Linus Torvalds on AI, Junk Patches, Humans, and Godzilla (zdnet.com) 17
Linus Torvalds once said LLMs might bring a 10X increase to programmer productivity. But speaking at Open Source Summit India 2026, he now says that number was "not scientific,"
reports ZDNet. "That was pulled out of my ass number, obviously."
Today, he continued, "we're at the point where hopefully it creates more productivity than it takes away," but "we certainly saw more junk being generated by LLMs than we saw useful code up until the like early this year.... it can actually be a huge drain on resources when it takes humans a lot of effort to figure out that, hey, this machine-generated report was not true." Even now, he said, "most of the good ones require more than just the LLM," because "we've had to push back quite a bit... if you find a bug with an LLM, it's not enough to just ask the LLM to make a bug report and then throw it over the fence to us. We want to see a suggested patch; we want to see the human who ran the LLM act as a kind of back-and-forth."
Torvalds described many AI-generated patches as "mindless band-aid kind of patches... they may fix the immediate problem, but the kind of bug remains, and it just is waiting in the hallway to hit you in another place." For his own toy projects, he uses LLMs as prototypers: "I use them as a way to prototype things... quite often the code is not usable in that form, but it's a great way to try something out," while insisting that for kernel-level fixes, "LLMs, in my experience, have not been at that level yet."
Torvalds acknowledged that some AI-found issues have been "absolutely, stunningly, I mean, interesting in a painful kind of way," especially security problems that "show up in the technology press two days later." Despite the embarrassment, he said, "I'm very much not a shoot-the-messenger kind of person. I think we're much better off with LLMs finding bugs, even when they are embarrassing, and they are things that we should probably have found two decades ago."
Torvalds also said he's using AI "for my own toy projects... Every time I travel to some new place, and this is the first time I've been to India, I send the kids pictures of where I am, and for some strange reason, Godzilla seems to follow me around and gets added to those pictures."
ZDNet notes that Torvalds concluded, "There are many useful and less useful uses for AI," and "I think Godzilla is a great place to stop."
Thanks to Slashdot reader joshuark for sharing the article.
Torvalds described many AI-generated patches as "mindless band-aid kind of patches... they may fix the immediate problem, but the kind of bug remains, and it just is waiting in the hallway to hit you in another place." For his own toy projects, he uses LLMs as prototypers: "I use them as a way to prototype things... quite often the code is not usable in that form, but it's a great way to try something out," while insisting that for kernel-level fixes, "LLMs, in my experience, have not been at that level yet."
Torvalds acknowledged that some AI-found issues have been "absolutely, stunningly, I mean, interesting in a painful kind of way," especially security problems that "show up in the technology press two days later." Despite the embarrassment, he said, "I'm very much not a shoot-the-messenger kind of person. I think we're much better off with LLMs finding bugs, even when they are embarrassing, and they are things that we should probably have found two decades ago."
Torvalds also said he's using AI "for my own toy projects... Every time I travel to some new place, and this is the first time I've been to India, I send the kids pictures of where I am, and for some strange reason, Godzilla seems to follow me around and gets added to those pictures."
ZDNet notes that Torvalds concluded, "There are many useful and less useful uses for AI," and "I think Godzilla is a great place to stop."
Thanks to Slashdot reader joshuark for sharing the article.
Leading to simple programming languages (Score:3)
The question has and is "How can we make programmers and/or writing code more productive, less bug free, with less security issues?"
That is on the output side the process.
We need to ask "How can we design programming languages, system APIs, application structure, data structures and database schemas so that it is a lot less likely to make a mistake in logic, modularity, data flow or data design?"
The follow-on to the second question is "How can we prevent the feature accumulation making programming languages
Re: (Score:2)
IT is fundamentally immature. You get crap like that with immature technologies. Maybe in 100 years things will look better, bit with the active efforts by the main profiteers to stop that from happening (because it means liability, qualification requirements, things becoming commodities, etc.), it may take quite a bit longer.
Re: (Score:3)
We need to ask "How can we design programming languages, system APIs, application structure, data structures and database schemas so that it is a lot less likely to make a mistake in logic, modularity, data flow or data design?"
"How can we design a (programming) language in which it's impossible for people to think bad thoughts" has been a wished-for silver bullet since at least the 1960s. This is why the industry uses SAST(ools), dynamic analysis, AI analysis, etc on whatever they've currently got, because after sixty years of not being able to produce this silver bullet people have accepted that you need to try something else.
Re: (Score:2)
Look at any "low code" approach before. Either you can only implement toy projects, depend on people building complex backend modules you need in other languages for your next project, or have a feature creep until it is almost a full programming language. And you still do not avoid logic bugs. A compiler won't find that you're missing authentication, because it doesn't know if you want to run the code on localhost only or on the web.
Re: (Score:2)
A compiler won't find that you're missing authentication
Nor should it need to. That's the concern of the web daemon.
Separation of concerns. Modularity. Keep it small and simple.
Re: (Score:2)
You're missing the point, that is some (most?) bugs are logic problems and won't vanish if you simplify the programming language.
Sure, but ... (Score:2)
Linus Torvalds on AI, Junk Patches, Humans, and Godzilla
What are his thoughts on GOTOs [xkcd.com] - and Velociraptors?
Words to live by (Score:2)
"I think Godzilla is a great place to stop."
Yup. Let AI make lots of Godzilla videos and cat videos, then call it a day.
I'm waiting for... (Score:2)
So far, all I see is stories of productivity.
Re: (Score:2)
Humans have never been able to accomplish this. LLM are trained off of human output.
Why would AI tools be able to do something that humans cannot do?
Agentic coding tools are definitely 10000% faster than a human and are about as accurate as a very experienced senior developer. But a senior dev is only as good as the specification they are writing to, just like with agentic coding tools. (note, I am talking about agents using frontier models like gpt codex and claude opus so if you use other
Re: I'm waiting for... (Score:5, Informative)
No, they are not as accurate as senior devs. They still hallucinate, peppering their work with some elementary school mistakes no human dev would ever make. The human supervisor needs to be able to catch those.
Re: (Score:1)
I think you need to catch up a few generations of LLM. Your comment is like telling "AI cannot do fingers".
AI images are still not perfect that's why people are bickering and commenting on smaller and smaller AI artifacts to make clear they can still tell, but fingers are fine by now. In the same way "elementary school mistakes" no longer happen in code LLM by now, even though you can still recognize LLM code if you look very closely and people are still working on LLM code becoming bettter and human skin l
Thinking (Score:2)
Although LLM's have been trained on the best thing that we could find - Humans - oddly, LLM's don't think in the same ways that we do.
An odd byproduct of stuffing the entirety of the human way of thinking into the LLM's is the LLM's ability to think in unusual ways. Because it is inherently logical, it combines some human rationality and finds both the good and the bad in a codebase. It also hallucinates just as much as humans do.
Hallucination by an AI is called lying by Humans. And Humans have lied quite a