Two hundred and thirty times a minute

AI is fantastic. Genuinely, without irony. We have all been handed the ability to vibe-code every idea we have ever had in the shower, and we are using it enthusiastically. I have a folder on my machine called "projects" that would be more honestly named "the graveyard", and inside it sit at least four tools that do roughly the same thing, written within a month of each other, by me, for me, because on each of those four evenings I was completely certain that this time I had finally figured it out.

It turns out I am not remarkable in this. GitHub added 121 million new repositories in 2025, which works out to more than 230 new repositories every minute. Around 4.3 million of them are AI related, close to double what it was in 2023, and 36 million new developers joined in a single year, roughly one every second. That is not a trend. That is a weather system.

What nobody can tell you is how many of those repositories were written by AI rather than merely about it, and the reason is quietly funny. There is no field for it. Git records an author, and the author is whoever the tooling says it is, which is generally a human with a laptop and an unread newsletter subscription. So researchers have to go looking for litter instead. They count CLAUDE.md and AGENTS.md files, commit messages that politely credit an agent, and agent config files that somebody remembered to gitignore. One analysis using exactly those signals found that roughly a third of newly created repositories in its sample carried at least one of them, and was honest enough to call that a floor rather than a finding. The closest thing to a real number comes from a paper in Science that went through more than 30 million commits and put AI generated Python functions in US based repositories at around 29 percent by the end of 2024, which measures code and not projects, and is therefore not quite the question anyone was asking.

I find this wonderful in a bleak sort of way. We have built the most productive software factory in human history and we cannot report its output, because the machines do not sign their work and the humans have stopped mentioning it.

And mostly the weather is warm. More people building things is good. Lower barriers are good. I have no interest in standing in the doorway of a newly opened field, complaining that too many people are running around in it.

But then you look at what we are actually producing. GitClear went through 211 million changed lines of code and found that refactoring dropped from about a quarter of changed lines in 2021 to under ten percent by 2024, while copy-pasted code climbed steadily in the other direction. In 2024, for the first time in their dataset, cloning beat reshaping. Blocks of five or more duplicated lines went up eightfold in a single year.

You can read that as a code quality problem, and plenty of people have. I read it as a picture of behaviour. We are adding faster than we are consolidating. We are starting far more than we are joining. And I recognise myself in it completely, because joining is boring. Joining means reading somebody else's abstractions, working out why they made a decision I would not have made, opening an issue, waiting three weeks, and then being told no. Starting means an empty directory, a prompt, and forty minutes of uninterrupted dopamine. There is no contest.

Security tooling has this in a particularly concentrated form. The space is full of decent projects. Very few of them feel finished. Most of them are somebody's proof of concept that collected a few hundred stars and then went quiet, and the next person who needed that capability did not fork it or fix it, they built their own, because their mental model was cleaner. I say all of this as someone who has personally contributed to the silence.

The economists have caught up, which is usually a sign that something real is happening. A preprint from January 2026 with the almost impolite title "Vibe Coding Kills Open Source" argues that widespread AI-assisted development undermines the open source ecosystem's ability to sustain itself, even while it produces more software than ever before. The mechanism is the engagement loop. You used to download a package, read the documentation, hit a wall, ask about it somewhere public, and occasionally stay and fix the thing. That loop is where maintainers earned their reputation and where projects found their next humans. Now we ask the model instead. The wall is gone, and so is the reason to stick around. Maintainers get the pull requests without ever getting the people.

So the scarce resource is no longer code. It never really was, but now it is obviously not. What is scarce is people willing to link arms, to pick up something that is seventy percent done and carry it the rest of the way, which is the least glamorous stretch of any project and always has been.

I do not have a program for this, and I am suspicious of anyone who proposes that the answer is a new platform or a manifesto, since that would just be another repository. But there is one small thing I have started doing, more out of embarrassment than virtue. Before I start, I spend twenty minutes genuinely looking for the thing I am about to build. Not a lazy search, a real one. Usually it exists. Usually it is seventy percent done, slightly abandoned, and written in a style I find mildly irritating.

And then I have a choice that is honestly uncomfortable, because my version would be better. I know it would. Everybody knows it would. That certainty is worth being a little suspicious of. It is the most expensive feeling in software, and we can now act on it two hundred and thirty times a minute.