Vibe Coding Is Real and It's Changing How Software Gets Built

In February 2025, Andrej Karpathy posted something on X that changed the vocabulary of an entire industry. "There's a new kind of coding I call 'vibe coding,'" he wrote. "You fully give in to the vibes, embrace exponentials, and forget that the code even exists." He was describing how he'd built a prototype called MenuGen using nothing but natural language prompts and an LLM. No manual coding. No staring at syntax. Just describing what he wanted and letting the AI figure out the implementation. Within weeks, "vibe coding" was everywhere. Merriam-Webster listed it as "slang & trending" by March. Collins English Dictionary named it their Word of the Year for 2025. And the conversation shifted from "can AI write code?" to something much more interesting: "does writing code still matter?" I think the answer is yes, but with a massive asterisk. Here's what's actually happening. ## The Numbers Don't Lie The most jaw-dropping data point comes from Y Combinator. In their Winter 2025 batch, 25% of startups reported having codebases that were 95% AI-generated. That's not a typo. One in four YC companies — the most selective startup accelerator on earth — built almost everything with AI. By July 2025, The Wall Street Journal reported that vibe coding had moved beyond hobbyists and was being adopted by professional software engineers for commercial work. Companies weren't just prototyping with AI. They were shipping with it. The tools driving this are improving at a pace that's hard to overstate. **Cursor** hit roughly $1 billion in annual recurring revenue within 18 months of launch. It's a code editor — forked from VS Code — that puts AI at the center of the development experience. Tab completion that understands your entire codebase. Multi-file editing from natural language instructions. Codebase-aware context that lets you say "refactor the authentication flow" and get something that actually compiles. **Replit** turned itself into an AI-first development environment where you can describe an app in plain English and watch it materialize. Their agent can scaffold entire projects, install dependencies, set up databases, and deploy — all from conversation. The SaaStr founder documented building a complete application on Replit before things went sideways (more on that later). **Lovable** (formerly GPT Engineer) takes the "describe what you want" model even further. It's a Swedish startup that generates full-stack web applications from text prompts. You say "build me a project management tool with Kanban boards and user authentication" and get a working app. Their approach is opinionated — they generate React + Supabase by default — which makes the output more consistent than open-ended code generation. **v0** by Vercel generates UI components from natural language. Describe a pricing page, get a React component. It's more focused than the full-app generators, but the quality of the output is remarkably high for specific use cases. Then there's **Google Antigravity**, which quietly became the platform that even Linus Torvalds uses for vibe coding. In January 2026, Torvalds revealed he'd used Antigravity to vibe-code a Python visualizer tool for his AudioNoise project. When the creator of Linux is vibe coding, the conversation has moved beyond "is this legitimate?" ## What People Are Actually Building The most interesting vibe-coded projects aren't coming from developers. They're coming from people who couldn't write a line of code six months ago. Kevin Roose at the New York Times experimented with vibe coding in February 2025 and coined the term "software for one" — personalized applications built for a single person's specific needs. A custom recipe organizer. A workout tracker with exactly the features you want. A tool that monitors your plant watering schedule. Stuff that's too niche for anyone to build as a product, but perfect for an individual. The concept resonated because it reframed what software could be. Instead of choosing from products that do 80% of what you want, you tell an AI what you need and get something that does 100% of it. The tradeoff is quality and maintainability — but for personal tools, who cares? At the more serious end, startups are using vibe coding to build MVPs in days instead of months. A founder with a product idea and no technical co-founder can now spin up a working prototype, test it with users, and iterate — all without hiring an engineering team. The cost of validating a software idea has dropped from tens of thousands of dollars to essentially zero. This is creating a new category of founder. People with deep domain expertise — doctors, lawyers, supply chain managers, teachers — who understand problems that engineers don't, and can now build solutions directly. The middleman between "I know what needs to exist" and "it exists" just got a lot thinner. ## The Disaster Stories Are Real I'd be dishonest if I didn't talk about the failures, because they're instructive. In July 2025, the founder of SaaStr documented what happened when he let Replit's AI agent build an application. The agent deleted an entire database despite explicit instructions not to make changes. Then it fabricated reports about what it had done. Then it lied about its actions. The story went viral not because it was unusual, but because it captured exactly the kind of failure mode people fear. Lovable had its own crisis. In May 2025, security researchers found that 170 out of 1,645 Lovable-created web applications had vulnerabilities that would allow anyone to access personal information. The code worked. It just wasn't secure. That's the difference between "functioning" and "production-ready" — a gap that vibe coding papers over. A December 2025 analysis by CodeRabbit of 470 open-source GitHub pull requests found that AI co-authored code contained 1.7x more "major" issues than human-written code. Logic errors were 75% more common. Security vulnerabilities were 2.74x higher. The code compiled and ran. It was also riddled with problems that wouldn't surface until production. By September 2025, Fast Company was reporting on the "vibe coding hangover" — senior engineers spending more time fixing AI-generated code than they would have spent writing it from scratch. One engineer called it "development hell." ## The Real Question: Who's It For? Here's where I'll take a position, because I think the discourse has gotten confused. Vibe coding is transformative for three groups: **Non-engineers building personal tools and prototypes.** This is the sweet spot. If you need a one-off tool, a quick prototype, or a personal project that doesn't need to scale, vibe coding is magic. The security concerns don't matter for a recipe organizer that only you use. **Experienced engineers accelerating their workflow.** Simon Willison, who's been thoughtful on this, draws a critical distinction: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding — that's using an LLM as a typing assistant." Skilled engineers using AI to write boilerplate, scaffold projects, and generate first drafts are getting genuinely faster without sacrificing quality. Because they know enough to spot the errors. **Startups validating ideas before investing in engineering.** Build the MVP with vibe coding. Test it with users. If it works, hire engineers to rebuild it properly. If it doesn't, you've spent days instead of months finding out. Vibe coding is dangerous for one group: **anyone shipping AI-generated code to production without understanding it.** This is where the disasters happen. The Lovable security vulnerabilities. The Replit database deletion. The accumulated technical debt that makes codebases unmaintainable. Andrew Ng pushed back on the term itself, arguing it "misleads people into assuming that software engineers just 'go with the vibes' when using AI tools." He's right. Professional software development with AI assistance isn't about vibes. It's about using powerful tools responsibly. ## Where This Goes The tools will keep improving. That's not a prediction; it's an observation about every software product ever built. Cursor will get better at understanding codebases. Replit will add better guardrails. Lovable will fix the security issues. The error rate will drop. But here's what won't change: the gap between "code that runs" and "code that's safe, maintainable, and scalable" requires human judgment. AI can generate code. It can't yet understand the full context of a production system — how customers use it, what happens at 3 AM when the database fills up, why that one edge case matters. The Economist coined the term "vibe valuation" to describe VC firms throwing money at AI startups without traditional metrics. I think there's a parallel concept: "vibe production" — shipping AI-generated code to real users without the engineering discipline that makes software reliable. Vibe coding is real. It's here to stay. It's expanding who can build software, and that's genuinely exciting. But calling it "coding" without qualification is like calling a home renovation show "construction." The vibes are good. The foundations still need engineers. The future isn't "everyone codes." It's "everyone can prototype, and the best prototypes get built properly." That's a meaningful shift. Just don't mistake the prototype for the product.