Even the Pioneer of 'Vibe Coding' Admits It Just Doesn't Hold Up in Real-World Challenges
Imagine this: You're a tech innovator who helped shape the future of AI, only to step back and realize that the very tool you championed might not be the game-changer everyone thought it was. That's the fascinating—and somewhat ironic—story unfolding around Andrej Karpathy, the OpenAI cofounder who left the company over a year ago and introduced the world to 'vibe coding.' But here's where it gets controversial: Karpathy himself has now turned his back on it, revealing its flaws in a way that could shake up how we view AI's role in programming. Curious to see why? Let's dive in.
For those new to the concept, 'vibe coding' is Karpathy's playful term for relying heavily on AI coding assistants to handle projects. Instead of meticulously writing and debugging code yourself, you 'farm out' the work to these tools, going with the flow—copying prompts, pasting error messages, and letting the AI generate solutions without fully understanding the underlying logic. Karpathy popularized this idea in a now-famous post, describing it as a low-effort way to build quick, disposable apps or websites. 'When I get error messages, I just copy-paste them in with no comment, usually that fixes it,' he explained. 'The code grows beyond my usual comprehension... Sometimes the LLMs can’t fix a bug, so I just work around it or ask for random changes until it goes away.' It's like letting the AI be your creative partner, but only for things that don't require deep inspection.
But—and this is the part most people miss—Karpathy recently put theory to test with his own project, an open-source model called nanochat. This isn't just any project; it's a 'minimal, from-scratch, full-stack training and inference pipeline' that empowers anyone to create a large language model complete with a ChatGPT-like chatbot interface. The catch? It can be done in just hours for under $100. And here's the kicker: Karpathy wrote every single line of its approximately 8,000 lines of code by hand, using nothing more than tab autocomplete for assistance. No AI vibes here.
In a candid admission, he shared on social media why he abandoned his own creation. 'It’s basically entirely hand-written... I tried to use Claude/Codex agents a few times, but they just didn’t work well enough at all and net unhelpful.' This stark contrast to his earlier enthusiasm for vibe coding is telling. Sure, he positioned it as ideal for 'throwaway weekend projects'—those fun, low-stakes experiments where you don't need bulletproof code. But nanochat, being a robust, foundational tool, demanded precision that AI couldn't deliver. It's a clear reminder that while AI might excel at rapid prototyping, it often falls short for complex, mission-critical development.
And this is where things get really intriguing: Despite the hype that vibe coding could revolutionize programming—think of it as the future where developers just 'see stuff, say stuff, run stuff, and copy-paste stuff'—reality is catching up. Take the findings from a recent survey by cloud computing giant Fastly: 95% of developers reported spending extra time fixing AI-generated code, sometimes more than the time saved by using the tools in the first place. Similarly, research from METR, an AI evaluation firm, showed that incorporating AI actually slows down developers on tasks. Not to mention, companies are increasingly hiring human experts just to clean up the messes left by AI-assisted coding bloopers, as reported in industry news.
To clarify for beginners, imagine building a house: AI might help sketch a quick blueprint or suggest materials for a simple shed, but for a sturdy, custom-built home, you'd still need skilled architects and contractors to ensure it doesn't collapse under pressure. Vibe coding works great for prototypes or prototypes, but when the stakes rise, the 'bad vibes'—like unfixable bugs or incomprehensible code—can derail the whole process. It's not that AI is useless; it's more about knowing its limits.
Of course, this doesn't mean vibe coding is dead in the water. Some argue it's a democratizing force, making coding accessible to non-experts and speeding up innovation in niche areas. But is it truly sustainable for advanced projects, or are we risking a future where sloppy code leads to bigger problems? And here's a thought-provoking question: Do you think AI will eventually bridge this gap, making vibe coding viable for everything, or is human oversight here to stay? What are your experiences with AI in coding—has it saved you time, or left you frustrated? Share your thoughts in the comments below; I'd love to hear agreements, disagreements, or even counterpoints that challenge Karpathy's shift. After all, in the fast-evolving world of tech, opinions like yours could shape the next debate!
