Even the Inventor of ‘Vibe Coding’ Says Vibe Coding Can’t Cut It

It’s been over a year since OpenAI cofounder Andrej Karpathy exited the company. In the time since he’s been gone, he coined and popularized the term “vibe coding” to describe the practice of farming out coding projects to AI tools. But earlier this week, when he released his own open source model called nanochat, he admitted that he wrote the whole thing by hand, vibes be damned.

Nanochat, according to Karpathy, is a “minimal, from scratch, full-stack training/inference pipeline” that is designed to let anyone build a large language model with a ChatGPT-style chatbot interface in a matter of hours and for as little as $100. Karpathy said the project contains about 8,000 lines of “quite clean code,” which he wrote by hand—not necessarily by choice, but because he found AI tools couldn’t do what he needed.

“It’s basically entirely hand-written (with tab autocomplete),” he wrote. “I tried to use claude/codex agents a few times but they just didn’t work well enough at all and net unhelpful.”

That’s a much different attitude than what Karpathy has projected in the past, though notably he described vibe coding as something best for “throwaway weekend projects.” In his post that is now often credited with being the origin of “vibe coding” as a popular term, Karpathy said that when using AI coding tools, he chooses to “fully give in to the vibes” and not bother actually looking at the code. “When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away,” he wrote. “I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”

Of course, nanochat is not a web app, so it makes sense that the strategy didn’t work in this case. But it does highlight the limitations of such an approach, despite lofty promises that it’s the future of programming. Earlier this year, a survey from cloud computing company Fastly found that 95% of surveyed developers said they spend extra time fixing AI-generated code, with some reporting that it takes more time to fix errors than is saved initially by generating the code with AI tools. Research firm METR also recently found that using AI tools actually makes developers slower to complete tasks, and some companies have started hiring human specialists to fix coding messes made by AI tools. The thing to remember about vibe coding is that sometimes the vibes are bad.