The Selfie Problem: Vibe Coding Wins Even if Experts Hate It

Some of the most important software being built right now is being built by people who can’t code. A biologist piecing together a carbon sequestration monitoring pipeline. A grid engineer prototyping a flexibility marketplace. A product person — me — who spent twenty years shipping other people’s ideas and finally had a tool to think out loud in.

I opened Cursor for the first time about six months ago with a specific kind of dread. Not the fear that it wouldn’t work. The fear that it would. What if this was great? What if I’d spent two decades accumulating judgment about software products and could now bypass the one thing that had always required me to hire someone else? What did that mean for the someone elses?

It was great. But in a subtler way than I expected. It didn’t make me a developer. What it did was remove a specific kind of friction — wrong syntax, misread documentation, the slow crawl of test-and-iterate — and replace it with something closer to thinking. I could observe, critique, redirect. I caught some things a real developer would have caught faster. I missed others that a real developer would never have missed at all. But I wasn’t doing it for production. I wasn’t doing it for exhibitions. I was doing it the way someone takes a selfie — personally, for the thinking, for the making, not for the output.

That frame matters. Because we’ve been here before.

The disgust is always about access

Every technology that hands creative power to ordinary people goes through the same ritual. First comes the disgust.

In 2013, when Oxford Dictionary named “selfie” its word of the year, critics denounced it as evidence of a narcissistic subculture.¹ In the Telegraph, Harry Wallop insisted it “perfectly captured the self-regard of our age.”² When the selfie stick took off in 2014, the world drew a collective heave of disgust.³ Serious photographers rolled their eyes. And they were right, in the narrow sense that most selfies are bad photographs. They were wrong about what that meant.

Go back further. In the late nineteenth century, photography was a complex and burdensome process accessible only to professionals.⁴ George Eastman built the Brownie and collapsed that barrier with a slogan — “You press the button, we do the rest” — that professional photographers found every bit as threatening as it sounds. What followed wasn’t a degradation of the craft. It was a transformation of what photography was for.

Or take desktop publishing. Before Aldus PageMaker arrived in 1985, producing professional printed materials required access to expensive typesetting equipment and specialist training, limiting publication to large corporations and dedicated typesetting houses.⁵ Designers were appalled by what non-professionals did with the new tools — too many fonts, no kerning, layouts that violated every principle of the craft. Desktop publishing resulted in mass layoffs of unionised typographers and paste-up artists, often replacing professionals with decades of experience with people who’d had weeks of training on PageMaker.⁶ The disruption was real and it was uneven. It also permanently expanded who could publish, what could be published, and what design itself meant.

The critics of vibe coding are making the same correct observation the typographers made: the output is often bad. They’re drawing the same wrong conclusion.

What the numbers actually show

The term “vibe coding” didn’t exist until February 2025, when Andrej Karpathy coined it⁷ to describe building software entirely through natural language prompts — describing what you want, letting AI generate the code, never looking too hard at what’s underneath. The uptake has been fast and a bit disorienting. 63% of vibe coding users are now non-developers.⁸ A quarter of Y Combinator’s Winter 2025 batch had codebases that were 95% AI-generated, and 44% of non-technical founders are now building their initial prototypes with AI coding tools rather than hiring developers.⁷

The scepticism is grounded. 45% of AI-generated code contains security vulnerabilities and 40% has exploitable bugs.⁹ Only 33% of developers trust AI code accuracy.¹⁰ The clearest case study in what goes wrong is Moltbook, which launched in January 2026 with its founder proudly stating he hadn’t written a single line of code. Three days later, security researchers found 1.5 million API authentication tokens exposed in client-side JavaScript.⁹ These are not edge cases. They are the growing pains of a technology that has arrived before the norms around it have.

