Linus Torvalds on LLM usage in kernel development
134 points by thang
134 points by thang
So Linus tries to sell the idea of using LLMs and when Laurent basically says "no thanks", Linus says we're all about technical reasons and if you don't have them, you have nothing, don't try to sell me your personal beliefs about LLMs (after doing that himself) which Laurent rightfully calls out.
This is a nice confirmation that appeal to authority is probably one of the most shallow and weakest arguments one can make and invoking Linus' name doesn't somehow make an argument stronger.
This is a nice confirmation that appeal to authority is probably one of the most shallow and weakest arguments one can make and invoking Linus' name doesn't somehow make an argument stronger.
Really? Linus has arguably led the most complex real-world software project of the past 30+ years. He managed to balance complex contributions from technically brilliant individuals, competing corporations, legal considerations and social issues in an open-communication style, all while keeping the project moving forward successfully. That's no small feat.
Now he made a judgment call, as he has many times before. You may disagree with that decision, but calling him "shallow and weak" is low IMO. If anything, he earned the right to make such calls and that's part of the reason most won't dare to cross him.
Umm, I didn't call Linus "shallow and weak". I called the argument from appeal to authority shallow and weak, regardless of which name one might invoke because I've read people say "if it's good enough for Knuth and Linus, it's good enough for me".
appeal to authority is probably one of the most shallow and weakest arguments one can make
I think the real question is whether Linus is in a position to decide the direction of the kernel development as a whole. If the answer is yes, then whether the argument is technical merit or appeal to authority is non-consequential.
I think the real question is whether Linus is in a position to decide the direction of the kernel development as a whole
I mean… mainline releases are cut from torvalds/linux. So long as that is true, Linus is in the position to decide the direction of the kernel development as a whole.
Linus has arguably led the most complex real-world software project of the past 30+ years.
Over those years, LT has driven contributors away through personal hostility and verbal abuse (paywall bypassed by turning off javascript). By doing that, he may have harmed the Linux project on a technical level, if only by creating a hostile culture for any would-be contributor who didn't want to be on the receiving end.
I don't follow the kernel mailing list, but from what I know he's become less directly hostile after that New Yorker piece from 2018. Still, reading the linked mailing thread, I see LT as having the emotional reasoning of an average user on an old-school programming forum (which is not very high), with cliches like (in my own words) "I had this experience [with LLMs] and so can you if you just try", ignoring that experiences can differ and be equally valid, "just get good [with LLMs]", which is condescending, and using "this is a technical discussion, your opinion is irrelevant" to shut down a discussion that hinges on his own opinion.
If anything, he earned the right to make such calls
Bluntly, his outstanding talent as a software engineer, and the rare in scale success of his work, has earned him the kind of celebrity that lets him get away with pretty much any call he makes. That he has also torn people apart in public, and in doing so may have significantly damaged people's open source careers is also likely
part of the reason most won't dare to cross him
I don't think that's a good thing for the project.
Everyone is fallible. I generally think Linus is on the right track with decisions in open source. I don't disagree with his conclusion that the tools are effective, or even that they're better for some tasks than many of the other solutions we have for mass review at the moment.
The horrible thing about our situation is that everyone arguing for both things is right in some way. The tools work, yes, but LLMs and derived tools are a comprehensive manifestation of the problems of the industry, many of which Linux was originally created to sidestep. It's disappointing to see that he is now so easily swayed by the appearance of convenience, and his willingness to ignore the problems that lead others to not want to use these tools in the first place because of that convenience.
many of which Linux was originally created to sidestep
Wasn’t Linux created just for the fun of building it?
When I actually read that thread it just boils down to email filtering/spam that somehow people feel icky about an AI system doing work with their patches. It's just another automated system to me. Those two FOSS rules shouldn't apply in this case and his equating the inclusive terminology commit acceptance as a similar issue is just silly.
And no, AI isn't perfect. But Christ, anybody who points to the problems at AI had better be looking in the mirror and pointing at themselves at the same time.
Because it's not like natural intelligence is always all that great either.
Natural intelligence isn't burning the planet. The use of natural intelligence doesn't require actively funding the torment nexus.
Sure, the social angle of working on open source is important and often a very motivating part of the project, but in the end that's a side benefit, not the point of the project.
[...]
In the kernel community we do open source because it results in better technology, not because of religious reasons.
I do think this is an interesting point to make though, because there's a big difference in culture between different open source projects, and I imagine the priorities of a project in this regard are a decent predictor for their stance toward LLM usage. Other projects place a much stronger emphasis on community, and it's not surprising that those projects would be more strongly biased against AI.
