Linus Torvalds on LLM usage in kernel development
lore.kernel.org
lore.kernel.org
Agentic cli based coding "got good" (around Oct/Nov 2025), and after just a short trial period, most devs in our company were convinced enough to start using it. The graph from "not using it much" to "95% of code written by AI" was incredibly steep. It was obvious from around that point, but particularly from around Opus 4.6 that this was actually, very, very good, provided you understand the failure cases and worked to eliminate them.
Ever since then the story has been about the % of developers denying it can write anything decent dwindling, and right now there are very few left who say this. Places like reddit and HN seem to attract a lot of them though, so weirdly, despite in the real world most professional software developers using these tools prolifically, it can seem as though it is 50/50 here.
When even someone like Linus Torvalds (the most crotchety of crotchetiers) is accepting it, you must know you're on the wrong side by denying it at this point. Surely the cognitive dissonance must be grating?
You can also read up here: https://lore.kernel.org/linux-media/20260715125058.3692aa03@... that this thread is about code analysis not generating it and the signal/noise ratio. It's not like his opinion is shared by everyone because these people actually have to look at the PRs generated and filter out all the noise.
The issue is not AI finding bugs is bad, it's that usually a maintainer gets offloaded 1000s of automatically generated PRs where 99,9% are BS and is then expected to read and check and understand all of them, immediately merging the 0,1 actually pointing out a real security issue.
You working in some "we replaced MS Access with another high-level language tool"-place is so far off this level of engineering i'm not surprised you only took the information "AI good" from this then told everyone else about how wrong they are to not use it.
You might have a point buried in your post somewhere, but I can't find it because your personal attack obscures it.
See if you use your time to write specifications so AI will stay on guardrails - why not just go all the way to memory layout and functional programming? It's great to know what actually happens on the machine.
With natural language or the 4323th framework that makes things "easier" you just end up yourself not actually knowing what your intent is your machine should fulfill. Shouting at some AI the stuff your boss shouts at you? Not exact logic and what people intend today may contradict what they wanted yesterday.
Having domain knowledge and knowing how to translate peoples immediate wishes, predicting what they will want after that into building solid software is a skill you can grow into, but switching the toolset all the time that hides the magic keeps you from learning what was true in the 70s and is true today.
He's not writing code, he merges 200 PR in 2 weeks. From this POV an AI eases his workload, but he makes no difference whether he dislikes AI generated code or is just fine with found bugs by it.
Still he has no idea whether the person suggesting a feature or fix has any actual knowledge about the code so when problems come up there's no one there taking responsibility and i think that's shortsighted.
I think Linus' position even pre-LLM-coding – basically not trusting any code he gets, and not writing any code himself, just reviewing a ton of it – is strikingly similar to how most people now do "software engineering" (or vibe coding/agentic engineering or whatever you want to call what we're doing nowadays).
From his perspective, the only things that changed is that he can more quickly/better review code and find bugs, and maintainers below him are somewhat more annoyed due to the influx of AI slop.