And they map almost exactly onto what happened in photography. Film cameras peaked in 1999 when consumers took around 80 billion photos worldwide.¹¹ Between 2010 and 2019, the global population took over 8.6 trillion — seven times more than the previous decade.¹² Today the number runs to roughly 1.72 trillion photos annually, at around 47,500 per second.¹³ Most are banal. Many are terrible. The point is what became possible at the edges, in the hands of people who had something to say and suddenly had the means to say it visually.

The professional industry didn’t collapse — it stratified. Reduced barriers to entry intensified competition and disrupted salaried roles, but work that required genuine skill, taste, and presence became more specialised.¹⁴ New categories emerged entirely: drone photography, computational image-making, content creation as a standalone profession. The floor dropped. The ceiling rose.

But the gate isn’t open to everyone

Here is where the photography parallel starts to strain, and where honesty matters.

The Brownie was a dollar. A smartphone is several hundred. The internet required a fixed line, then a mobile network, neither of which arrived evenly. Every democratisation in the history of technology has had a geography: it democratised within the places that already had the infrastructure to receive it. The IMF warns that AI could exacerbate cross-country income inequality, with growth impacts in advanced economies potentially more than double those in low-income countries.¹⁵ 60% of the world’s top supercomputers sit in the United States, China, and Germany. There are no supercomputers in East Africa.¹⁶

Vibe coding runs on compute. Compute concentrates. Token-based pricing in commercial AI services already creates disproportionately higher costs for speakers of underrepresented languages, because current tokenisation schemes systematically favour English and other high-resource languages — with efficiency variations of up to thirteen-fold.¹⁷ If you’re a biologist in Lagos with a brilliant idea about mangrove carbon monitoring, the tools theoretically exist. The infrastructure to run them affordably and reliably may not.

This matters not just as an equity concern — though it is one — but because it shapes the argument about democratisation itself. Vibe coding is democratising access to software creation within already-connected, already-resourced contexts. That is genuinely significant. It is not the same as universal access. The selfie spread because the smartphone spread, and the smartphone spread because mobile networks covered populations that fixed broadband never reached. Whether AI compute follows a similar trajectory — cheap, distributed, available — is not guaranteed. It is a political and infrastructure question as much as a technical one.

The selfie wasn’t the problem. It was the prototype.

What critics missed about the selfie — and what critics of vibe coding are missing now — is that new creative forms emerge from the bottom of the capability curve, not the top. The selfie looked like vanity. Depending on who held the phone, it was also a tool for documenting illness, for protest, for identity assertion by people whose images had previously been controlled or suppressed. The denunciation of the selfie as narcissistic subculture triggered a cultural response — people adopted it precisely because it had been identified as impertinent.¹ The form was claimed by the people it was built for.

The software equivalent is already forming. Not the toy projects — the serious ones. A non-technical founder building a marketplace for grid flexibility doesn’t need the codebase to be beautiful. They need it to exist, to be testable, to tell them whether the idea holds. A materials scientist who can prototype a data pipeline for tracking embodied carbon across a supply chain in a weekend, without an engineering queue, without a six-month roadmap — that’s a structural change in who gets to ask hard questions with software.

The creative differentiation will work exactly as it did in photography. The tools give everyone the button. The question is whether you have anything worth pointing at. That comes from domain knowledge, from taste, from two decades of understanding how an industry actually fails its own problems. None of that is in the prompt. The person who spent twenty years in energy systems or industrial logistics will build something different with these tools than someone who learned to prompt last Tuesday. The former has a subject. The latter has access.

Force multiplier, not replacement

The selfie didn’t make everyone a great photographer. It made everyone a photographer. The great ones were freed to operate at higher levels — more experiments, less friction, new forms that hadn’t existed before.

Vibe coding won’t make everyone a great software builder. It will make everyone a software builder. The great ones — those with genuine insight into hard problems, with the aesthetic judgement to know when something works, with domain knowledge that goes deeper than the prompt — will be freed to take on harder missions, to cross-pollinate between fields, to attempt things that were previously too expensive to prototype.

