Linus Torvalds Says Vibe Coding is Fine For Getting Started, 'Horrible Idea' For Maintenance (theregister.com)
- Reference: 0180143777
- News link: https://developers.slashdot.org/story/25/11/19/1832208/linus-torvalds-says-vibe-coding-is-fine-for-getting-started-horrible-idea-for-maintenance
- Source link: https://www.theregister.com/2025/11/18/linus_torvalds_vibe_coding/
Torvalds told Dirk Hohndel, head of open source at Verizon, that computers have become more complicated than when he learned to code by typing in programs from computer magazines. He said vibe coding offers a path into computing for newcomers. The kernel maintainer is not using AI-assisted coding himself. He said his role has shifted from rejecting new ideas to sometimes pushing for them against opposition from longstanding maintainers who "kind of get stuck in a rut."
Rust is "actually becoming a real part of the kernel instead of being this experimental thing," he said. Torvalds said AI crawlers have been "very disruptive to a lot of our infrastructure" because they gather data from kernel.org source code. Kernel maintainers receive bugs and security notices that are "made up by people who misuse AI," though the problem is smaller than for other projects such as curl.
[1] https://www.theregister.com/2025/11/18/linus_torvalds_vibe_coding/?td=rt-3a
But it's a self-defeating loop (Score:5, Insightful)
A beginner using AI to "vibe code" isn't going to understand when it's making things up out of whole cloth (like how it has a tendency to hallucinate Powershell cmdlets) and is going to spend more time whack-a-moling the result to get a kinda-working program than they would if they just started with a basic foundation and got better as they go.
You learn to build a skyscraper by first learning to build a house, not by ChatGPTing a skyscraper and then figuring out what you need to shore up to prevent it from falling down.
Re: (Score:2)
"Hey, Sora, show me a video of what my building would have looked like at the grand opening if it hadn't fallen on the orphanage. Oh, and ChatGPT, I need a recipe for toilet wine."
Re:But it's a self-defeating loop (Score:4)
Indeed. We should also remember that while Linus is a good engineer and a stellar project leader, he is not an engineering-teacher.
Re: (Score:2)
It's not very good at course-correcting a mistake during generation but if you ask it after every code generation prompt to pretend someone else wrote it and find the mistakes you have a shot at it fixing it for you. Still not a useful tool if you don't plan to learn anything while using it.
Re: (Score:1)
> You learn to build a skyscraper by first learning to build a house, not by ChatGPTing a skyscraper and then figuring out what you need to shore up to prevent it from falling down.
I love this quote!!!!
Re: (Score:2)
This is a really good point and one of the reasons why I'll happily keep to vim rather than using an IDE which I felt might have done too much for me. When coding with vim I like to keep the man pages in a terminal and other libraries on a split screen. I think using a code generator would just feel much worse than even letting the IDE suggest library functions to you.
Maybe for somethings it's ok, but I've not found much. It seems a lot of uses are for asking it to give boiler plate back, but why bother, it
Fine for getting started (Score:4, Informative)
I'll agree with that, especially if you're younger / less experienced and don't have a lot of code you've written banked from which you can pull. Cleaning up, or at least heavily reviewing, the vibe code for production may be a good way to hone your skills. The environment is kind of like that already with the existence of sites like Stack Overflow -- none of which were around when I was in university and getting started.
Way back then, the system administrator (4.3BSD on VAX-11/785 and, later, also earlier Sun systems) was also a very knowledgeable programmer and would answer SA/coding questions - eventually. His first answer was always, "Did you read the man page?" [Off to read man pages.] His second was, "Did you read the source code?" [Off to read BSD source.] Then he would lean back, with the keyboard still on his lap, and scribble something helpful on the whiteboard... It actually was a good, if annoying, learning process as I read a LOT of man pages and BSD source -- which helped me a immensely when I became a SA and systems programmer.
Vibe coding is an intermediate step that will die (Score:2)
I don't think vibe coding is going to last long as a thing, because it's just a sort of intermediate step to telling the AI to do what you want and having it do that. Right now, people are telling the LLM to write code to accomplish a thing and then running the code to see how it works, then telling the LLM to refine it, but that's a lot of unnecessary extra steps. I'm sure that in the not-too-distant future people will just tell the LLM what they want to do, which may require creating a custom user inter
Re: (Score:3)
Prompt engineering is already mostly dead. You need to learn to communicate clearly and efficiently for AI to understand and that skill is useful with humans as well. You needed prompt engineering when models had like a few thousand characters for input and output combined, such that you prompt needed to be short. The system prompt of some of the large company's AI models are pages long of instructions given as well-structured human readable text without looking like anyone was counting tokens when writing
Vibe Coding != AI Coding (Score:4, Insightful)
Vibe coding is not the same as general coding with AI.
With "Vibe coding" you basically instruct an agent to create a document containing a plan for your program and then to step by step create the required components. For some projects and some programming languages that works some time, but eventually these projects get to the point when it stops to work because the LLMs lose the overview, even if they would in general be fit for the project.
But there is also the option to have the AI as a tool in your IDE. Select a piece of code and then ask the LLM about it, or instruct the LLM to change your tail recursion to a loop or similar things. The piece of code is manageable and and you are watching the LLM program like a pair programmer and instantly stop it if it begins to do things you didn't want. You can also use it as StackOverflow with fewer insults, by asking the LLM "Is there a STL function that can wrap this for-loop more idiomatic" and get an answer for your code even may be no drop-in snippet but more an explanation.
Don't just buy what large companies want to sell you, but only take the pieces that are useful for you and leave out the others.
Is it? (Score:2)
Learning requires quality input data to your learning process. That's why school exist. If AI gets it wrong or convoluted roughly half the time, is AI really the right knowledge source to recommend to potential student?
News at 11: Blowhard bloviates obvious bias (Score:2)
Blowhard bloviates obvious cognitive biases. News at 11.
Why does he keep doing this?
It can't be good for the Linux kernel.
Depends... (Score:5, Insightful)
On what you are maintaining, and how you approach it. The code I maintain, if I need to add new functionality to a function, and I can describe it to claude easily enough then I'll tell claude code what file, what function add a marker where I want the code and explain what I want. I don't let it just edit on it's own I have it show me what it wants to do, and if it looks good I'll have them do it.
Just telling claude hey add this functionality to my code base, and hoping it gets it right, which from what I am understanding seems a lot of people are doing, is a good way to get bad results.
Re: (Score:1)
I just use git to undo any harm and let claude do it's thing. Also, if certain code has no tests, I ask it to write tests for what's currently implemented first before changing anything
Re: (Score:3)
There's even more to maintaining than that. The maintainer has to read and understand the existing code, not just to expand on it, but to fix bugs. These bugs might be coding errors or logic errors.
It's one thing to maintain a code base where the code follows a particular style, methodology, and way of thinking. But, when AI injects various blocks of code with different styles, and various random ways of thinking, it becomes dramatically harder to understand the code and find the issues. It's like well writ