Natural intelligence burns more than 8K KJ per day and that energy doesn't come free from environmental footprint which is far greater. It also directly led to this planet "burning". The main thing you can argue for it beyond us being fond of, well, us, is that it is a sunk cost and we may as well use it.
Are ... are you equating human existence with training and operating LLMs? And somehow implying that for LLMs energy "comes free from environmental footprint" whatever that means?
I feel like personal choices made by individuals are a completely different kettle of fish than equating resources used by training/operating LLMs with having children. Yes, both contribute to environment degradation, but ethically they're as different as apples and, not even oranges, but something like dumbbells.
People can make personal choices to not have children and it's still reprehensible for someone else to proselytize to others that "they should have less children", if that makes sense. While it's not morally reprehensible to tell governments and CEOs to stop their feeding frenzy of propping up a technology which consumes resources disproportionate with its returns.
No, I'm not and I don't have creative enough mind to read this in my words either.
What I am saying is that we were burning this planet well before LLM became a problem and saying natural intelligence isn't burning it is not true even if living like use of tech would be reduced only to its energy consumption. This doesn't excuse LLM, other technologies or us and focusing only on energy use is not particularly useful.
Nothing comes free. Nature mostly doesn't "care" what the source of CO2 is.
Yes, but LLMs are the newest and one of the more virulent consumers of energy, water, electronics, what have you, and, like I said previously, didn't yet bring the return of investment expected by its proponents.
People investing in "AI" companies keep expecting that inflection point where suddenly it will be able to take over a large part of humanity's intellectual endeavours, and lead everyone into a utopic, care free future. Yet nothing materialized so far except for greed and poverty for the people whose jobs got displaced in the process.
For regular people like me it feels like these people have fallen pray to sunk costs fallacy and are willing to burn the whole world in the hope that someday "AI" is going to magick a way to undo all the harm.
Nobody in the history of humanity equaled this hubris. Not agriculture, not having kids, not eating meat, not even gasoline powered vehicles. LLMs are a net minus on everyone's benefit outside of the 1% that invest in them at break neck pace.
Trying to present them as "yet another" in a string of environmental issues is missing the point by a very large margin in my opinion.
I agree with pretty much everything you've said except the last paragraph. I am painfully aware of negative impact because I am in the middle of a job hunt and AI has clearly — for me — negatively impacted hiring beyond just reducing the number of opportunities. I am certain there are fewer remote positions available because of it and I expect my search to take a while. Societally that's a minor problem that I wouldn't even bring up generally compared to the rest I expect we will all go through.
I reacted to the line that I read as suggesting we do not have comparable negative impact which to me is absurd since we are living in anthropocene epoch. I do think environmental issues are real, but as someone else said like dubious copyright not nearly among the main issues with this technology.
My original comment admittedly wasn't very helpful and I underestimated how much will be read into it.
Maybe vegetarians should get to use LLMs guilt-free then?
Sure, and because I'm such a nice person in general, I should get to punch annoying people in the face guilt-free /s
People making life decision (vegetarianism or something else) for environmental reasons are likely sensitive to and willing to avoid/reduce many other environmentally-problematic things, like power-hungry AI.
We all need to eat, we don't need AI.
This is a bad argument. We all need to eat, sure—but we all don't need to eat meat! We don't need AI either, but it's far less environmentally problematic than meat production.
I think the only thing I can add here is that I don't think this pans out long term. I don't think you are arguing this, but I thought it would be useful to mention.
Currently we don't have an effective way to update models with new information. Sure you can give it a ton of context about a library it doesn't know about and hey presto, it can use that library.
But that (A) uses up a lot of context, and (B) results in worse performance than having that library in the training data.
It won't pan out long term, models will perform worse and worse as time passes, and the only real solution (short of a new innovation) is to retrain a new model from scratch.
I find it really hard to view this as a good faith argument. There's clearly a difference between AI data centers measurably accelerating the climate crisis, and humans needing energy to think and survive (as they always have). I know that you know the difference.
That's what I don't like about modern debating. People can't stay topical, and instead bring in every semi-related topic (which are highly contentious and debatable on their own) and dilute discussions to the point of nihilism.
In the end, the real problem (the gang of corporations forcibly planting it on everybody) laugh all the way to the bank while the laity are busy grabbing one another by the throat.
Natural intelligence [...] also directly led to this planet "burning".
Do I understand your argument is that since we made mistakes in the past, we should keeping making them?
