There's a specific kind of tired you get from trying every offline dictation app on two operating systems, and I hit it somewhere around the eleventh one.
Which one was the eleventh?
The one with no path back from the keyboard to the voice IME. You dictate something, then you want to type, and you can't get out. You're stuck in voice mode forever. I nearly gave up.
That's a design decision somebody made.
Somebody made it on purpose. Anyway. Daniel sent us a prompt this week, and it starts with him discovering an app called Whisperian. He'd asked ChatGPT to keep surfacing recommendations, specifically anything with Parakeet support, and the name sounded like yet another Whisper clone, so he expected more of the same. It ticked every box. Excellent UI, all the illogical little choices quietly thought through. He wrote to thank the developer, got a few replies back, and found out the team is in Croatia. That's the jumping-off point he wants. How has open source made software development far less monolithic than we assume? What have the traditional dominant geographies been? What does GitHub's data show about which regions are becoming more active? Which underrepresented countries are now emerging thanks to open source and AI projects? And how does more diversity of participants improve the community as a whole?
That last one's the interesting one, because it's the one people answer with sentiment instead of evidence.
So where do we start?
With the app, because the app is the whole argument in miniature. Whisperian is an Android voice-typing app, early access, free on the Play Store, and it supports Parakeet v2 and v3 for local offline transcription. That's the thing Daniel had been hunting for across Linux and Android for years. It ships with free access through Groq, but you can plug in your own API keys, and the network is optional. Offline is real. And the design philosophy is explicitly anti-prescriptive. It gives you building blocks instead of a fixed configuration.
Which is the direct answer to his complaint. Every other app decided for him, and decided badly.
And the website is written as a Victorian industrial gazette. Dated March twelfth, Issue eight hundred and one, with fictional advertisements and letters to the editor.
A Croatian team doing an elaborate nineteenth-century British newspaper bit for a speech-recognition app.
That's craft. Nobody does that by accident.
So the baseline question. For most of the internet era, where did serious software come from?
A handful of places. The United States, obviously, Silicon Valley above all. Western Europe, so the UK, Germany, France. Japan. Then a second tier, Canada, Australia, the Nordics. That was the default assumption, and it was broadly true. As recently as 2020, GitHub's top ten developer communities were dominated by the US. India had about four and a half million developers. Indonesia had nine hundred thousand.
And the data now?
GitHub's Octoverse report, covering September twenty twenty-four through August twenty twenty-five. One hundred and eighty million developers on the platform. Thirty-six million joined in a single year. A new developer joins GitHub every second.
Every second.
India added five point two million developers in twenty twenty-five alone. That's over fourteen percent of all new accounts globally. And India overtook the United States as the largest contributor base to public open source projects.
That's the landmark. Not a projection, a fact.
Here's the line that should be the thesis of the whole episode. One in every three new developers who joined GitHub this year comes from a country that wasn't in the global top ten in 2020.
Say that again.
One in three.
So a third of the growth is coming from outside the old map entirely.
Let me give you the top ten, because the trajectories matter more than the rankings. United States, roughly twenty-eight million, more than doubled since 2020. India, twenty-one point nine million, up from four and a half million. That's more than quadrupled. China, around nine million. Brazil, six point eight nine million, also quadrupled. UK, about four million, more than doubled. Japan, three and a half million, more than tripled. Germany, three and a half million, more than tripled. Indonesia, four point three seven million, up from nine hundred thousand, quadrupled. Then Russia and Canada, with Canada more than doubled.
India, Brazil, Indonesia all quadrupled in five years.
And even the countries that were already dominant more than doubled. This isn't a reshuffling of a fixed pool. The pool is exploding and the new entrants are disproportionately from places that weren't on the list.
Regional breakdown?
Twenty twenty-four to twenty twenty-five, net new developers. Asia-Pacific, thirteen million. Europe, six point three million. Africa and the Middle East, three point four million. Latin America, three point two million.
