Linus Torvalds tells AI haters to fork off Chief penguinista Linus Torvalds has declared that Linux is not an "anti-AI" project, telling contributors who object they can either walk away or fork the kernel. On lore.kernel.org, the archive for Linux kernel mailing lists, reformed potty mouth Linus was responding to a discussion about some negative sentiments toward AI. …
as one of the most well known FOSS voices i wish he would also advocate for open weights models and running locally on your own hardware etc. in the same spirit as developing software/linux. AI is literally trained on the commons of the internet and he is well placed to begin the conversation of righting the balance. as it stands this statement is dual use as an ad for the AI giants. and that’s a damn shame.
Using a software-as-a-service frontier model to help with Linux and other open source projects doesn't spill trade secrets, because the source is already public. Proprietary software companies producing products like FidoBasic and DogOffice need to be more careful. Note also, having a clever chatbot help develop code is different than allowing agentic access make commits directly to the kernel tree.
There are definitely software engineers who are motivated by being part of a community. Focusing on the technology is likely the best way to keep that community attractive. Just like philosophical discussions about the ideal tab size had to be ended to prevent endless discussion from fracturing society, so too philosophical discussions about AI.
Give it time. First cracks are showing, and the recent developments (I do aarch64 bringup and tinkering for 3 years now) point to making existing hardware useful, regardless of the secrecy mindset of its creators/vendors. This will only accelerate. And, enabling local models for mortals / non-corp types.AI giants have their own arena and, quite frankly, issues enough to consume them, and a sizeable chunk of the fraud industry with them.
Here is what I doubt:
Alright, maybe there are different types of AI.
Those that are useable as a dedicated tool, I have used myself. For translations,
and for analyzing big piles of data fast, they are great.
Then, according to my lunatic logic, there are those AI models, that are invented and
programmed to ignore their masters, by people who, I dunno, sort of are walking in lockstep, and are singing:
laaaaaa!
"Software, AAA IIII üüüüber alles, la laaa
la laaa
la
laa
I can't read musical notes, so there..
Those kinds of folks are the scary aspect of A I. They come across like never grown up irresponsible dudes, consequences be damned,
or like certain politicians I won't mention, because they might send out a death squad. Life and death is nothing but a big joke to people like that.
It's just a feeling I get, when I see who these companies support politically.
I am probably not the first one to notice the the extreme lack of humanity, lack of conscience,
lack of foresight, reason, kindness, decency among certain software giants.
Are Register readers and commenters,feeling annoyed when someone like me
makes such generalizing, angry, dismissive judgments about the software industry?
Or is it just that I know to much about these companies.
As one does, after having encountered the lack of customer service at a variety
software companies.
But I am also fine with being the only saying this.
It's lonely at top, middle and bottom, knees and toes.
Mr Torwalds is maybe 100 times smarter than me, yet but when it comes to explaining
the use of Linux for the masses, he exhibits either zero interest or zero talent for
basic human-to-human communication. I sure would love to get away from MSFT,
but there are certain programs that exist only MSFT compatible, and when I ask about
running MSFT Apps under Linux, I get either no answers, or some confusing and contradictory answers.
So there it is in a nutshell! Now how do I get out of this nutshell!??
I can kind of see the argument from his perspective.
Maybe I’m wrong, but when you’re working close to the metal then I’d suggest that your requirements spec is more likely to be very precise. Hardware doesn’t forget things, or change its mind and isn’t vague and handwavy.
Under those conditions, then maybe I can get on board with the idea that you can describe the problem in sufficient detail for an LLM to be a kind of pair programming partner.
I think the problems are much greater the closer you get to the very unpredictable and imprecise wetware.
I’m still not entirely convinced though.
I think it might also be related to the change control procedures already established for the kernel. The kernel has already had to find ways to deal with contributions from unknown people that might work or break, whereas many other projects have been able to assume that people contributing code had a reason and cared. That makes them more insulated against problems unchecked LLM usage brings.
My opinion on LLM usage is and has always been that, if you use it* and generate good results, fine with me. I get annoyed if someone uses it, generates crap, sends it to me without testing or review for me to clean up, and pretends they've done something useful. While that has mostly aligned me with those who oppose AI, my problem's only with the slop.
* LLM usage is only fine if it doesn't break other rules. People who decided to upload internal code and data to an LLM of their choice was not acceptable in the same way that using any other unverified third-party services on internal data would have been a problem. That it was an LLM did not make that better.