The way I read the arguments against the LLM hype-machine, is that we are consuming way to much energy for the benefits it brings. The technology and business around them are a net negative for humanity. If people disagree with that statement, and they want to discuss the topic, I think it's fair to address those concerns directly.
To what end though? Why not target the companies actually bringing about this technology instead of consumers (a lot of whom are being forced to use it at work)? This simply feels like pointless virtue-signalling.
a lot of whom are being forced to use it at work
You can say "no" to things at work.
Not saying no to things leads to things like Palantir and software written for guiding missiles. As an industry, developers need to start saying no to things and grow a back bone. I think there should be a programmer's guild with a clear deontology that should include not working for purposes accelerating Earth's decline or murdering of others.
You can say "no" to things at work.
Not saying no to things leads to things like Palantir and software written for guiding missiles
Here’s the problem with what you’re proposing for a programmer’s guild: you can have a small one that follows your very specific set of yes/no rules to a T, or you can have a large one, but I don’t think you can have both. I agree with you on Palantir. I disagree with you on missile GNC. I work on technology that has serious potential to be dual-purpose; my employer is currently 100% in the civilian market but could change that tomorrow and there’s nothing I could do about it with respect to my past work, it’s theirs.
To make the thin-slicing problem very clear: as a strong guideline (I won’t call it a rule or ethical hard line) I will by default refuse to work on technology for law enforcement, even completely unarmed technology (e.g. tactical situational awareness). I have no qualms by default working on military technology, but could have qualms about specific types of military tech (I wouldn’t, you know, work on SmartNapalm).
My examples were pretty spur of the moment, I would imagine that if such an ethical body would at some point exist, the rules would be selected after some more consideration than I gave it.
Anyway, the implication is that ethics are a thing and keeping to an ethical standard, whatever that is, implies sometimes saying "no", and, as you underscored, sometimes being selective in the work that you choose to do.
You can say "no" and be replaced immediately by someone who will say "yes" of which there are plenty.
There's a wide spectrum of other possible choices to make and I don't begrudge someone who needs to make a living and didn't ask to be in this situation having to weigh them.
The significant difference between you saying yes to doing the bad thing or you saying no to doing the bad thing and someone else stepping in to do the bad thing is that in the first case you did not do the bad thing.
The term for this - holding your principles despite any consequences - is martyrdom. You can choose to prioritize your ethics and be a martyr but I don't think it's so cut and dry that everyone should, especially if they have dependants. Until we have an effective way to decide collectively not to do things, individual choices like this won't have much impact on anyone else.
Edit: There are degrees here. Someone mentioned Palantir and frankly I would just not put myself in the position of working for them, and can't imagine being desperate enough for that to change. "Using AI because of a sudden employer mandate" feels like a very different degree.
I agree that there are degrees, but personally my lodestone is my own moral compass. If I feel that what I am being asked to do is wrong or harmful or I think that others whose opinions matter to me would feel that way, then I don’t think I should do it regardless of whether I think it’s likely that someone else will.
be replaced immediately by someone who will say "yes" of which there are plenty.
We all know how how the prisoner's dilemma goes.
I agree these are structural issues with perverse incentives for individuals. At the same time, I do not want to accept the argument that we cannot, at least try to, fix or improve a bad situation. The concerns surrounding this technology seem valid to me, and I have mostly failed to have found arguments to assuage them. A lot of it seems to be rather dismissive, akin to "this is inevitable, I accept it might hurt you; stop whining and help us make it happen".
Do I understand your argument is that since we made mistakes in the past, we should keeping making them?
I read it as suggesting that a blanket ban on this energy usage is too broad.
If we're to prohibit LLMs due to energy usage, do we also prohibit contributors using AC, or driving to work?
Blanket bans on particular energy usages isn't the approach that's being taken to reduce emissions, so it's not surprising that a project like Linux isn't willing to take on this type of approach.
There are other harms of LLMs that might hold more weight, but in this comment I'm specifically replying to the comment thread re energy usage.
I read it as suggesting that a blanket ban on this energy usage is too broad.
I am afraid I do not understand how the thread could be read that way. I cannot remember the last time someone argued for a complete ban on the technology. On the contrary, a lot of the voices against the mainstream LLM expansion do acknowledge the benefits it brings, for niche cases (for example, medical and pharmaceutical research with protein folding). What I feel people are against, is the push to use this technology everywhere, regardless of whether it fits, and how much that costs, both in energy, money, or raw materials.
If we're to prohibit LLMs due to energy usage, do we also prohibit contributors using AC, or driving to work?
This feels like a straw-man to me. Even if we agree that people are arguing for a "blanket ban on this [LLM] energy usage", how do you make the logical jump to extend that to other energy usages?