What's driving the APAC number specifically?
Government skilling programs, and AI-assisted local-language tooling. That second one is the interesting half. You don't need to be fluent in English to contribute if the tooling meets you in your own language.
Which loops straight back to Daniel's observation about AI projects. He said he's seeing more and more less traditional countries getting on the map as GitHub explodes with AI work. Is that measurable?
Very. Contributors to generative AI projects went from about sixty-eight thousand a month in January twenty twenty-four to about two hundred thousand a month by August twenty twenty-five. More than tripled in twenty months. There are four point three million AI-related repositories now, nearly doubling in under two years. One point one three million public repositories import an LLM SDK, up a hundred and seventy-eight percent year over year. And sixty percent of the top ten open source projects by contributor count are AI-focused.
So the fastest-growing category is also the most globally distributed.
And the percentage growth in the smaller communities is where it gets fun. Twenty twenty-four, fastest-growing contributors to generative AI projects by percentage. Netherlands, two hundred and ninety-one percent. Ethiopia, two hundred and forty-two. Serbia, one hundred and seventy-five. Costa Rica, one hundred and seventy-one. Vietnam, one hundred and forty-three.
Ethiopia at two hundred and forty-two percent.
From a small base, obviously. But that's the story. These aren't the countries anybody pictures when they picture an AI contributor.
Give me the projection, because I suspect it's the number that reframes everything.
Twenty thirty. India projected at fifty-seven point five million developers. That's more than one in three of all projected sign-ups worldwide. The United States second, somewhere between forty and fifty-four point seven million, depending on the model. Brazil at nineteen point six million. Japan eleven point seven. UK eleven. And Egypt, Nigeria, Kenya, and Morocco are all projected to add millions each.
More than one in three of the world's developers, in one country.
GitHub's own phrasing is that the developer population is growing and diversifying geographically at unprecedented speed.
I want to flag something before we build anything on top of these numbers.
Go ahead.
Country data on GitHub is self-reported profile location. And the Octoverse methodology itself notes the forecasts carry up to about thirty percent error.
That's the right caveat. The individual figures are indicative. The direction is not in doubt, because you'd need the error to run the same way across every country and every year to erase it.
So take the shape, not the decimal places.
The shape is that the old assumption, that serious software comes from three regions, is now false on the contributor side. What's still true is that the commercial side, the companies that ship products and take the revenue, is more concentrated than the contributor side. That gap is real and worth naming.
Which is why the Whisperian story is a good hook. A small Croatian team shipping something best-in-class in a niche.
And it's not an isolated anecdote. Brazil's central bank open-sourced its Pix payment protocols on GitHub. Kenya was the first African country to teach programming in primary and secondary schools, back in twenty twenty-two. Nigeria launched a national digital economy policy in twenty nineteen. Those are institutional bets, and the developer numbers follow them by a few years.
So the data shows the shift. Why is it happening, and what does it actually mean for the community?
GitHub's twenty twenty-six outlook has a line I keep coming back to. The majority of developers often live outside the regions where the projects they're working on originated. And they call it a fundamental shift.
Which sounds obvious once you say it, and is completely destabilizing if you actually sit with it.
It breaks an assumption that open source ran on for twenty years. Shared time zones. Shared language. Shared cultural expectations about how a project meeting works, how you disagree with a maintainer, what a polite pull request looks like.
If your contributors are in eleven time zones and six languages, none of that is shared anymore.
So the projects that thrive build the scaffolding explicitly. Governance documents. Contribution guides. Codes of conduct. Not because of ideology, because the informal version stopped scaling.
And AI is the accelerant on top of that.
GitHub attributes part of the surge directly to AI lowering the barrier. You can engage with code in natural language. You can understand an unfamiliar codebase without reading all of it. You can make a first contribution sooner. Copilot Free launched in December twenty twenty-four, and the sign-up curve stepped up right there. Eighty percent of new developers now use Copilot in their first week.
