Summary: Andrej Karpathy, cofounder of OpenAI and the person who popularized “vibe coding,” recently released an open-source project called nanochat. Despite his previous enthusiasm for vibe coding—relying heavily on AI to generate code with minimal oversight—Karpathy revealed he hand-wrote nearly all of nanochat’s 8,000 lines of code. This candid admission highlights the limitations of AI-assisted coding, especially for complex projects, and reflects broader developer experiences where AI-generated code often requires significant manual fixing.
What Is Vibe Coding?
Over a year ago, Andrej Karpathy left OpenAI and introduced the term “vibe coding” to describe a style of programming where developers delegate coding tasks to AI tools. The idea is to “go with the vibes”—letting AI generate code snippets and integrating them with minimal scrutiny. Karpathy described this approach as ideal for “throwaway weekend projects,” where speed and experimentation take precedence over code quality.
Karpathy’s Nanochat: A Hand-Coded Project
Earlier this week, Karpathy unveiled nanochat, an open-source, full-stack training and inference pipeline. It enables anyone to build a large language model with a ChatGPT-style chatbot interface quickly and affordably—potentially in just a few hours and for as little as $100.
Surprisingly, Karpathy revealed that he wrote almost all of nanochat’s roughly 8,000 lines of code by hand. He explained, “It’s basically entirely hand-written (with tab autocomplete). 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 contrasts sharply with his earlier vibe coding philosophy.
Why Vibe Coding Falls Short for Complex Projects
Karpathy’s experience with nanochat underscores a key limitation of vibe coding: AI tools often struggle with complex or nuanced programming tasks. While vibe coding can speed up simple projects, it may not be reliable for building robust, production-quality software.
In his original post about vibe coding, Karpathy admitted that he often doesn’t fully understand the AI-generated code and sometimes resorts to trial and error—copy-pasting error messages or asking the AI for random changes until the code works. This approach can be frustrating and inefficient, especially when bugs persist.
The Reality of AI-Generated Code in Development
Karpathy’s experience aligns with broader industry findings. A recent survey by cloud computing company Fastly found that 95% of developers spend extra time fixing AI-generated code. Some even reported that debugging AI code takes longer than writing it from scratch. Research firm METR also found that AI tools can slow down developers, and some companies are hiring specialists specifically to clean up AI-generated coding errors.
Ultimately, while vibe coding offers an exciting glimpse into the future of programming, it’s important to remember that AI-generated code isn’t a magic bullet. Sometimes, the vibes just aren’t right.