Vibe Coding Pioneer Returns to Traditional Methods
OpenAI cofounder Andrej Karpathy, who popularized the term “vibe coding” to describe delegating programming tasks to AI assistants, has reportedly written his latest project entirely by hand. According to reports, Karpathy’s new open source model called nanochat contains approximately 8,000 lines of hand-written code despite his previous advocacy for AI-assisted programming approaches.
Nanochat: A Minimalist Approach to LLM Development
Karpathy described nanochat as a “minimal, from scratch, full-stack training/inference pipeline” designed to enable anyone to build a large language model with a ChatGPT-style interface within hours and for under $100. Sources indicate that the project represents a significant departure from his previously stated coding philosophy, with Karpathy admitting he tried using AI coding assistants but found them “net unhelpful” for this particular project.
The Limitations of Vibe Coding Revealed
While Karpathy had previously championed the concept of fully embracing AI assistance without scrutinizing generated code, his experience with nanochat appears to have exposed practical limitations. According to his statements, he discovered that for complex, foundational programming work rather than “throwaway weekend projects,” current AI tools couldn’t meet his requirements for clean, reliable code.
Industry Research Confirms Productivity Concerns
These developments align with recent industry findings about AI coding assistance. A survey from cloud computing company Fastly found that 95% of developers reportedly spend extra time fixing AI-generated code, with some indicating that error correction takes longer than the time saved through initial code generation. Additionally, research from METR suggests that using AI tools may actually slow developers’ task completion rates.
The Evolving Landscape of AI-Assisted Programming
Karpathy’s shift in approach highlights the ongoing evolution of AI tools in software development. While he previously described vibe coding as involving minimal code review and simply pasting error messages back to AI tools until issues resolved, his nanochat experience demonstrates that different projects require different approaches. Analysts suggest this pattern reflects broader industry learning curves as developers determine optimal use cases for various coding methodologies.
Open Source Development and AI Integration
The nanochat project, as an open source model, represents the type of complex development work where AI assistance currently shows limitations. Industry reports indicate that some companies have begun hiring human specialists specifically to address coding issues created by AI tools, suggesting that while AI assistance has its place, human expertise remains crucial for certain development challenges.
Future Implications for Development Practices
The contrast between Karpathy’s earlier vibe coding advocacy and his current hand-coded approach provides valuable insights into the maturation of AI programming tools. As the industry continues to evaluate the practical implementation of these technologies, developers are reportedly developing more nuanced understanding of when to leverage AI assistance versus when traditional coding methods remain preferable for specific project requirements and quality standards.
This article aggregates information from publicly available sources. All trademarks and copyrights belong to their respective owners.
