The Productivity Tip

Andrej Karpathy is at a conference. He's on stage, relaxed, that signature open smile. He tells the room he's been using AI coding tools and that he's stopped checking the output. He used to correct it all the time. Now he doesn't need to. It's a productivity tip. The tip is: give up reviewing the code.

The room nods. This is progress.

Then, almost immediately, he tells a second story. An AI agent working on his app called Menu Genen made an assumption that didn't make any sense , something about reusing emails. "The kind of thing," he says, "that if I didn't catch it, would have been really catastrophic."

He tells this as a charming anecdote. An isn't-that-funny moment. The room laughs.

He coined the term vibe coding. He is one of the founders of OpenAI. He helped build the modern AI stack. And he just told us in the same breath that he stopped reviewing AI output, and that when he doesn't review AI output, catastrophic things happen.

The Heart Attack

Later in the same interview, Karpathy says this: "When you actually look at the code, sometimes I get a little bit of a heart attack. It's not super amazing code necessarily all the time. It's very bloaty. There's a lot of copy-paste. There are awkward abstractions that are brittle. It works. But it's just really gross."

He has a medical event, he says, every time he opens a pull request and looks at what the AI wrote.

This is the same technology that is supposed to be replacing software developers. Writing production codebases for companies with real users. The founder of vibe coding is having a heart attack looking at vibe-coded code.

He describes his actual workflow: write an extremely detailed specification document in Markdown. Every edge case. Every precise requirement. Then give it to the AI. Because if you are not that precise, the AI produces something that works but is structurally terrible. He asked the AI to simplify bloated code it had written. The AI said it couldn't.

The Thing Engineers with Jobs Won't Say

Karpathy is unusually honest about this because he doesn't have anything to protect. Most engineers who are publicly enthusiastic about AI have a commercial interest in the enthusiasm. They are riding the wave. The CEOs they work for don't look at the code. The gap between the demo and the pull request is a secret the industry is quietly keeping.

Karpathy names it. The code is brittle. The abstractions are awkward. It works, but it's gross. This is the frontier , not in spite of these limitations, but including them. This is what the leading edge actually looks like right now.

What He Said About Getting Hired

He had one piece of advice for people trying to get jobs as developers in 2026 that landed differently than the rest of the interview.

Most companies, he said, haven't updated their hiring process for agentic engineering. They are still running LeetCode interviews and abstract puzzles. But what actually matters now is writing specs. If someone tells you to build a Twitter clone in an interview, the question is: what does your spec document look like? If you've never practiced writing specs at home, you'll produce something simple, and the AI will choke on everything you didn't specify , token expiration, rate limiting, password reset flows.

Practice writing specs. Not algorithms.

The Part He Couldn't Answer

Near the end of the interview, the host asked him what skills would still be worth learning if AI keeps getting better. What things remain valuable when the models improve?

He gave a response. It was empty. Not because he was hiding anything , because he genuinely doesn't know. Nobody does. The person who coined vibe coding and co-founded the company that built GPT-4 does not have a clear answer for what matters in a world where AI writes most of the code.

If you are feeling a little lost about that question, you are in good company. Even Andre feels a little lost.