It's really hard to accept the environmental argument at face-value when people are openly lying and trying to mislead others about water usage while saying little about lawns.
As for the power argument, there is the fact that the #1 country providing AI inference and training right now has backed out of its promise to lower emissions in general on the backdrop of it historically failing to provide affordable public transportation, effectively forcing most of its citizens to use cars while actively promoting wasteful tank-sized gas guzzlers as a lifestyle. And promoting airplanes and trucks over trains.
Inference currently happens mostly during waking hours, so how about lobbying for those data centers to instead put up enough solar panels to cover their needs and then some, and mandate taxed surcharge for inference during night as well as extra tax on training through the night?
Wouldn't that actually be beneficial to everyone? Electricians can make a living, everyone gets cheap inference, little CO₂ gets emitted and who knows, maybe using those models proves economical in the long run. Or not and you get cheap power during summers to run your AC from.
And you can use that tax to build more transformers (the electromagnetic kind) and wires for power distribution you desperately need.
so how about lobbying for those data centers to instead put up enough solar panels to cover their needs and then some,
Every "clean" watt that you direct towards data centers is a watt that you could have directed towards shutting down fossil power plants instead. As long as there is brown power on the grid at all, every load in isolation is effectively brown.
(And that's ignoring that there is no such thing as completely green power. Even solar cells need to be manufactured, installed, and maintained.)
Wouldn't that actually be beneficial to everyone? [...] everyone gets cheap inference,
Green or not, I wouldn't count "everyone gets cheap access to the misery machine" as particularly beneficial.
So what do you propose our friendly US peers should actually do about the fact that their oligarchs have decided to build data centers using rather drill, baby, drill non-green power sources that are breaking our common ecosystem?
I know what I propose: limit damage, quickly. Something that is manageable under the upcoming democratic presidency. I am pretty confident our US peers won't be picking up cough, uh, woodworking and solving the underlying problem anytime soon.
Maybe they would at least try to stop killing our elderly and babies with those heat waves. I don't really care about their living standards anymore, it's their choice.
You can just shut them down.
If you can impose conditions on their planning permission, or their grid connection, or whatever mechanism you'd propose for actually enforcing that solar mandate… you could just refuse it outright instead.
You don't have to give them an out.
I cannot do that. I do not have that power. Neither do you. And neither will the upcoming administration. Please realize that those people over there truly believe that the billionaires who are steering the economy deserve to be in power.
As in, they believe those should be allowed to spend their money. They totally miss that that means ordering people to do stuff under thread of homelessness and hunger.
I always tell people that "dubious copyright" and "environmental harm" are the two weakest arguments one can make against LLMs and this is a good example why. If a person questions the environmental harm of AI data centers and LLMs, apparently they'll get questioned about the environmental impact of their own existence, which is really something.
Well, do you make number go up as much as the llm? No? So why should the system allow you to live rather than putting those resources into the llm? /s
What would you suggest as strong arguments? I generally agree with you that those are two weak arguments. And much to my dismay (I was anti-LLM on a quality basis for a long time, until I wasn’t), I have been genuinely stunned at the ability of these tools to produce excellent high-quality work with appropriate guidance and steering, whether that’s producing new code or doing analysis on existing code or datasets produced by existing code.
Linus is right and isn't. Pretending it's all about technology is a fallacy, it never was. Linux started for licensing reasons, Linus said (asked about MINIX).
So it was about choice from the beginning. And still is.
Technology is the means, freedom the end.
I think the reality is that there is not really room for meaningful compromise here.
The maintainers of the project believe that the agentic code review is helpful, stands on its own, and does not create ethical problems. If your belief is that you should not have to read that output, then you're asking them to decide which is more valuable: the ability to do what they think is best for the project, or your ethical opinion that they do not share.
If it were merely about using the tool on your own, you can coexist--no one really knows what you do on your own computer. I am not saying you should use LLMs on a project like Zig that rejects them. You should be honest--what I'm saying is that a project that views AI as acceptable can have some people using it and others not.
However, for shared infrastructure, shared code review? There has to be rough agreement. The same goes in reverse--if a project rejects AI code review, and I have my bot review things, I am breaking the norms of the project and will probably not be welcome.
So if you object, and there are enough people like you, you may carry the day, and they may abandon the tool because they do not want to lose you. Or perhaps you'll even convince them to share your ethical concerns, though people have a way of digging in their heels. Otherwise, you'll have to decide whether to stay or go--swallow your objections, or stick to them at the price of not participating.
