The CEOs Who Coded
On a Tuesday in early December 2025, Sebastian Siemiatkowski, the CEO of Klarna, walked into a meeting with something unusual under his arm — not a slide deck, but a working prototype he'd built himself the night before. It had taken him twenty minutes. "I no longer have to beg someone more technically proficient to create a product," he said. "I can just build it." [Read more →]
He wasn't alone. Across Silicon Valley, in boardrooms in Stockholm and São Paulo and Singapore, a peculiar thing was happening. Founders who hadn't touched a code editor in years were suddenly shipping demos. Executives were building internal dashboards before the weekend. Non-technical product managers were deploying web apps before lunch. The Wall Street Journal called it the "vibe coding revolution." Collins English Dictionary named it their Word of the Year.
The cause, everyone agreed, was November 2025. Two AI models had arrived in quick succession — Anthropic's Claude Opus 4.5 on November 24th, and OpenAI's GPT-5.2-Codex just weeks later. Developer and writer Simon Willison captured the feeling exactly: "It genuinely feels to me like GPT-5.2 and Opus 4.5 in November represent an inflection point — one of those moments where the models get incrementally better in a way that tips across an invisible capability line where suddenly a whole bunch of much harder coding problems open up." [Simon Willison →]
The anecdotes were everywhere. A quarter of Y Combinator's Winter 2025 batch had codebases that were 95% AI-generated. [YC Winter 2025 →] GitHub itself was reporting that a new developer joined the platform every second — 36 million in a single year. [GitHub Octoverse →] By early 2026, 63% of active vibe coding users weren't developers at all. The story told itself.
But stories, however compelling, are not data. And this one had an obvious empirical test: if the revolution was real — if lapsed coders were genuinely returning to their keyboards, and dormant GitHub accounts were suddenly lighting up with commits — you should be able to see it in the code.
So we looked.
How to Measure a Comeback
Testing whether a specific event triggered a behavioral change in millions of developers is harder than it sounds. The obvious approach — check whether commit activity went up after November 2025 — immediately runs into a measurement trap. If you sample developers who were active just before November, you've already selected people near their activity peak. Their subsequent decline looks dramatic, but much of it is just regression to the mean. [Methodology →]
To avoid this, we ran a carefully designed two-stage sampling study using the GitHub REST API. We sampled 1,762 developers across four separate collection batches, comparing their commit activity in two equal three-month windows: before November 2025 (August–October) and after (November 2025–January 2026). The core design question was: what fraction of developers showed a meaningful increase?
We also ran a crucial debiasing step: a second, smaller sample drawn from developers active in March–April 2025 — six months before the hypothesized break — to check whether the main sample's overlap with the pre-window was distorting the results. And we ran a placebo sweep, asking: if we moved our hypothetical break month to July, or August, or October — months when nothing especially dramatic happened in AI — would November still stand out as exceptional?
Six Percent
The answer came back with unusual consistency across every configuration we tried.
In the main 1,600-user sample, exactly 96 developers — 6.0% — met the bar for a meaningful increase. In the smaller, carefully debiased 162-user sample drawn from an earlier, non-overlapping seed period, the number was 10 out of 162 — 6.2%. The two figures are so close they could be the same measurement taken twice. The 95% confidence intervals overlap substantially. This is not noise. This is a stable signal.
And that signal says: no broad surge.
If there had been a genuine wave of returning coders, we'd expect something like 20%, 30%, perhaps 40% of developers to show a marked uptick. Campaigns that meaningfully change population-level behavior — public health interventions, major platform launches, viral trends — show up unmistakably in the data. Six percent is not a wave. It's barely a ripple.
The dormant returners: a rounding error
Most striking of all was the cohort we were specifically hunting for: the dormant returners. These are developers who had effectively gone quiet before November — minimal activity in the pre-window — and then suddenly reactivated. The vibe coding narrative predicted hundreds of them. We found ten.
Just ten. Out of 1,600 sampled users. That's 0.6%.
The picture that emerges is not a homecoming parade. It's more like a slow exhale. The largest cohort — 851 developers, 53% of the sample — were declining: meaningfully less active after November than before. A quarter landed in "other" (roughly flat). Only 15% were consistently active throughout. The returnees were statistical ghosts.
The humans behind the six percent
Who are the 6% who actually did increase? Browsing through the individual cases reveals something interesting: they're not the lapsed professionals who stopped coding and picked it back up. They're developers who started new projects — or dramatically intensified work on existing ones. Here are a few representative examples from the sample.
