What’s interesting is that the discussion in which Torvalds decides to interject in is when Pinchart brings up the ethics of using LLMs in development as Torvalds quickly points out that the goal of the project is not to be a “social warrior” (barely indistinguishable from calling others SJWs), he even makes a reference to veganism as a metaphor.
In the end, I think Giacomo Tesio’s response is the most poignant (https://lore.kernel.org/linux-media/20260716112048.2dc10a3f@hermes.development.it/)
On the other hand, legitimizing any of the several AI issues is a greet concern for the LF members [2] that includes companies such as Google, Meta, Microsoft and Anthropic that are collectively exposed for trillions of dollars over the AI bubble. It’s not that you must be “pro-AI”, but if you are not, you will be countered or silenced in every influential technical spaces.
So neither inclusiveness or developers’ alienation are relevant ethical issues in this list, just political tools or not.
Linus faithfully serves the interests of his employers, so no argument, either ethical, technical or whatever, really counts here whenever its effects could annoy them.
The LF patrons are all embroiled in the western AI bubble, so it’s difficult to ascertain whether Torvalds could ever not take a positive stance or even directly advise others to use LLMs. It’s probably not worth taking Torvalds’ “anti-ethics” arguments seriously and instead look at the bigger picture: The Linux Kernel is a titanic (pun intended) C project which hits the limits of the guarantees that the C compiler and coding guidelines can provide. Using an LLM to find memory bugs is necessary considering attackers can do the same thing (with the recent Copyfail and dirtyfrag attacking vulnerable modules). In a way, LLMs are just damage control for poor engineering decisions.
The non-LLM answer was to incorporate a stronger compiler in the form of Rust for future drivers and modules (even better would be to rewrite the kernel in Rust), which eliminates this repetitive class of memory vulnerabilities. In a more centrally planned computer science field, C would be deprecated rather than become the parasitic bedrock of everything above it.
It’d also be better if we actually had multiple kernels that were supported on the level of Torvalds’ Linux git tree rather than a monoculture where everyone is destined to hover around Torvalds’ tree because “out-of-tree” linux kernel stuff is a pain, much less a fork. Also if your foundation receives millions in grants and employment contracts from silicon valley, telling others to “make a fork” isn’t a realistic measure.
The claim that LLMs are simply damage control for poor engineering decisions is a gross misrepresentation of the reality of maintaining a codebase the size of Linux. No human can hold the full state space of the kernel in their head. Memory safety is one class of bug, but the most subtle vulnerabilities are logic bugs such as race conditions, incorrect state transitions, misuse of APIs that cross module boundaries, or behavior changes during a refactor. None of these would be caught by Rust’s borrow checker or by typical static analysis tools. These problems emerge from interactions between subsystems written by different maintainers who were solving separate problems and weren’t aware of how features might interact in negative ways.
LLMs, by contrast, can look across a far larger context and identify interactions across the entire codebase. They can trace the execution of a path through a driver, spot where a lock is held too long, or detect that a function’s contract is violated by a caller many levels deep. Humans simply cannot scale this kind of analysis to millions of lines because we can only hold so much information in our heads. Calling LLMs damage control is frankly dishonest in the extreme.
Rust is a powerful tool for eliminating issues like buffer overflows and use-after-frees in new code, but it’s by no means a silver bullet. On top of that, the Linux kernel already has millions of lines of C that will never be rewritten. A Rust rewrite of the entire kernel would be a fantastical idea, and even if that magically happened you’d still have many kinds of problems such as logic errors, algorithmic complexity attacks, or unsafe blocks needed for hardware interaction which Rust would not help you with. The reality is that LLMs help find the same memory bugs in C code today, and many of these problems would simply not be found otherwise.
The whole idea of having multiple supported kernels to break the monoculture is likewise fantastical, and ignores the sheer amount of work that goes into maintaining a project of that scale. It’s also completely orthogonal to the LLM question. If we had ten kernels then each would still be a giant codebase needing the same kind of automated analysis.
Dismissing LLMs as damage control ignores the fact that much of all engineering is damage control, and the real question is which tools give us the best return on effort. LLMs currently provide a unique ability to surface hard to see interactions that no other tool catches.
Reading the whole message, his position seems totally reasonable.
yup, it’s a rational stance on a new technology
I think calling
I expect maintainers who want to act on sashiko reviews to triage and verify them first before bothering authors, yes.
“a very anti-LLM position” is absurd and a reason for concern.
You can’t say that here.
I one got banned from an anti ai com for mentioning video game npc ai.
There’s obviously a huge difference there. Video game NPC AI is closer to a complex flowchart than to genAI, at least traditionally - who knows what’s coming in the near future.
I’m not sure if you were being unreasonable by drawing the comparison (or by using the false comparison to intentionally instigate) or if they were being unreasonable by equating the two and banning you, but someone was.
If i recall it was as an example of how “ai” can be more then just llm chatbots. I thought the com was specifically against llms but they didn’t like general machine learning either.
Long time ago though.
It’s going to be a sad day when Linus finally leaves the kernel
Based. But the AI slop machine is something to be aware of. AI can get you 90% of the way there nowadays but without the remaining 10% error checking you’re not gonna have a good time.
I’d argue it’s more than 10% checking. You really do have to be engaged in the process, and you can’t farm out thinking to the LLM. It’s a great tool for generating code, but you have to be making conscious decisions at the developer. My process has been to come up with a step by step plan, where there are clear and focused deliverables at each stage, and then do commits for each one and review the diff. This way I have a clear context of what the task is doing, and a reasonable amount of code I can read through to do a proper code review. And it’s easy to actually test the functionality out to see that it’s working. If you take this approach, then the tool really can save you a lot of time.
Yeah, my impression is that LLMs are okay at writing code, but not good at software architecture.
It’s an advanced autocomplete, not something that can think, at least not on the level that humans do.
Yup, and as you mentioned in another comment, this tech is being marketed as something that it’s not because companies pushing it want to convince other companies that it will replace human labor.
the issue on ai is actually about tradition vs. modernity, in most cases. if you add up to that the fact that it’s coupled with automation, which a lot of people consider a bad thing because it eliminates jobs from the market. it was capitalism all along…
exactly, pretty much every serious argument against this tech boils down to capitalist relations rather than the tech itself
Imo the biggest problems with LLMs is that they’re being sold as something that can replace humans. The people peddling the tech are trying to hide the fact that it should never be used without HIL (human in the loop). Not for anything important.
Absolutely agree, there is a big disconnect between how these tools can actually be helpful in service of a human using them, and the narrative AI companies sell to CEOs about replacing human labor with them. This was a good write up from Doctorow incidentally on the topic. https://pluralistic.net/2025/09/11/vulgar-thatcherism/
How separable are the two when scale is needed to run llms effectively? If you’ve used ollama on your own computer you know the limitations of alternatives. When you use big tech’s offerings you are helping to train their models further. Adding to their control. My suggestion isn’t necessarily abstention, but here tech and capitalism seem to be inseparably intertwined.
We’re basically in the mainframe era of this tech, but if you have used local models then you know that progress has been absolutely breathtaking in the past year. Qwen 3.6 27b that you can run on a laptop is straight up better than frontier models that were available just a couple of years ago and required a data centre to run. Also, you don’t have to use American models. You can use open models from China, they’re very capable.
Socialism takes advanced, industrial production beyond the profit motive and into a consciously planned and pro-social direction. You can have industrial scale without having capitalism.
It was a totally measured and well thought argument. Not singing Nirvana nor Apocalypse when it comes to AI but embracing technological progress in a measured way
