Torvalds explains where AI helps him and where it endangers the Linux kernel

Linus Torvalds, who turns 56 this year, and the Linux kernel he created, which turns 35, have weighed in on AI in programming. His formula is simple: "I use AI to do the things I'm bad at." He called vibe coding, meaning writing programs by describing a task to a neural network without closely reviewing the result, "a wonderful way to find joy in programming." But he immediately added a caveat: that approach is not right for managing Linux development. Using AI to find bugs is fine with him.

A tool for weak spots

This position contains neither enthusiasm nor rejection. Torvalds does not say AI will replace programmers. He describes it as a way to fill gaps where he has neither the skill nor the desire to build it. According to press reports, he previously used an AI assistant on a small hobby project: the assistant wrote an auxiliary part in a language Torvalds barely works with, while he wrote the main code himself.

The logic is clear. Someone who has written low-level C code all his life is under no obligation to become an expert in visualization or scripting. If a neural network does that work in minutes and the cost of a mistake is low, the trade is a good one. The word "joy" in his assessment matters too: vibe coding lowers the barrier to entry and restores the pleasure of watching an idea quickly become a working thing.

Why the kernel is in a different league

The problem starts where code becomes shared infrastructure. The Linux kernel runs on servers, smartphones, embedded devices and in the cloud. Every patch goes through maintainer review, and human attention remains the project's scarcest resource. Code that is generated without understanding and accepted "because it works" shifts the costs onto the people who have to review it.

This is our own conclusion from Torvalds' statement, not a quote from him. But it is consistent with what he said directly: vibe coding is fine for personal enjoyment, not for steering kernel development. There, what counts is predictability, readability and the author's responsibility for every line.

In this scheme, bug hunting looks like a natural exception. If AI points to a suspicious spot, a human still makes the decision. A wrong hint costs a few minutes of checking. A missed bug in the kernel costs far more.

The position in facts

• Source: ZDNet, article dated October 9, 2026.
• Key quote: "I use AI to do the things I'm bad at."
• Vibe coding: "a wonderful way to find joy in programming."
• Linux kernel development: vibe coding is not suitable.
• AI-assisted bug hunting: acceptable.
• Age of the project: Linux first appeared in 1991.

How to apply this approach on your team

Torvalds' position easily turns into a working rule. First, sort tasks by the cost of a mistake. Prototypes, internal scripts, visualizations and learning experiments can be handed to AI almost without restrictions. Code that will reach production must go through normal review, and the author must be able to explain every decision.

Second rule: find your own weak spots. If you are a backend developer and dislike writing frontend code, AI will give you the biggest payoff there. In your own area of expertise, it more often generates needless noise, and you will quickly spot its mistakes.

Third: start with analysis, not generation. Static analysis, hunting for races and leaks, and explaining someone else's code are tasks where AI saves time while the decision stays with a human. If your project already has a policy on AI patches, follow it. If not, it is worth writing one before the first generated commit lands in the main branch.

In the end, Torvalds urges neither abandoning AI nor trusting it blindly. He poses an honest question: how good are you yourself at the thing you are about to hand to a model?