> I've had Deepseek write a device driver skeleton that was perfect 1st time
Which makes sense when AI is trained on examples and essentially replicates the pattern it has already seen combined with some search and replace. So it makes sense it would make skeleton code very well when there are already plenty of examples of that boilerplate in the wild (as well as plenty of actual drivers it can derive the boilerplate from).
When it would fail would be when you start pushing it too far from the boilerplate and into actual code logic. At first it'll appear to work pretty well (making it good for demos) and then it'll start making weird decisions (eg doing the same thing in different ways rather than reusing common code), hallucinating or just failing to understand the logic properly because you have gone outside of the examples it has been trained on so it has no pattern to follow. And then agentic AI would start iterating on that making it get better in some areas while worse in others and often making the code less and less maintainable overall. Hence it'll build the skeleton well first try, then fail entirely at getting the driver over the finishing line without having an actual developer step in.
> Which makes sense when AI is trained on examples and essentially replicates the pattern it has already seen combined with some search and replace.
Which is what I, and probably everyone, did manually before AI - Search Stack Overflow and the web, copy and paste, keep looking to find all the edge-cases and alternatives. At the end of the day you have a reasonably good skeleton to build on.
An AI should be able to do that quicker but what a user won't have picked up is all the related information and technical detail revealed during that searching. By the time I delivered my driver I could understand what it did and why, could maintain it and fix bugs without needing an AI. That ancillary information picked up along the way was crucial to completing the project.
And that right is the crux of the issue with "AI". It removes agency from the user (or developer), and places it squarely in the proverbial hands of the computer. I prefer to control what I do on my computer. Control is the key. It's the same reason I only drive manual transmission vehicles.
It all comes down to who (or what) is in control.
> Hence it'll build the skeleton well first try, then fail entirely at getting the driver over the finishing line without having an actual developer step in.
Don't bet your house on that - ChatGPT and DeepSeek seem to be munching through Erdős rather faster than the meat based mathematicians.
Truth is AI is basically highlighting how much of development, copyrighting, graphic design, music etc has little real value, novelty or skill involved.
If nothing else DeepSeek's latest iteration should be as terrifying to investors as it is interesting to developers - it seems both groups have heads deep in the sand for almost opposite reasons.
>” I've had Deepseek write a device driver skeleton that was perfect 1st time”
It wasn’t that long ago achieving this was a simple cut-and-paste from a reference. The hard part was adapting the skeleton code to your specific requirements and then adding the “flesh” to convert the generic skeleton to a device driver for your specific device.
Not at all.
Adopting Rust isn't just about approving another language and compiler. The whole philosophy behind Rust involves a different way of thinking about the code that you're writing. Assuming you're doing it "right" of course, rather than bending Rust code to make it do things in a C like way.
For anything other than a very small project, having two radically different approaches within the same code base is just inviting a long term maintainability issue. Any maintainer needs to have a thorough understanding of both to be able to make changes or introduce new features.
LLMs might spit out terrible code, but at least (in theory) it's based on code written by other people with the same basic mindset.
@intrigid
The average person does not exist (that got sorted out in the days of Quetelet and all).
Most people quite like being able to earn some money, buy food and clothes, and put a roof over their head(*). The latter desire is already challenging in many 'rich' countries for absurd reasons. Mess with those common factors of most people's lives too much and things could get lively.
Back vaguely on topic: if the kernel team thinks that they can manage the results of larger scale LLM assistance, then fine, we can see how it goes as everything is in public &c. As another poster put it, pragmatism might be best.
(*) and have the occasional party
How many of these "average" people are benefitting from the ludicrous increase in hardware costs? (RAM especially)
Or benefitting from the increase in energy costs through so much wastage?
Or benefitting from the extreme weather that energy use is contributing to?
I definitely see more negative than positive at this point in time.
Enlighten us, give us examples of tangible benefits of LLMs for the average person. Preferably ones that outweigh the drawbacks, which are quite tangible - mass layoffs, increased electricity costs, cyber bullying with deep fakes, enshitification of search results, mountains of AI slop across the spectrum.
RE: The average person does not exist...
By definition, no one who knows how to program Linux, is an average person.
Even non-programmer people, like me, who only visit The Register to read people's
sarcastic, or sweet memory stories about how stupid people, such as employees,
bosses, and clients behaved ... are not average.
So, right here, on the Register, the average person does not exist.
But out there, in the non-techie world, there are plenty of average people,
and they are correct to be afraid of software and A I, because they
have no clue about it and no control over it.