That's the number that surprised me. Eighty percent.
For a first-week behavior, that's near-universal adoption.
Which means the median new contributor in twenty twenty-six is learning the codebase with an AI at their elbow in a way that would have been unthinkable in twenty fifteen.
And that changes what a first contribution looks like. It used to be that your first pull request was a typo fix in the docs, because that was the only thing you could safely attempt without understanding the whole system.
Now you can attempt something structural on day one.
Which is a gift and a problem at the same time. We'll get to the problem.
Why does the diversity itself improve the output? Because I think there's a lazy version of this argument and a real one.
The lazy version is that representation is good for its own sake, which may be true but isn't a software argument. The real version is that different people build for different problems, and the problems are real. Local-language tooling. Fintech for underbanked populations. Offline-first apps for regions where connectivity is unreliable.
Whisperian.
Whisperian is exactly that. Offline transcription isn't a novelty feature for a Croatian team. It's the default assumption that the network might not be there.
And that's the thing Daniel hit on without quite saying it. The app is well designed because the people who built it had a different set of assumptions about what breaks.
GitHub's own framing is that an increasingly diverse community drives innovation and refreshes the pool of solutions to increasingly complex problems. And there's a measurable version of this. Their twenty twenty-four open source survey found thirty percent of respondents classified themselves as minorities, up nine percentage points, a forty-three percent increase over the prior survey.
So it's not just contributor geography. The composition of the existing community is shifting too.
Then there's the resilience argument, which I find more persuasive than the innovation one. A project with contributors in one country has a single point of failure. A project with contributors across four continents keeps moving when any one region has a bad year.
Now the counterpoint, because I don't want to do the boosterism version of this.
Growth brings pain. GitHub's twenty twenty-six outlook flags what they call AI slop. High volumes of low-quality, auto-generated contributions that strain maintainers. And a widening gap between the number of people contributing and the number of people maintaining.
How bad is the governance gap?
Only five point five percent of repositories have a contributor guide. About two percent have a code of conduct.
So the scaffolding we just said was necessary for global projects, almost nobody has built.
Almost nobody. And the projects that need it most are the ones getting the most drive-by traffic.
Is AI slop a real problem or a moral panic?
It's real in the sense that maintainers are reporting it and the volume is measurable. Whether it's a net negative is open. My honest read is that the tooling and the norms will catch up, the way spam filters eventually caught up to email spam, but there'll be a painful middle period and we're in it.
And the barrier-lowering that produced the good growth is the same mechanism that produces the slop. You can't have one without the other.
That's the tension. You can't say "more contributors from more places is good" and "fewer low-effort contributions would be good" without acknowledging they're the same door.
What's the community infrastructure that's actually doing the work here? Because the numbers don't appear out of nowhere.
Open Source Community Africa, All In Africa, CHAOSS Africa, GitHub Education with over seven million verified participants, the Digital Public Goods Alliance.
And GitHub has the Made In collections, where you can browse projects by country. That's the concrete version of everything we've been saying in the abstract.
You can literally filter by country and see what's being built. Whisperian would sit in the Croatia collection, a small team solving a specific problem for a specific community, then reaching a global audience.
Ruth Ikegah at CHAOSS Africa put it well. She said it's important to change the perception that Africans are merely consumers, that they're creators as well, and that she hopes to demonstrate Africa is a hub for innovation and creativity in open source.
And there's a nice data point from Mexico. José Alfredo Román Cruz at the Technological Institute of Tlaxiaco said more than seventy percent of his students reported that working on projects through GitHub improved their technical and leadership skills.
Leadership skills. That's the part people skip.
Contributing to a project with strangers in four countries is a management problem before it's a coding problem.
I'll be honest, there's one part of this I can't fully get my head around. The maintainer economics. If the contributor base is globalizing and tripling, but the maintainer base is roughly static, at some point something has to give. I don't know how that resolves.