When I used it, the friction that disappeared wasn’t the interesting friction. I still had to decide what to build. I still had to know whether what came back was solving the right problem. I still missed things I didn’t know to look for. The tool didn’t give me taste or knowledge or the right question. It just got out of the way of the thinking.

That’s not a small thing. It’s also not a revolution in who has something worth building. The disgust, then, is worth acknowledging — the disruption to professional developers is real, the equity gaps are real, the security risks are real — and then setting aside. Most of the 47,500 photos taken every second are forgotten before the hour is out.¹³ That’s not the point. The point is what becomes possible at the edges, in the hands of people who have something real to say and now have a new way to say it.

The gate is off its hinges. The question is what you build with the opening.

References

“Selfies|me: Glimpses of Authenticity” (Academia.edu) — on the selfie as an identity practice and the cultural backlash that first framed it as narcissistic. Source: Academia.edu. ↩ ↩²
spiked, “Dear Oxford, we need to have a word about ‘selfie’” (22 Nov 2013) — the critical reaction to Oxford Dictionaries naming “selfie” its word of the year. Source: spiked. ↩
Inc., “The History and Future of Selfies” — on the selfie stick’s 2014 rise and the disdain it drew. Source: Inc.. ↩
Penji, “The Rise and Fall of Kodak” — how George Eastman’s Brownie (“You press the button, we do the rest”) collapsed photography’s barrier to entry. Source: Penji. ↩
NOVEDGE, “Design Software History: Aldus PageMaker” — how 1985’s PageMaker took professional publishing from expensive typesetting houses to the desktop. Source: NOVEDGE. ↩
LinkedIn, “How did desktop publishing revolutionize print?” — on DTP displacing unionised typographers and paste-up artists, often replacing decades of experience with weeks of PageMaker training. Source: LinkedIn. ↩
“The Rise of Vibe Coding in 2025” (Medium) — traces Andrej Karpathy’s Feb 2025 coining of the term and adoption stats: a quarter of YC’s W25 batch had 95% AI-generated codebases, and ~44% of non-technical founders now prototype with AI tools. Source: Medium. ↩ ↩²
Hostinger, “Vibe coding statistics” — reports that 63% of vibe-coding users are non-developers. Source: Hostinger. ↩
Superframeworks, “Vibe coding tipping point” — reports ~45% of AI-generated code carries security vulnerabilities (40% exploitable bugs) and recounts the Moltbook breach: 1.5 million API authentication tokens exposed in client-side JavaScript. Source: Superframeworks. ↩ ↩²
“State of Vibe Coding 2026” (Hashnode) — reports that only 33% of developers trust AI code accuracy. Source: Hashnode. ↩
IEEE Future Directions — film photography peaked in 1999 at roughly 80 billion photos taken worldwide. Source: IEEE. ↩
Mylio — photo-volume statistics: more than 8.6 trillion photos taken globally between 2010 and 2019, about seven times the previous decade. Source: Mylio. ↩
Lapse of the Shutter — photography statistics: roughly 1.72 trillion photos taken per year, about 47,500 every second. Source: Lapse of the Shutter. ↩ ↩²
CareerPlanner — photographer job outlook: as barriers to entry fell, salaried roles were disrupted while work demanding genuine skill, taste and presence became more specialised. Source: CareerPlanner. ↩
CSIS, “The Divide in Delivery: How AI Can Serve the Global South” — on AI’s risk of widening cross-country inequality, citing IMF analysis that growth impacts in advanced economies could more than double those in low-income countries. Source: CSIS. ↩
Tony Blair Institute, “State of Compute Access: How to Bridge the New Digital Divide” — documents the concentration of top supercomputers in the US, China and Germany, and their absence in regions such as East Africa. Source: Institute for Global Change. ↩
arXiv paper on tokenisation inequality — current tokenisation schemes systematically favour English and other high-resource languages, with efficiency variations of up to ~13× that raise cost and latency for speakers of underrepresented languages. Source: arXiv. ↩