I predict projects will continue to bifurcate on this question. As time goes on, it will more and more be the case that they'll be free of AI code review, or it will be enforced.
I think that what I have written above should be common ground to people who like AI and people who reject it.
A good analysis of the situation, and points to the heart of these debates: the network effect of the tool use or boycott is what makes it a problem.
Nobody arguing about, say, vim vs Emacs cares what their co-developers use to make their commits. They may look down on people in the other camp, but it's not a reason to leave the project as it has no meaningful effect on them directly, whereas LLM usage or boycott typically does affect everyone.
Mechanically, I don't see why that'd be the case? To accommodate someone who doesn't want their code to be LLM-reviewed, they can simply say "i opt out" and then they can have a human maintainer review it instead. Unless you're not expecting to have any human code review the cost doesn't really go up, and assuming llm review results in better code the overall codebase quality gets only as bad as it was pre-LLM.
I could see a culture fit argument, if this were a consolidated dev group. but lots and lots of different people need to contribute to the kernel, even on a drive-by basis.
To accommodate someone who doesn't want their code to be LLM-reviewed, they can simply say "i opt out" and then they can have a human maintainer review it instead
Or they can have nobody review it instead, if there is not even a trace of spare human maintainer capacity left to accommodate such requests (which, incidentally, was the reason LLMs were adopted for review in the first place).
The solution is to make sure those LLM tools help maintainers instead of just causing them pain. There's no question on that side.
People have complained about the GPL and its “viral” effects for decades, but now generative AI has effectively hijacked the whole software industry and no one bats an eye.
Think about it, devs who would rather be working on the kernel have now to make a decision and leave, fork, or accept to be guinea pigs for OpenAI/Anthropic/Google/Meta/X/whatever.
No? They can just continue to not use any LLM tooling if that is their preference.
You use LLMs. I don’t.
You submit a patch. I review it and request changes.
You will feed that into an LLM.
May I choose not to review your LLM-generated PRs? Yes. But if overwhelming majority of them are, too, then I have to make a choice on what is the best course of action as I’ll cease to be an effective contributor to the project.
That seems like the other person using an LLM and not you, but certainly if you not want to review such patches you can also not do that as you say.
By those same standards, I’m talking alone at the moment and your input makes no difference to the conversation.
...no?
You might be right, I might have skipped some steps.
An agent bot submits a merge request.
I am too lazy to check the author—I thought Claude Bot was a real guy with a silly surname giving its first shot at contributing to the kernel.
I request changes, Mr. Bot replies back.
Did I help the LLM then?
Does it matter if a human intervened during the process and Mr. Bot was kind enough to submit the patch for him?
Does it matter if a human was kind enough to send the patch on behalf of Mr. Bot?
Edit: I changed the question from “did I use an LLM?” from “did I help an LLM?”. ESL && terrible at proofreading.
If someone hooks up an agent directly to the mailing list and submits slop patches and replies directly somehow I hope that would get blocked as spam.
That is a good reply to the easy question, I’d like to hear your opinion on the hard ones.
If a human uses an LLM related tool during the writing of their patch, I wouldn't consider the reviewer to have used an LLM any more than I would have considered the reviewer to have used Emacs just because the author did.
However if the reviewer feels differently they could refuse to do a review on such a patch. The question of how the reviewer feels is more relevant than the question of how I might feel.
The question of how the reviewer feels is more relevant than the question of how I might feel.
Good that you can empathize with them.
Now if the absolute majority of contributors use LLMs, does it change the fact I stated in my original message, that they have to make a decision on whether they must leave, fork, or accept to be guinea pigs for those companies?
You said it doesn’t matter how you feel, but how they do.
I’m not asking what do you think is fair. I’m asking if they do have any options besides these three. That was my original statement.
Btw replying to myself bc this is getting too abstract so I’ll drive the point home as we’ve already entered “it’s just my opinion” territory in defense of generative AI companies.
AI is just a tool, yeah. Well, a hypothetical computer made of human heads in a jar screaming to get out would also be a tool.
Would some devs still review a kernel patch written in such a concoction? Apparently yes. And we shouldn’t be bothered by that.
OK. I might have gone too far, didn’t draw a clear line. AI is nothing like that. It’s just like Emacs. Also, The FSF will aggressively scrape the internet for content. And there is no opt-out.
You don’t want the FSF, er, AI companies to scrape your data or to upload the data I will copy-paste into my agentic UI either today or in the future? Well, too bad. You’re entitled to your opinion as long as you’re willing to review my damn patch.
The option to refuse to review any such contributions? I don't see how there being a lot of them removes this option.
