Summary: Andrej Karpathy, who popularized the term “vibe coding”—using AI tools to write code with minimal oversight—recently revealed that for his latest project, nanochat, he wrote all the code by hand. This candid admission highlights the limitations of relying solely on AI for coding, especially for complex projects. While vibe coding can be fun and useful for quick experiments, real-world development often demands a more hands-on approach.
What Is Vibe Coding?
Over a year ago, OpenAI cofounder Andrej Karpathy left the company and introduced the term vibe coding to describe a laid-back style of programming where developers farm out coding tasks to AI tools and go with the flow. In his own words, vibe coding is about embracing the “vibes”—copying and pasting AI-generated code without deep scrutiny, fixing errors by trial and error, and letting the code evolve beyond one’s full understanding. Karpathy described it as ideal for “throwaway weekend projects,” where speed and experimentation matter more than perfection.
Karpathy’s Nanochat: A Hands-On Approach
Earlier this week, Karpathy released nanochat, an open-source project that offers a “minimal, from scratch, full-stack training/inference pipeline” to build a large language model with a ChatGPT-style chatbot interface. Impressively, he wrote about 8,000 lines of “quite clean code” entirely by hand, relying only on tab autocomplete. He candidly admitted that he tried using AI tools like Claude and Codex agents to help but found them “net unhelpful” and not up to the task.
“It’s basically entirely hand-written (with tab autocomplete),” Karpathy explained. This hands-on approach contrasts sharply with the vibe coding method he once championed.
Why Vibe Coding Doesn’t Always Cut It
Since nanochat isn’t a simple web app, the vibe coding approach didn’t fit this project’s complexity. This example underscores the limitations of vibe coding, despite the hype around AI as the future of programming. While vibe coding can speed up certain tasks, it often falls short for projects that require precision, optimization, and deep understanding.
The Reality of AI-Generated Code
Supporting this, a recent survey by cloud computing company Fastly found that 95% of developers spend extra time fixing AI-generated code. Some even report that debugging AI-written code takes longer than writing it from scratch. Research firm METR also discovered that AI tools can actually slow down developers. In response, some companies are hiring specialists specifically to clean up AI-generated coding messes.
The takeaway? While vibe coding can be a fun and efficient way to experiment, sometimes the vibes just aren’t good enough for serious development.