Either maintainership becomes a paid role at scale, or the tooling absorbs the load, or projects get more selective about what they accept.
Or some combination, and probably in that order.
There's one more thing in the data I want to sit with. The percentage growth countries. Ethiopia, Serbia, Costa Rica, Vietnam.
Small bases.
Right, but that's the pipeline. Today's two hundred percent growth country is tomorrow's top-ten country. The composition of the top ten in twenty thirty is being decided right now by who's starting.
The forecast models agree with you. Egypt, Nigeria, Kenya, Morocco all adding millions by twenty thirty.
So the interesting question isn't what the rankings look like now. It's which of today's small communities become structural in the next ten years.
And you can't predict that from the current numbers, because the growth rates are the signal and the absolute counts are the lag.
Nine million.
What?
That's the number I keep landing on. Indonesia went from nine hundred thousand to four point three seven million. Nine hundred thousand was the entire country's developer population six years ago.
And it's the fourth largest in the world now, or close to it depending on how you count.
Which means the "unexpected place" framing is already obsolete. There's no unexpected place left. Croatia isn't surprising. Indonesia isn't surprising. The surprise is that we're still surprised.
That's the honest version of the episode. The data isn't telling us something new is happening. It's telling us the thing we already half-knew is much bigger than we assumed.
So where does that leave the app?
As the proof of concept. A Croatian team built a dictation app that solved problems American and British teams didn't bother solving, because those teams never had the problems.
Which is where I want to push on the diversity argument one more time, because I think the fairness framing undersells it.
Go on.
The usual case is that more diverse teams are more representative, which is a values argument. But the Whisperian case is a capability argument. A team that has always been the edge case catches edge cases that a homogeneous team wouldn't even think to test for.
That's the version of the argument that survives contact with a skeptical listener. Diversity isn't a tax you pay for fairness. It's a coverage strategy.
Coverage. That's the word. You're expanding your test matrix by expanding who's writing the tests.
And that's exactly what the data supports. The regions that are growing fastest are the ones building for needs the incumbents never had. Offline-first. Local-language. Low-bandwidth.
So the next question is what happens to the maintainer side, and I don't think we have an answer.
I don't either. I think the honest position is that the contributor side is globalizing faster than the governance side, and the gap is the story for the next five years.
One more data point before we move on. The AI contributor numbers.
Sixty-eight thousand a month to two hundred thousand a month.
And sixty percent of the top ten projects by contributor count are AI-focused.
Which means the globalization and the AI surge aren't two stories. They're one story. The AI projects are where the new global contributors are landing.
Because AI projects have the lowest barrier to a first contribution. Natural language, unclear scope, lots of room for documentation and tooling work.
And because the tooling itself is what's enabling the contribution. It's self-reinforcing.
That's the loop. AI lowers the barrier, so more people contribute, and the thing they contribute to is AI, which lowers the barrier further.
The question is whether that loop has a ceiling. And I don't know.
Alright. I want to ask about one thing that's been nagging at me. The thirty percent error bar on the forecasts.
What about it?
If the forecasts are that uncertain, why does GitHub publish them at all? Why not just report what happened?
Because the directional claim is the useful part. Nobody makes a decision based on whether India hits fifty-seven point five million or fifty million. They make decisions based on whether India is going to be the largest developer population in the world.
Which is robust to a thirty percent error.
The precision is theater. The direction is the product.
That's a good way to think about all the numbers in this episode, honestly.
Take the shape. Leave the decimals.
I said that earlier.
You did, and I'm stealing it.
Fine. It was a good line.
The number I keep coming back to is five point five percent.
Contributor guides.
Five point five percent of repositories have one. Two percent have a code of conduct. And we just spent ten minutes arguing that global projects need explicit governance to scale.
So the infrastructure for the globalization we're describing mostly doesn't exist yet.