So “leaving” is the answer. Thanks
Edit: forgot the ELI5 section, wrote it one hour later. I don’t expect a reply
“Kernel devs are smart, those who don’t want to review LLM patches will automate it with a (allow|deny)list”. That’s what we think will happen.
Reality is, we should not expect a “hard” LLM/no-LLM split based on individual kernel components if an individual maintainer is allowed to use LLMs based on their own preferences.
Other than that, a no-LLM user today may become a LLM user tomorrow but doesn’t want to mention it because he’s not forced to and he doesn’t plan to use it always. That would be dickish behavior knowing that some contributors want to have nothing to do with LLMs but it’s not technically impossible so the list would be ineffective anyway.
If one doesn’t wish to review LLM-generated code, they cannot block emails altogether, otherwise they will lose visibility on the changes to their component and become a less effective contributor. The other option (which is arguably worse) is to be hit with surprise code after a git pull and have to search for context afterwards, reviving e-mail threads they didn’t want to be involved in. They will burn some social capital on doing that because they voluntarily decided to be out of the original discussion.
Then a LLM-generated PR with a nasty code execution vulnerability hits their inbox and they spot it. What are they supposed to do? “Oh, just do the sensible thing and report it”.
If they do, they will be sacrificing their own beliefs for the greater good. That’s beautiful but LLM users did not give them the same grace today. That might generate some resentment on no-LLMs devs.
They’ll be told to grow over it.
When they’re finally burnt out they will have to make a decision on whether to leave, fork, or be guinea pigs for AI companies. We’re back to square one.
But at this point they will most likely give up and leave.
We're so cooked, chat. Since looking at Open Slopware a few months ago, Lisp against the (LL)Machine yesterday, and now this, I'm pretty convinced it's impossible to avoid LLM-generated code in daily computing, unless you stick to a self-consciously retrocomputing environment and freeze it in time. I think the Tower of Babel is going to collapse rather messily, but it's not actually possible to participate in society without going along for the ride. All we can do is not contribute to it, and preserve old systems for after the tower falls. It's a heavy lift for us to fork Emacs. Forking Linux is unthinkable.
Maybe it's time to move fully to NetBSD?
https://www.netbsd.org/developers/commit-guidelines.html
Code generated by a large language model or similar technology, such as GitHub/Microsoft's Copilot, OpenAI's ChatGPT, or Facebook/Meta's Code Llama, is presumed to be tainted code, and must not be committed without prior written approval by core.
I dunno; we were already under mountains of slopware written humans. At least now a well intentioned and solid engineering professional can swim upstream with the help of even locally run models.
The solution is to make sure those LLM tools help maintainers instead of just causing them pain.
This is exactly how I’ve approached LLMs at work: how can I use these tools to make life /easier/ for my coworkers? This obviously precludes slop-bombs. It admits motivated refactoring, better test coverage, and bespoke tooling that we wouldn’t otherwise have the time to create. There are ways to do this where we end up with better code at the end. Focusing on the human on the other end helps a lot.
100% agree... but a great many people don't seem to approach it that way, and I worry that's a weakness in Linus's position here. When he says "We're not forcing anybody to use it, but I will very loudly ignore people who try to argue against other people from using it," the risk is that people who argue specifically against the anti-social use of it get loudly ignored by reflex.
In the kernel community we do open source because it results in better technology, not because of religious reasons.
Ironically, those "tool" are one of the most opposite of open source. Even calling an Anthropic subscription a "tool" is debatable.
I don’t think that’s ironic. Linus developed git because he could no longer effectively use BitKeeper, not because he was annoyed that it was closed source.
Thank you for mentioning BitKeeper. I just realized that Linus has "been here before" with respect to the degree of adoption of proprietary technologies.
[...] if somebody has issues with that, they can do the open-source thing and fork it.
Because that's such a trivial thing to do.
I think you're trying to be sarcastic here but I'm not sure why. It's pretty easy, and it routinely done on the regular by many different groups.
Forking code is very easy. Forking a community (i.e. creating a maintained fork where development continues) tends to be a lot harder, and happens much more rarely.
Especially when your reason for forking is not a technical reason but a personal ethical reason that the majority of contributors don’t agree with, by definition: if the majority did agree, there’d be no reason for a fork because the policy could be swayed without forking.
Forking the Linux kernel with the goal of having your own independent development and not keeping the contributions made by LLMs cannot be considered easy.
Forking it to add a small feature, sure. But hard forking? Completely different beast.
People know that and leverage the fact that it is technically possible but unsustainable. False naivety. You see that everywhere, not only in tech.
