is complete for you — but that version is never ranked or compared with anyone. A number nobody else's figure includes has no business sitting next to theirs.

The split

November 2022, the month ChatGPT was released. The same date for everybody — no per-person judgement about when someone “started using AI”, because there is no honest way to know that from a contribution graph.

The calculation

Everything starts from the two averages. They are what the pages lead with, because they need no explaining: 86 → 300 a month says more than any ratio of it can.

Two figures come off that pair.

The change is what the board ranks by and what the badges are cut from. The pace is the same movement as a multiplier, and it is the figure on your own page, sitting next to the two averages it came from.

Averages per month rather than totals, because the two stretches are not the same length — and the “after” side gets one month longer every month. Comparing totals would mostly reward whoever has been alive longer inside the AI era, which is not a fact about anyone.

The +15 added to both sides is the part worth explaining, so here it is in full.

Why we add 15

Divide by a “before” average close to zero and you get a number close to infinity. That is not a hypothetical: one real developer on this list came out at ×13,810 against a baseline barely above zero — six hundredths of a contribution a month. They had hardly used GitHub before 2022 and used it normally afterwards. The multiplier says nothing about them; it is an artefact of dividing by almost nothing.

There were two ways out. Exclude anyone whose baseline is too thin — which means withholding real people, with real contributions, from a board about their own contributions. Or fix the arithmetic. We fixed the arithmetic: adding 15 to both months gives a thin baseline something solid to divide by, while barely moving anyone who already had a real history. That developer now reads ×58; dhh went from ×3.5 to ×3.1, matz from ×2.9 to ×2.5. Nobody had to be removed.

It matters less than it used to, now that the board ranks the change rather than the ratio — a thin baseline no longer buys a high position. The smoothing keeps the multiplier on your own page from reading as nonsense.

One case still has no figure at all: a developer with no pre-ChatGPT months in the window. There is no “before” to compare against, and inventing the comparison would be worse than declining it. Those graphs show activity and no comparison.

Who is on the list

It started as about a hundred hand-written profiles: people known for building AI, with a note on what each is known for. Then it grew to roughly three thousand more, assembled by follower count and pruned by hand — obvious bots, mirrors and long-dead accounts taken out.

Worth being straight about: that pruning was a manual pass, not an automatic filter. Nothing here detects an account that mostly automates its own commits. If you spot something that clearly is not a person, tell us and it comes out.

You do not have to be on any list to appear. Look yourself up and you are measured on the spot and kept — no follower threshold, no volume threshold, no approval.

How the ranking works

One requirement to be ranked: enough history that a “before” exists. That is genuinely all. There is no minimum follower count and no minimum number of contributions.

Order is by the change in monthly contributions, then by total contributions where two people tie. Not by the multiplier, and that is a deliberate correction. A ratio rewards a small “before” rather than a large change: of the top fifty by multiplier, twenty-seven had a pre-AI baseline under five contributions a month, and one sat at zero. A developer who went from 1,009 → 7,307 a month ranked 74th, while someone who went from 0.1 → 863 ranked 5th.

The badges follow the same figure. They used to be cut from the multiplier while the board ordered by the change, which let someone sit high on the board wearing nothing and someone mid-table wear AI Supercharged. The bands are in contributions per month too — +150 and up for the loudest badge, +40 for the next, +8 for a modest one — and The Veteran reads the pre-AI pace instead, since “already shipping before AI” is a claim about the past rather than about failing to qualify for anything else.

Measurements go stale on purpose: a rotation re-measures the list every couple of hours, and anyone not measured in the last thirty days drops off the board rather than sitting there with an old number.

Why we show both figures

Because neither one is the whole answer, and the honest thing is to admit it.

The change is the fairer thing to rank. It is in units nobody has to convert, and it cannot be gamed by having been quiet beforehand. But it does favour people already working at volume: adding two hundred contributions a month is easier from four hundred than from four.

The pace is the fairer thing to compare against yourself. It measures you against your own past rather than against Linus Torvalds, which is the only comparison most people actually want. But on a ranking it collapses — a thin baseline beats a real career.

The interesting claim about AI is that it flattens the act of writing code: that the gap between a famous engineer and everyone else narrows when a model does the typing. Testing that needs both readings — how much moved, and how much it moved relative to where you started. So the board ranks one and your page shows the other, and neither pretends to be the verdict.

What this cannot tell you

• Volume is not value. A hundred commits can be one good refactor or a hundred typo fixes. The graph cannot see the difference, and neither can we.
• Public work only. Anyone whose real work sits behind a company firewall looks quiet here, however much they ship.
• Not just code. The calendar counts issues and reviews too, so a number can rise on a year of good review work without a line being written.
• November 2022 was not a controlled experiment. A new job, a funded startup, a burnout year, or simply moving work onto GitHub all land on the same side of the split as AI does. The number shows that something changed. It cannot show that AI is why.
• The window moves. We fetch five years of history, so the “before” side is a shrinking slice of time that gets a month shorter with every month that passes.