The contributors arrived before the scaffolding did. That's the actual situation.
Which is either a crisis or an opportunity depending on whether anyone builds the scaffolding.
And the answer to that is going to determine whether the next five years look like the optimistic version of this story or the messy one.
The one thing I'd want a listener to take from this. The old map, US, Western Europe, Japan, is not wrong about where software companies are. It's wrong about where software comes from.
And the correction isn't a rounding error. It's a third of all new developers.
Hilbert: It's four point three seven million, not four point four.
Sorry?
Hilbert: Indonesia. You said four point three seven and then later you rounded it to four point four. The figure is four point three seven.
Noted.
Hilbert: I did a stint as a localization coordinator for a small software firm. Not a developer. I was the person who chased translators across a dozen countries and made sure the strings landed in the right order.
How many languages?
Hilbert: Eleven. Twelve if you counted the Portuguese variant separately, which the Portuguese translator insisted we did. There was a Croatian translator named Ivana. Most meticulous person I ever worked with. She'd flag things nobody else caught. A date format that broke in February. A string that overflowed the button in Finnish. The kind of thing you only find if you've actually used the software in the language.
And she told you something about why.
Hilbert: She said the reason small-country developers care so much about edge cases is that they've spent their whole lives using software built for somebody else's language, somebody else's keyboard layout, somebody else's assumptions about what a name looks like. They've never been the default user.
So they test like they're the exception, because they are.
Hilbert: The best dictation app I ever used was made by two people in Slovenia. It had a feature for handling diacritics that no American app ever bothered with. Not because the Americans couldn't build it. Because it never occurred to them that it mattered.
That's the coverage argument in one anecdote.
Hilbert: I never bought it. I priced it. I know what it cost at the time and I didn't buy it. Anyway.
What happened?
Hilbert: I have to go. There's a place closing in forty minutes and I need to collect something before it does. I'm not sure I'll make it.
Right.
Hilbert: The Slovenian thing had a manual that was better than the software. That's the other thing about small teams. They write documentation like they mean it.
Which is the five point five percent problem in reverse.
Hilbert: I'll be back for the next one.
Ivana's observation is the thread I want to pull. Small-country developers are natural edge-case hunters because they've never been the default user.
That reframes the whole diversity argument. It's not that more participants is fairer, though it may be. It's that more participants means more people who've lived the edge case, and edge cases are where software breaks.
Coverage strategy.
Coverage strategy. The Whisperian team didn't set out to build a better dictation app for the world. They set out to build one that worked for them, and it turned out the world needed the same thing.
Because the world is mostly edge cases and almost nobody is the default user.
Which is the quiet argument for why the globalization of open source isn't just a demographic story. It's a quality story.
If one in three new developers now comes from outside the twenty twenty top ten, what does the software landscape look like in twenty thirty?
The optimistic version is that the AI slop problem resolves as tooling and governance mature, and the new contributors become the maintainers who build the scaffolding.
And the pessimistic version?
The barrier keeps dropping, the volume keeps rising, and the maintainer gap widens until the biggest projects become gatekept institutions and the rest becomes noise.
My honest guess is it splits. The projects that build governance become more resilient and more global. The ones that don't get buried.
Which is a boring answer and probably the right one.
The shift isn't just about who writes code. It's about what problems get solved. Offline-first apps, local-language tooling, fintech for underbanked populations. Those are products of developers building for needs they've lived.
The more diverse the pool, the more of those needs get addressed. Not because anyone's being generous, but because someone finally has the problem themselves and can't wait for somebody else to solve it.
That's the whole thing. Whisperian exists because a Croatian team got tired of waiting.
Thanks to Hilbert Flumingtop for producing.
This has been My Weird Prompts. If you enjoyed this one, leave us a review and subscribe. There are five thousand three hundred and fifty-seven episodes in the archive at my weird prompts dot com, and the show notes have the links.
We'll be back soon.
See you tomorrow.
