He coined the term “vibe coding,” but he’s not blown away by it.
At a talk at Sequoia Capital released on Wednesday, former Tesla AI head and OpenAI founding member Andrej Karpathy said that code written by AI can still be messy and needs human supervision.
“Right now, the answer is that the agents are like these intern entities,” he said. “You basically still have to be in charge of the aesthetics, the judgment, the taste, and a little bit of oversight.”
That’s because the AI-constructed code is far from perfect, he said.
“Sometimes I get a little bit of a heart attack because it’s not like super amazing code necessarily all the time,” he said. “It’s very bloaty, and there’s a lot of copy-paste, and there’s awkward abstractions that are brittle, and it works, but it’s just really gross.”
Karpathy, who now runs the AI-powered education platform Eureka Labs, introduced the concept of “vibe coding” in a post on X in February last year. He described it as a highly AI-assisted style of development in which builders barely touch the code themselves. In November, Collins Dictionary named vibe coding as its 2025 word of the year, cementing its importance in both tech and non-tech circles.
On Wednesday’s talk, the OpenAI cofounder said that humans will remain in charge of some high-level development decisions, while the AI does a lot “under the hood.”
Karpathy added that there’s nothing “fundamental” preventing AI from writing clean code; it’s that labs haven’t focused on this problem during model training yet.
The vibe coding wave, which took off even before Karpathy coined the term, has disrupted how tech companies hire and reward employees and shaken software stocks. It’s prompted an influx of companies building tools for professional and novice developers, and sent the valuations of startups such as Lovable, Cursor, and Replit well into the billions.
Despite this virality, professional developers discourage an overreliance on AI because it can produce messy, untested code. Recent security stumbles that risk exposing company data have also given users another reason to be wary of vibe coding.
Last week, the Swedish vibe coding platform Lovable said it had detected a security error after user backlash on X.
“Unfortunately, in February, while unifying permissions in our backend, we accidentally re-enabled access to chats on public projects,” Lovable wrote on X. “Upon learning this, we immediately reverted the change to make all public projects’ chats private again. We appreciate the researchers who uncovered this.”
