I don’t disagree that LLMs are very useful when debugging and such, but it’s disappointing that he talks down to people trying to do the right thing and fighting so we have jobs and we don’t end the world. I’m not out there doing any of that but I appreciate folks that do.
Thanks fir directly linking the mailinglist response from Torvalds instead of linking to an article referencing it! Always better to point to the source of truth.
Overall I agree with Torvalds standpoint: If you do Not want to use AI -> that’s fine. If you want to use it and you are producing good results -> also fine If you want to use it and produce Bad results (e.g. do not review/Check theory output) -> Not good and your contribution will be rejected
It is a tool. Use it as such.
Of course this does not cover the economics of AI and If these huge Investments will pay for themself in the Future .
If they can’t pay for themselves they will fail. I wonder which ones will survive, several are spending far more than they earn and are racing to make the best AI before they run out of money
There’s a significant possibility that the technology will not require enormous amounts of compute to train sooner than some might think. None of these companies want to spend money they don’t have to.
Well fucking said, Linus.
his words are what i experience in “real life”. only online do people get religous and emotional to the point of throwing out any critical thinking.
Mhm yes, just let us brush of the valid concerns, job loss and health risks and brain anthropy associated by LLMs and data centre. Its just the terminal online luddities talk about.
brain anthropy
case in point ;)
Its a hard word okay? :)
case in point
Example B
Yep, I’ve never experienced that kind of anti AI zealotry IRL. (Because if you interpret someone saying “only online are people extreme” as “let’s brush off the concerns” that is zealotry)
deleted by creator
I don’t think many LLMs have been trained on much code that is closed source
Oh I’m pretty certain none of their code is written by AI. If you read what he says carefully, he’s most impressed with how it can identify issues and guide the project to what needs attention. They’ve loudly complained about low-effort AI submissions previously, so AI code is not what he’s really talking about here.
AI is absolutely killer at issue identification and is getting better. He knows, as they all know, they won’t be able to survive without it going forward because if it can identify vulnerabilities for them it can identify vulnerabilities to anyone.
deleted by creator
How do you verify that no human has given a proprietary snippet? 🤔
deleted by creator
It’s the same as when you use any generative tool: if your name is on it, it’s your problem, tools aren’t accountable.
The problem is that in this case human delivering the code would have no idea that the code is proprietary.
Doesn’t matter, that human is still the responsible party putting it there.
The whole premise here is weird anyway; unless you’re using a RAG designed to quote code, the chances that you end up in a perfectly identical environment needing exactly the same operation, AND it’s proprietary somehow, AND you’re the one working on it and wouldn’t know there was a proprietary feature (whatever that even means here) is astronomically low.
If you were writing a book with a handful of other people and you used an LLM which PERFECTLY reproduced a chapter of another book on the topic without attribution, whose fault would that be? It’s not like there are software 2 liners that are just off limits for use, so besides the fact that it would be crazy unlikely to reproduce a perfect copy of a huge block of code, the chances of you working in the Linux kernel for example, getting your commit added to the official version, and not knowing it’s proprietary by looking at it is, IMO, so unlikely it’s not an argument against culpability.
Maybe if you’re vibe coding an app from scratch on your personal GitHub or something (even then the odds of perfect recapture are stupid low), but it’s not really possible or a good argument here.
Edit: I guess the obvious exception here is if some LLM provider DID train a model to use whole pieces of code and strip attributions and licensing. In that case, I don’t see how they survive getting sued into the ground.
While a program is eligible for copyright, anything small enough to be considered a snippet likely wouldn’t be unless it is sufficiently expressive or original.
as long as they don’t vibe code the kernel 😛