Notice what these cases have in common. These aren't dormant developers returning. They're people who were already committing code before November — and then accelerated. The most evocative case may be rjwalters/vibesql: someone who named their project after the very movement we're studying, and whose commit pace nearly sextupled after October. The irony is almost too neat.
And on the other side — the developers who slowed down or stopped entirely — the picture is equally instructive.
The sheer scale of decline in the top cases reflects a pattern common to GitHub: many repos function as automated data stores — receiving thousands of commits from scripts or bots rather than human keystrokes. When those scripts stop running, the commit counts collapse to zero. This is why looking at the share of developers who changed is more meaningful than the raw numbers: it's the human pattern we're trying to detect, and raw totals are swamped by automation.
November Was Actually the Weakest Month
Here is where the data gets genuinely strange — and genuinely informative.
We ran the placebo test. We asked: if we defined our "break month" as July 2025 instead of November, what fraction of developers would look like they'd increased their activity? What about August? September?
The results are remarkable.
In July 2025 — months before the big AI model releases — 13.9% of developers looked like they were ramping up. By August, still 12.8%. The share steadily eroded as we moved toward November. By the time we reach November itself, only 2.2% looked like they were accelerating. December: 1.7%.
This is the opposite of what a true November surge would produce. If the Opus 4.5 release had genuinely brought developers back to their keyboards, November should show an uptick in the increase share — not the nadir.
The aggregate looks fine — until you look underneath
There's one more thing worth noting. If you plot raw monthly commit totals across the sample, November actually looks decent — activity holds steady at levels last seen in July and August, before a pullback in December. For a moment, you might think: maybe there is something here.
But this aggregate hides a severe concentration problem. In our sample, the top 10 most active users generated 83.8% of all monthly commits. Ten people out of 1,600. The "stable November" isn't a broad trend — it's the behavior of a handful of prolific contributors whose patterns happen to hold up, while the typical developer was quieting down.
This is why looking at medians matters. The median developer's commit count dropped dramatically after October. The mean is buoyed by outliers. The typical developer was not having a coding renaissance. They were having a coding winter.
The Revolution Is Real — Just Not Here
At this point, you might reasonably ask: are we just calling the vibe coding revolution fake? Is this all just hype?
Not exactly. The revolution is real. But it's playing out somewhere other than the public GitHub commit logs of established developers.
The GitHub Octoverse 2025 report tells a fascinating adjacent story: 36 million new developers joined GitHub in a single year — more than one per second. [GitHub Octoverse →] Commits on GitHub grew 25% year-over-year. Pull request merges grew 23%. These are genuine, enormous numbers.
But these are new developers — people arriving on GitHub for the first time, people who never had a GitHub account before AI made coding accessible. They're the Klarna CEO building prototypes. They're the Y Combinator founders shipping 95% AI-generated codebases. They are, in many cases, people for whom GitHub is a new destination, not a place they're returning to.
That distinction matters enormously. Our study could only observe developers who already had a GitHub presence — people seeded from prior public commits. We deliberately filtered to users with pre-window activity. If the vibe coding wave was primarily recruiting new entrants rather than reactivating dormant ones, our study would not catch it by design. And the cohort data suggests that's precisely what's happening: just 8 "newly active" users in our 1,600-person sample, because our sampling method couldn't capture the truly new arrivals.
Think of it this way. The vibe coding narrative has two versions. Version one: lapsed developers who used to commit code but stopped are now coming back. Version two: people who never really coded before are starting for the first time. The GitHub Octoverse data strongly supports version two. Our analysis cannot rule out version one, but finds no evidence for it in existing users' commit patterns.
Why the "returner" story was so easy to believe
Andrej Karpathy coined "vibe coding" in February 2025 — a term that, by his own description, meant "giving in to the vibes, embracing exponentials, and forgetting that the code even exists." [Karpathy →] The term wasn't describing dormant experts returning. It was describing a new mode of engagement with software — one that didn't require traditional programming expertise. Collins Dictionary named it Word of the Year. Merriam-Webster listed it as slang trending.
But the anecdotes that followed — the CEO who built a prototype, the founder who shipped a demo — were overwhelmingly about people who had never been active GitHub users. They were new entrants, not returnees. The narrative of the "returning developer" was a seductive misread of what was actually a story about lowering the entry barrier, not raising the return rate.
