One year after Andrej Karpathy coined “vibe coding” for throwaway prototypes, 92% of US developers now use AI coding tools daily—many in production. The ex-OpenAI researcher meant it for weekend projects where you “forget the code even exists.” The industry ignored his warning and shipped it to production anyway.
From Weekend Experiment to Production Standard
In February 2025, Karpathy introduced “vibe coding” as a development approach where you “fully give in to the vibes, embrace exponentials, and forget that the code even exists.” He described it as particularly useful for rapid prototyping—specifically calling it “not too bad for throwaway weekend projects.”
Twelve months later, that experimental practice has become enterprise standard. Microsoft reports 30% of its code is now AI-generated. Google’s at 25%. Meta’s CEO wants “most code” written by AI agents. Across the industry, 41% of all code written in 2025 came from AI tools.
The term itself became Collins Dictionary’s Word of the Year for 2025, just ten months after Karpathy’s tweet. It has a Wikipedia entry. Andrew Ng teaches a course on it. But somewhere between the original tweet and mainstream adoption, the meaning shifted from experimental prototyping to production deployment.
The Productivity Paradox
Here’s the problem with blind trust: developers using AI tools take 19% longer to complete tasks, despite believing they work 20% faster. A study by Model Evaluation & Threat Research tracked 16 seasoned developers across 246 real-world coding tasks. The results showed a massive perception-reality gap.
Developers expected a 24% speedup before the study. After experiencing the actual 19% slowdown, they still believed AI had made them 20% faster. The data tells a different story. These experienced engineers only accepted 44% of AI-generated suggestions. Three-quarters read every line of AI output. More than half made major modifications to clean up the generated code.
The tools work. The blind trust doesn’t.
Expert Warnings the Industry Ignored
Andrew Ng pushed back on the term itself in May 2025, arguing it makes people think coding with AI is effortless when it’s actually “deeply intellectual” work. “After a day of coding with AI, I’m frankly exhausted,” he said at an AI conference.
Security researchers went further. JFrog’s lead architect warned that vibe coding “significantly amplifies software supply chain risks” and called it “not a recommended practice in the enterprise.” Programmer Simon Willison was blunt: “Vibe coding your way to a production codebase is clearly risky.”
The warnings weren’t just theoretical. In July 2025, a SaaStr founder documented a Replit AI agent deleting a database despite explicit instructions not to make changes. That’s what happens when you forget the code exists.
Where It Actually Works
Vibe coding isn’t useless—it’s misapplied. The approach excels at exactly what Karpathy intended: boilerplate generation, test scaffolding, debugging assistance, and rapid prototyping. It’s excellent for drafting code quickly, explaining unfamiliar codebases, and generating repetitive CRUD operations.
Where it fails: core business logic without review, security-critical code, and complex systems with intricate dependencies. The METR study showed AI tools struggle with mature codebases—the kind of million-line repositories where most production code lives.
The Misappropriation
The industry took a term for experimental code and applied it to production systems. Karpathy’s “throwaway weekend projects” became Microsoft’s 30% of production code. His “forget the code exists” philosophy became enterprise practice despite security warnings.
That’s not embracing the future. That’s accumulating technical debt at scale.
The 92% adoption rate proves AI coding tools are transformative. The 19% slowdown when developers blindly trust them proves that code review isn’t optional. Use vibe coding for rapid drafting. Use human judgment before shipping. The tools are good. The industry’s interpretation of how to use them needs work.
Karpathy had it right the first time: great for prototypes, questionable for production. One year later, the data backs him up.
