The man who named the future of coding just admitted it writes ugly code.
The Summary
- Andrej Karpathy, who coined "vibe coding" in February 2025, says AI-generated code is "bloaty," "brittle," and full of "awkward abstractions" despite the approach becoming mainstream enough that Collins Dictionary named it 2025's word of the year.
- AI coding agents function as "intern entities" that require human oversight for "aesthetics, judgment, and taste."
- The timing is notable: as Lovable launches mobile apps for on-the-go vibe coding, the term's creator is pumping the brakes on expectations.
The Signal
Karpathy's comments at a Sequoia Capital talk land like cold water on the vibe coding hype cycle. He describes opening AI-generated code and getting "a little bit of a heart attack" because of rampant copy-paste, bloated logic, and fragile abstractions. It works, he says, but it's "really gross."
This matters because Karpathy isn't some skeptic throwing stones from the sidelines. He's the former Tesla AI director, an OpenAI founding member, and now runs Eureka Labs, an AI-powered education platform. He invented the term vibe coding to describe highly AI-assisted development where humans barely touch code themselves. If anyone should be bullish on AI agents writing production code, it's him.
"You basically still have to be in charge of the aesthetics, the judgment, the taste, and a little bit of oversight."
Yet here we are. The workflow he named has gone so mainstream that Collins Dictionary crowned it word of the year for 2025. Companies like Lovable are building entire mobile products around it, letting developers "vibe code web apps and websites on the go." But the quality gap between what agents can ship and what human developers would accept is still wide enough to trigger heart attacks.
The "intern" metaphor is telling. Interns can be productive. They can close tickets, ship features, and move projects forward. But you wouldn't leave an intern unsupervised on architecture decisions or let them set coding standards for the team. That's the current state of AI coding agents: useful, sometimes impressively so, but fundamentally requiring human taste and judgment in the loop.
Key issues Karpathy flags:
- Code bloat from AI defaulting to verbose solutions
- Copy-paste patterns instead of proper abstractions
- Brittle code that works now but breaks when touched
- Security concerns from generated code (mentioned in original context)
This isn't just an aesthetic complaint. Bloated, brittle code compounds. It slows teams down six months later when they need to modify it. It increases attack surface. It makes debugging harder. The short-term velocity gains from vibe coding can create long-term technical debt that humans have to service.
The Implication
Developers using AI coding tools need to treat them like what they are: very fast interns who need review. The workflow isn't "prompt and deploy." It's "prompt, review, refactor, then deploy." Teams that skip the middle steps will ship faster today and pay for it in six months when that brittle code starts breaking.
The bigger signal: we're entering a phase where the AI coding hype meets reality. Vibe coding is real and useful, but it's not replacing human developers anytime soon. It's changing what developers do, shifting more time toward architecture, code review, and taste-making. If you're building a company in this space, that's the product to build for: tools that help humans review and refactor AI code, not tools that promise to eliminate humans from the loop entirely.