“Even the creator of ‘vibe coding’ agrees: it’s not a one-size-fits-all solution”
The Evolution of “Vibe Coding”
It’s been a year since OpenAI co-founder Andrej Karpathy coined the term “vibe coding” to describe the practice of delegating coding projects to AI tools. This week, he launched his own nanochat model, admitting that he wrote everything by hand without using vibes.
Nanochat: A New Approach
Nanochat is designed for anyone to build a large language model using a chatbot interface like ChatGPT in just a few hours and at a low cost of $100. Karpathy wrote about 8,000 lines of clean code for the project by hand, as he found that AI tools couldn’t meet his needs.
Rethinking the Strategy
Karpathy’s approach with nanochat contrasts with his previous embrace of “vibe coding” for weekend projects. While vibe coding involves surrendering to vibes and not looking at the code, this strategy didn’t work for nanochat. A survey by Fastly found that 95% of developers spend extra time fixing AI-generated code, with some taking longer to correct errors than if they had written the code themselves. METR’s research also shows that using AI tools can actually slow down developers, leading some companies to hire human specialists to fix AI-generated code issues. Sometimes, vibes may not be the best approach after all.