The sample comparison, sortable
Below is the full cross-batch comparison. Click any column header to sort. The "early-seeded" group — seeded from March–April 2025, six months before the break — is the least susceptible to sampling bias and should be weighted most heavily.
| Sample Group | Users | Meaningful Increase | 95% CI (increase) | Meaningful Decrease | Median Pre | Median Post |
|---|---|---|---|---|---|---|
| Main batch | 900 | 5.7% | 4.3–7.4% | 47.1% | 36 | 1 |
| Page-2 batch | 700 | 6.4% | 4.8–8.5% | 50.1% | 34 | 2 |
| Combined overlap-seeded | 1,600 | 6.0% | 4.9–7.3% | 48.4% | 35 | 1.5 |
| Early-seeded (debiased) | 162 | 6.2% | 3.4–11.0% | 32.1% | 31.5 | 7.5 |
Anecdotes Are Not Trends. But Trends Don't Explain Anecdotes.
Here's what the data can and cannot say.
It can say, with considerable confidence, that there was no broad wave of dormant GitHub users reactivating their accounts after November 2025. The surge was not in the commit logs. November was not a structural break in the activity patterns of existing developers. The specific claim — that lapsed coders returned to GitHub in droves — is not data-backed.
It cannot say that nothing happened. Something clearly did. GitHub's own data shows the platform growing faster than ever, with new developers arriving at a rate unmatched in its history. AI-assisted coding became genuinely easier in November 2025. Founders really were building prototypes in twenty minutes. That is all real.
The distinction is subtle but important. The vibe coding revolution appears to be a revolution of new entrants, not returning exiles. It's lowering the drawbridge, not raising the dead. The GitHub commit logs of people who were already using GitHub don't show the revolution because those aren't the people having it.
1. No broad post-November surge in public GitHub commit activity among sampled developers.
2. Only ~6% of existing developers showed a meaningful increase — stable across sampling designs.
3. November 2025 does not stand out as a positive structural break; it was the weakest candidate month in the placebo test.
4. Aggregate commit totals are dominated by <1% of users — raw totals mislead.
5. The vibe coding revolution most likely reflects new entrants to GitHub, not reactivation of dormant users.
Malcolm Gladwell once observed that the most dangerous kind of story is the one that feels truest. The returning developer narrative felt true because it matched something real: AI coding genuinely got dramatically better. But "better tools" and "lapsed developers returning to GitHub" are not the same claim. One is a fact about models. The other is a claim about human behavior. And human behavior, it turns out, is stubborn. The developers who had stepped back from GitHub didn't suddenly flood back. What changed, perhaps, is who's arriving for the first time.
That's a different story — and in some ways a more interesting one. The era of coding as the exclusive province of professional software engineers may genuinely be ending. But you won't find the evidence for that in the commit patterns of people who were already on GitHub. You'll find it in the 36 million new accounts. In the CEO who walked into that meeting with a working prototype. In the Word of the Year that didn't even exist two years ago.
The revolution is real. The returners are ghosts.
What This Means — For Developers, Researchers, and the Industry
Platform-level growth (36M new users) and account-level reactivation are different claims. Verify which one your sources are actually measuring before reporting either.
The meaningful-increase rate was stable at ~6% across multiple sampling designs. Any future study that finds a much higher rate should scrutinize sampling overlap with the pre-window.
Of all commits in our panel came from just 10 users. Aggregate metrics about developer activity can be wildly misleading — insist on medians and percentile breakdowns, not means.
The vibe coding revolution is best understood as a story about new entrants to coding, not returning ones. Your growth is at the top of the funnel, not the re-engagement layer.
The November 2025 inflection is genuinely real for individual capability — models got better. But better models ≠ more commits from existing developers. Don't confuse tool quality with user behavior.
The strongest next test: compare new-account commit rates pre and post November 2025. That's where the revolution should show up, and it's currently unmeasured by this study.
How This Study Was Done
We used the GitHub REST API with an authenticated token to sample developers from public commit activity. Users were seeded via /search/commits queries across defined date windows, then counted using the /repos/{owner}/{repo}/commits endpoint with per_page=1 and Link header parsing — a technique that infers commit counts efficiently without downloading full commit histories.
Four batches were collected: two "overlap-seeded" batches (seed period May–October 2025, overlapping the pre-window) totaling 1,600 users, and two "early-seeded" debiased batches (seed period March–April 2025, non-overlapping) totaling 162 users. Pre-window: August 1–October 31, 2025; post-window: November 1, 2025–January 30, 2026.
Change-point detection used the ruptures Python library with PELT algorithm. Confidence intervals use the Wilson method. The placebo sweep tested candidate break months from July to December 2025. Full replication code is in scripts/github_activity_analysis.py.
Scope limits: Public commits on default branches only, for users seeded from searchable commits. Does not capture private repositories, non-default branches, or developers who had never previously made a public commit. The 36 million new GitHub joiners are not captured by design.