Summary: Andrej Karpathy, who popularized the term “vibe coding”—the practice of relying heavily on AI tools to write code—recently revealed that he wrote his latest open-source project, nanochat, entirely by hand. This admission highlights the current limitations of AI-assisted coding and reminds developers that sometimes, a hands-on approach is still necessary.
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 lean on AI tools to generate code with minimal oversight. In his own words, vibe coding involves “fully giving in to the vibes,” often copying and pasting AI-generated snippets without deeply understanding or reviewing them. This approach, he noted, works best for “throwaway weekend projects” rather than complex applications.
Karpathy’s Nanochat: A Hands-On Approach
Earlier this week, Karpathy released nanochat, an open-source project described as a “minimal, from scratch, full-stack training/inference pipeline” that enables anyone to build a large language model with a ChatGPT-style chatbot interface quickly and affordably—potentially for as little as $100.
Despite his history with vibe coding, Karpathy admitted that he wrote the approximately 8,000 lines of code for nanochat entirely by hand, using tab autocomplete but not relying on AI coding assistants like Claude or Codex. He explained, “I tried to use claude/codex agents a few times but they just didn’t work well enough at all and net unhelpful.” This hands-on approach was necessary because the AI tools couldn’t meet the project’s specific demands.
Why Vibe Coding Isn’t Always Enough
Karpathy’s experience with nanochat underscores the limitations of vibe coding, especially for complex or foundational projects. While vibe coding can speed up development for simple or experimental tasks, it may fall short when precision and deep understanding are required.
The Reality of AI-Generated Code
Supporting this, a survey from cloud computing company Fastly found that 95% of developers spend extra time fixing AI-generated code. Some even report that debugging AI-assisted code takes longer than writing it from scratch. Additionally, research from METR indicates that AI tools can actually slow down developers in completing tasks. As a result, some companies are now hiring specialists specifically to clean up coding issues introduced by AI tools.
Final Thoughts: When to Trust the Vibes
While vibe coding offers an exciting glimpse into the future of programming, it’s clear that the approach isn’t foolproof. Sometimes, the vibes just aren’t good enough, and a careful, hands-on coding process remains essential. Karpathy’s nanochat project serves as a timely reminder that AI is a powerful assistant—but not yet a replacement for human expertise.