Which is why Linux forks for SOC support are always kept up to date, with security and performance patches merged /s
Not that it's the way I want it to happen, but I wonder if this dividing line will drive more developers to contribute to Redox. I get good vibes from that project and want to see it succeed. I'd hate for it to grow only by Linux shrinking, but I'll take it if it comes to that.
Time to bring Gentoo/FreeBSD back!
Anyway, what does it mean for Gentoo Linux?
And NetBSD which doesn't accept LLM AI tainted code
I wish! I have fond memories of the (admittedly brief, circa 1 year) time I ran NetBSD when I was younger. I tried to install it recently, but I realized that I no longer have a "desktop" computer, I only use laptops these days. Since the Wi-Fi support is a bit behind compared to other systems (FreeBSD included), I'm effectively cut out from using NetBSD. The only realistic options I have are running it in a VM – and at that point, why bother? – or using a USB dongle to regain Wi-Fi capabilities that I already have.
In case there's a third option I'm not seeing, I accept suggestions 🙂
There's this WiFi renewal project for better support https://blog.netbsd.org/tnf/entry/wifi_renewal_restarted
I saw that and I'm eagerly waiting on it. I will toss a coin to my Witcher the NetBSD project. If I had the chops, I'd help with development.
FreeBSD devs are definitely using LLMs. Some of them anyway
I mean NetBSD also doesn’t forbid the use of LLMs for development, it only forbids LLM code. Using it to review and debug your code is fine.
And yes, I feel very strongly about this, not because I feel strongly about AI per se, but because we have a long history interacting with the FSF.
They have their "ethical" reasons too, and use them as a weapon, and as a way to drive away sane people.
It's why Linux is not GNU/Linux, and why we call things "open source" instead of "Free Software".
That's an amusing example to bring up, considering that the other shoe hasn't dropped yet, and we might still live in the timeline where the FSF rules that "GPL remains applicable & enforceable on slopware", so GNU proclaims "go forth and vibeth ye faithful".
(Sorry, nothing more substantive to say, that just genuinely made me chortle, even tho Linus's point is more "we're not like the FSF generally" than "we don't agree with the FSF on this specific topic".
I remain morbidly mesmerized by how readily FLOSS maintainers of all people can adopt a blackbox they can only tune, inspect or modify via Markdown thoughts & prayers, a fortiori the guy who cared so much he made his own kernel, VCS, and editor, but I suppose we can all only care so much. I imagine Linus has little cause to care personally; the irony is just more palpable for the peanut gallery who needs to clear the Anubis gates to read his position on LLMs & check out his Emacs fork)
Linus did not create microemacs. He created Git because the license to use BitKeeper for the kernel was revoked, not because he didn’t like using it.
In his original talk about Git he explicitly says he makes open source software because that’s the only right way to make software but for using software he wants the best tool for the job.
Linus did not create microemacs.
Thanks for setting the record straight! Hyperfocused on the README.md. Incidentally the LICENSE included in the commit adding that README does a good job retracing that project's lineage.
In his original talk about Git he explicitly says he makes open source software because that’s the only right way to make software but for using software he wants the best tool for the job.
Heh; I can sort of see the logic but that would imply that "the best tool for the job" (LLMs) does not have to be "made the right way" (open source; going to exclude "open source AI" considering the hassle involved in modifying them), so I don't know how convincing I find this.
BitKeeper. Not Open Source, but, in Linus's opinion, the best tool for the job, even though it was not "made the right way"
Just because something was made wrong way does not necessarily mean it is not the best tool for the job.
It seems to me you're trying to conflate process with results. You can have a terrible process with good resulting software, and vice-versa. Bitkeeper's development process may have been suboptimal, but that doesn't mean that it wasn't the best tool for managing the Linux kernel.
Linus then went on to prove that the bitkeeper way of doing VCS was the best tool, by making git "the right way" and it going on to eat the world.
It seems to me you're trying to conflate process with results.
Or maybe I'm trying to willfully ignore results! I confess I don't know the first thing about making software that eats the world. All I know is I like fixing my tools, and I've never found a yak I would not shave; whereas by all accounts the process of working with LLMs sounds ignobly dull, to me; results notwithstanding.
Linus then went on to prove that the bitkeeper way of doing VCS was the best tool, by making git "the right way" and it going on to eat the world.
How similar is Git to BitKeeper, actually?
FSF rules that "GPL remains applicable & enforceable <...>
FSF does not have a say in the matter.
FSF is not a court and therefore it cannot rule on the matters of law (such as copyright law).