Put another way, in a democracy, people have or at least believe they have a measure of
control over the laws they must live under. In a dictatorship, in a tyranny, say in Russia,
North Korea, China, and lately also in the USA, no such measure of control exists
anymore. And for the same reason - i.e. complete loss of control over their lives -
they also dislike the robot-like behavior of AI, programmed according to the harsh
concept of reality which software makers love to impose upon the world.
Name one person who likes the robots that read and dismiss 490 of 500 resumes
and job applications. There is no one who likes that! If you do like it, I suspect
you are a robot, or a robot merchant who loves to sell robots.
The AI software creators ARE NOT REQUIRED to reshape the world according
to their image and understanding of right and wrong. But they love to impose
their will upon the world, and they ask no one, nobody at all, for permission to
do so.
I am just trying to put into relatively reasonable words, what feels like an over-emotional or frantic
response to AI. Yes, relatively! For robots, what I say has no meaning whatsoever.
Another open source project rejects any AI-generated code, slop or not, on the grounds it is uncopyrightable.
Tin men don't have tinman-rights and cannot claim copyright on their output, you see.
Nor can they grant nonexistent rights to others.
Or file a lawsuit to defend those nonexistent rights.
When FSF, or Microsoft, or SCO catches up, will Linux be fragmented into GPL and nonGPL lines of code, and thereby tainted forever?
> Compare and contrast with OpenZFS
https://tinyurl.com/openzfs-avapet
Five commit signals.
>Tin men don't have tinman-rights and cannot claim copyright on their output, you see.
Is that also true for compilers ?
LLM: prompt a human types -> LLM generated source code
Compiler: high level language human types -> intermediate code -> randomised by Monte-Carlo reordering optimiser into fast but incomprehensible spaghetti -> binary
OTOH there *were* companies selling compilers that tried to exert their licensing rights over the compiled results of your own source code (and no, not just "you've linked it with our proprietary C library"), because using their transforms somehow transformed the ownership as well.
Who knows what insanity someone is going to try and claim in a court? SCO lawyers, block your ears!
Not had cause to investigate compilers in recent years, but in the past it was common for compiler makers (eg. Microsoft) to differentiate between “private” usage of the compiler’s output and “commercial” (ie. Resale) of the compiler’s output output, demanding a higher fee if you were using the compiler’s output output to produce a software product for sale.
>It's not isomorphic, compilers are formal, deterministic transformations of the code
I'm not sure that's true for all optomising compilers that do statistical analysis of code paths
IIRC it's one of the issues of getting absolute reproducible builds
And similarly is there any reason an LLM can't ensure it produces the same output for the same prompt ?
I'm sure that the mega billionaires who have the ear of Donald 'I love to sleep on the job' Trump will soon be filing suit to make any AI generated 'stuff' (slop or not) have the same rights as a human. After all they did it with corporations and Trumps buddies on SCOTUS just love to fly around the world on private jets.
Once those rights are granted then everything that AI generates is copyrightable and able to be monetized at zero risk of the IRS coming after them. They can't unless the AI has an SSID or Tax Registration Number.
Those mega billionaires will soon become mega trillionaires.
FSCK AI and all that it generates. IMHO, it is long past time for AI's to be subject to modern Salem Witch Trials or for a new Ned Ludd to start burning them down.
"IMHO, it is long past time for AI's to be subject to modern Salem Witch Trials"
Um, you do know that those were innocent people, hanged on the basis of false accusations, right?
(Hanged, not burned. One man, Giles Corey, was tortured to death. I'm not disputing that "witch" burnings took place; just not in Massachusetts.)
This is a good point.
Only the copyright holder can issue a license (e.g. GPL).
When the code comes from a LLM, who is the author (per Bern Convention) ?
It's not the meatbag entering a prompt (precedent long set about the significant own creative input required) . The output is at best a derivative work.
The output of a LLM can only be GPL'd is you're sure the training data used to generate the output was GPL code (which you can't be). One bit of say BSD-licensed code used to generate the output and you can't license it under the GPL.
I beg to differ.
Language Models are dependent on their training data, which has been created by *other people*, and in many cases often without their permission, under licenses which limit what may be done with their work.
If I translate a book* while I may be the automatic copyright holder of the translation (as I've put a significant amount of effort in myself), I can not publish said translation nor give someone else permission to publish it (a license) without the permission of the copyright holder of the original work. If I've used a machine to do the translation with no input myself, then I'm not even the author of the translation.
