The most honest statement a platform can make is often the one where it fails to explain itself. In its annual report, GitHub officially admitted: AI is driving interest in software development, but this doesn't explain the surge in the data, and they will keep investigating. The numbers: 100 million users in early 2023, 180 million in fall 2025, 225 million by July 2026. Meanwhile, the number of people who write code for a living is estimated at 36.5 million. Even if every developer had three accounts, the math doesn't add up. The extra cohort doesn't write code. So what are they doing on GitHub?
Five Case Studies
A life manual. HowToLifeBetter, created September 7 this year, hit 27,000 stars in one month—by this blog's own check on October 6, it had 46,554 stars and climbing. Its content covers longevity and disease prevention, first aid, saving money, legal red lines, health insurance, and relationships. Every item lists cost, benefit, evidence grade, and original source, citing only journal papers. It can be installed directly as a skill into your Codex or Claude.
A poetry library. chinese-poetry, 53,560 stars, founded in 2016. Tang poems, Song lyrics, and Yuan verses are all inside; it's the underlying data source for many poetry apps on the market. It contains no code in the "modern" sense.
A prompt list. awesome-chatgpt-prompts, now renamed prompts.chat, 172,196 stars. The largest open-source prompt library in the world, cited by courses at Harvard and Columbia. Again, not a single line of executable code.
A lobster. OpenClaw, 391,516 stars. A personal AI assistant framework with a lobster mascot, created in late November last year.
A string of departed coworkers. colleague.skill (同事.skill). The README's first line: A coworker left, leaving behind a mountain of unmaintained docs? An intern departed, leaving only an empty desk and a half-finished project? Distill their chat logs, documents, and code commits into a skill and load it into your agent—the person leaves, the experience stays. Launched in late March, it broke 10,000 stars within days; verified at 25,338. A derivative universe followed: boss.skill, ex.skill, self.skill, immortality.skill, Nüwa.skill (33,682 stars—distill anyone's way of thinking), and even Crush.skill (363 stars—teach your agent to analyze your crush).
None of these five are traditional open-source projects. In the React era, GitHub was a programmer's workshop. Now it's something else.
What colleague.skill Is Really Mining
This project deserves a closer look, because it mines deeper than it appears. The README's scenario is explicit: a departing coworker, unmaintained docs, an intern's half-finished work. This isn't a new problem—organizational memory loss has been studied for decades in management research, and the solutions have always been documentation, knowledge bases, and handover checklists. All of them depend on the departing person's willingness to cooperate, and someone leaving has no incentive to cooperate.
Skill-ification flips the direction: don't ask for cooperation—distill existing traces instead. Chat logs, code commits, meeting notes are all already there; AI turns them into reusable mental models. The person never left a manual, but the traces are enough to draw one. When ten thousand people star this project, they aren't starring the code—they're starring "I wish I could keep that person too."
This is also the first time the carrier of personal experience has become a mass consumer product. Turning a person's experience into something machine-executable is called an agent memory system in the lab, skill encapsulation in engineering—and in the hands of workers, it's one sentence: put your departing colleague on a USB drive and take them with you.
The Most Counterintuitive Money Is Actually the Most Rational
The most striking number in the source material: Cursor was acquired by rocket company SpaceX for $60 billion, officially announced June 16, 2026, in an all-stock deal that Forbes called the largest venture-backed startup acquisition ever. This blog verified it through multiple sources.
Why would a rocket maker buy an IDE? Because SpaceX judged the battlefield isn't in the sky. The acquisition coverage named Anthropic and OpenAI as the competitors—coding agents are their thickest revenue stream. Cursor's annualized revenue grew from $100 million to $2 billion in 13 months; buying it means buying a ready-made entry point into the "let agents write code" channel. $60 billion against $2 billion annualized revenue is a 30x multiple—a bet that the channel itself will keep widening.
The numbers on the non-professional side are steeper. Lovable targets people who don't code at all; officially, 80% of its users have no technical background. Its valuation went from $1.8 billion last July, to $6.6 billion in December, to $13.3 billion this August, with $500 million in annualized revenue. The source said "from 1.8 to 6.6 billion in 8 months"; the verified timeline is 8 months from launch to $1.8 billion, then 5 more months to $6.6 billion—the numbers are all correct, just compressed.
What Stars Are Hoarding
Put the three explanations side by side. GitHub's explanation is that AI lowered the barrier. The source author's explanation is a "high-resolution era": in the past, a need had to be shared by tens of millions before it was worth building into a product; now one person's need can become a project, and the star count is a headcount vote of everyone with the same need worldwide. This blog's verdict: all three are right, but at different layers—AI lowering barriers is the fuel, individual needs becoming projects is the demand-side change, and what actually makes all of this settle is the star mechanism itself.
Stars are currently the cheapest validation signal. A repo with 20,000 stars means 20,000 real accounts voted with their actions. It's imperfect—star manipulation has happened and GitHub has acted on it—but consensus is hard to fake: behind chinese-poetry's 50,000 stars are real downstream apps; behind prompts.chat's 170,000 are real classroom citations. This signal used to only price code; now it's starting to price "life manuals," "departed coworkers," and "classical poetry." HowToLifeBetter's stars rose from 27,000 in the source material to 46,000+ in this blog's check—nearly doubling in a month. The validation is still in progress.
Looking back at GitHub's official "we can't explain it": they can't explain it because they still assume they're a code hosting platform. That cohort of non-coders among the 225 million accounts isn't hoarding code—they're hoarding "credible, free, immediately usable things." The trust mechanism the open-source world spent fifteen years building—traceable, verifiable, reusable—is being loaded with something else. The author of HowToLifeBetter tags every life tip with an evidence grade and journal source; the density of that practice on GitHub is higher than on any lifestyle platform.
It's not that GitHub went mainstream. It's that the open-source methodology did.