OpenAI Cofounder Andrej Karpathy Builds AI Model By Hand, Challenging “Vibe Coding” Trend
Despite popularizing the practice of leveraging AI for code generation, OpenAI cofounder Andrej Karpathy recently completed an entire project – a new open-source model called nanochat – through traditional, manual coding. This revelation casts a spotlight on the limitations of current AI coding tools and the enduring value of human expertise.
Karpathy, who left OpenAI over a year ago, has been a leading voice in the emerging field of “vibe coding,” which refers to outsourcing coding tasks to artificial intelligence platforms. However, his latest project demonstrates a surprising return to fundamentals.
Nanochat, described by Karpathy as a “minimal, from scratch, full-stack training/inference pipeline,” is designed to empower users to construct a large language model with a ChatGPT-style chatbot interface quickly and affordably – potentially within hours and for as little as $100. The project comprises approximately 8,000 lines of “quite clean code,” all meticulously written by hand.
“It’s basically entirely hand-written (with tab autocomplete),” Karpathy wrote. He explained that while he experimented with AI coding assistants like Claude and Codex, they ultimately proved inadequate for the task. “I tried to use claude/codex agents a few times but they just didn’t work well enough at all and net unhelpful.”
This admission is particularly noteworthy given Karpathy’s advocacy for AI-assisted development. It suggests that, even for a seasoned engineer like himself, current AI tools are not yet capable of handling complex projects independently. The experience underscores the need for a nuanced perspective on the role of AI in software development – one that acknowledges its potential while recognizing its current limitations.
The development of nanochat highlights a critical juncture in the evolution of AI and coding. While AI continues to advance, the human element remains essential for innovation and the creation of robust, reliable software. This project serves as a reminder that “vibe coding” is not a replacement for skilled programmers, but rather a potentially useful tool within a broader skillset.