They can indeed not decide whether or not the GPL can still be enforced. But they can decide the policy for whether or not to accept slop contributions, and have indicated that that policy is going to depend on their understanding of how it interacts with GPL enforcement.
FSF is not a court and therefore it cannot rule on the matters of law (such as copyright law).
Not in that sense, no, and I apologize for the shorthand. s/the FSF rules (".*",)/the lawyers that the FSF is currently consulting conclude \1/?
I remain morbidly mesmerized by how readily FLOSS maintainers of all people can adopt a blackbox [...]
I'm not very familiar with this lore thread, but doesn't a lot of corporate-backed developers work on the Linux kernel these days? Even the Linux Foundation is an umbrella of companies, not citizens gifting away their time for FLOSS.
I'm not very familiar with this lore thread, but doesn't a lot of corporate-backed developers work on the Linux kernel these days? Even the Linux Foundation is an umbrella of companies, not citizens gifting away their time for FLOSS.
It is a fact that many kernel developers are corporate-backed, yup, but I don't know how much that should correlate with whether those developers value control over their tools. Tho I suppose $DAYJOB can desensitize you to helplessness by imposing tools (thereby making willing adoption of LLMs more likely; I guess there's also the "$DAYJOB forces LLMs on you" hypothesis but that's not the prompt for this thread, AFAIU).
I honestly don't know what to think. I thought Linus was being a bit addle-headed, but reading the other person's comments, I think Linus is the sane one.
Can you expand on that, because I came out of the thread thinking the opposite?
All Laurent is asking for is for the community to consider that people may have personal ethical concerns over using generative AI (emphasis mine):
I consider that, today, there's no ethical justification for the use of generative AI in FOSS development. That's a personal opinion[...]
What I'm asking is for the community to recognize there are people with ethical concerns, listen, and try to see if a reasonable middleground exists where some of those concerns could be addressed.
Linus misreads this as a demand for the entire community to adopt those ethics:
You can choose not to use AI, but that's your PERSONAL choice.
It has absolutely no impact on anybody else, and you should not expect it to have any.
So keep your ethics where they belong - in your personal life. Don't try to enforce your ethics on others.
Then there's the vegetarian parallel where Linus likens Laurent to someone who's trying to force everyone to be vegetarian, to which Laurent responds:
What I'm asking is to not be forced to eat meat
To be honest I think the reasonable middle ground is already to ignore those personal opinions. I'd also be pretty angry to push back on this. Imagine the same argument made with christian or muslim or judaic opinions: a "middle ground" isn't halfway there, the middle ground is already "you can do whatever you like on your own system."
What I'm asking is to not be forced to eat meat
No what he's asking is to not forced to be spoken at by meat-eaters. And he isn't! He can step out of the room. Or just ignore them! The tooling to solve this locally is present in every email client.
Now, if one really doesn't really want to receive e-mails from a particular sender, a block list solves it
I think Laurent is being forced to use AI (in a way) due to AI reviews possibly being forced onto maintainers. I am not certain but I think that is where this discussion started.
Well in the sense that they're delivered to his inbox via the ML. In this case it's awkward because I think a maintainer forwarded AI reviews to the ML? So Laurent isn't "forced to use AI" so much as "forced to confront the reality that others use AI". I didn't read through the whole chain though.
My understanding is that the kernel is going to start using an automated tool where all patches go through ML review. It's done by a bot, it isn't simply other (human) reviewers open up claude, type a prompt and copy paste it back to Laurent
Yes, pretty much what FeepingCreature said in his comment.
As far as I know, there is no mandate to use Ai for the kernel, right? If so, what's the problem? Everybody has their own opinion about it, but so long as one is not being forced to use (or not use it), I don't think there is real contention here.
Also, I was a bit disappointed about the reasonings provided by Laurent. When discussing technical issues, why not focus on technical arguments?
It's good to hear some voice of reason and not just emotional outbursts how the world isn't the way people would like it to be.
The whole debate of personal choice has been repeated over and over again and we always end up with individualism as the way to go. Because everything else is worse, not because it's great.
Did...did we read the same mail? I'm definitely not going to tell Linux maintainers what they should do, but if that post was expressed with the voice of reason, then I really would not like to see him arguing because of spite(edit: not saying it's written out of spite).
Ghostty, Godot, Zig, curl, they all had much more reasonable posts about this """debate of personal choice""", but those weren't in the context of a mailing list.
If we are grading on a curve, yeah that was really reasoned for the author when dealing with a contentious subject. Or semi-contentious. Or really anything.
I wonder which obscure OS projects will take loud, principled stands against usage of AI tools in response.