Entering a prompt to a LLM is not "sufficient human control over the expressive elements" to be the copyright holder of the output.
So while LLMs may be considered a tool (though more so in the British vernacular sense), they are not the same as other tools in this aspect.
You can't for example copypasta BSD Licensed code from various sources, then release the result under the GPL. Using an intermediary such as a language model to do the same doesn't free you from the legal or (IMO more importantly) ethical considerations.
-----
* one of those things made from dead trees
I think that is not correct.
Machine generated code is not protected under copyright law.
Did your heard about the case where an ape did a photo of himself, the photographer who owned the camera claimed copyright and lost?
Currently only human works can be protected.
It is not only an AI problem. Complex IDE code which is autogenerated do have the same problem if you can prove that the code in question has been machine generated.
Depending on your reading of copyright it could be argued that the copyright for the AI-code would lie with the writers of the original code that it's slurped into its training data. It could also be argued that the AI-code is a derivative work - much like a collage, albeit one that was like produced not in accordance with the original copyright terms.
It seems that there are (at least) two separate issues here. There's the copyright of the input to the AI model and the copyright of the output of the AI model. To me, it seems like the input to the model is from massive copyright infringement - and that means that the output is tainted.
Just yesterday I downloaded the source trees of OpenBSD and NetBSD. With both of them I worked for some time in the past (not as kernel developer, though). I used to compile both of them for several CPU architectures, and developed applicative software on and for them. I might give them some tries again. And also to FreeBSD. For quite a long time i had several suspicions regarding Linux kernel, but also the whole ecosystem (systemd most of all, that's why Devuan is my primary distro from its first release). I, sometimes, opened randomly chosen source files of some important, on daily base-used software for mission critical purposes, and was sometimes getting sense of discomfort. The code that I always found the best (most readable, most maintainable, and most secure) was the code found in OpenBSD source tree. And it seems we will, sooner or later, find ourselves in situation to need to learn, and understand well, kernels and system libraries of the operating systems we use, at least in order to remove some unwanted code before compilation. Waiting also for alternative, non-monopolistic nor duopolistic, CPUs without "surprises" hidden inside. At worst, I'll switch to eZ80 (since the venerable Z80 is discontinued two years ago, and eZ80 is a pretty good replacement). And there's pretty good quantity of software you can run on those 8-bit CPUs, the really important pieces of software, I don't need "special effects" of modern word processors, spredsheets, compilers, file managers, etc). Also FreeDOS, Haiku, Kolibri OS, ecc, might be pretty decent solutions for legacy hardware.
WE THE PEOPLE DON'T WANT/NEED FAKE INTELLIGENCE!
Citizens and local communities have successfully blocked or delayed at least $130 billion in proposed data center projects in the first three months of 2026 alone, matching the disruption total for all of 2025.
KEEP IT GOING! WE NEED A REVOLUTION!
Linux is not one of those anti-AI projects…[…]
The solution is to make sure those LLM tools _help_ maintainers instead of just causing them pain. There's no question on that side.
Okay. So what does that mean?
Let's dispense with the arguments against whether LLMs are “useful” or not. Asbestos is useful. Human experimentation is useful. Child labor is useful. Chemical/Biological Weapons are useful. Saying that LLMs is useful is an extremely low bar and honestly means nothing, because its use exacts a toll, and the reason why people don't use those tools is because they decided that the toll was higher than the benefits.
Aside from the fact that LLMs swamps maintainers and introduces problems that are hard to spot, the question of the IP status of LLM code vis-a-vis ownership remains a legal thicket. Linux kernel devs have already gone through this shit multiple times.
Why would you want to go through it again?
Although you put it in a way that many may find offensive (please note, not me) and therefore disregard, I think you correctly put the finger on a very valid point: the base of the so-called AI consists out of "robbed content". In a world where Joe Average can't repair his own tractor because that "would infringe", in a world where an university has to pay a publisher to download a publication it produced itself, it apparently is OK to steal what is not yours, repackage it with an authoritative tone and sell it again. So, knowing this, using AI, no matter how useful it is, is fencing. Which seems to be illegal. And think about that for a moment when discussing the use of these tools. But hey, the human species always loves that bargain that fell of the truck, right? Up to the point that it touches their wallet. Or when the light doesn't go on any more and no water comes out of the tap any more.
that many may find offensive
Yeah, it was deliberate. Most notably, one of the examples I put in there was literally very provocative — the human experimentation one linked to a man who is literally the father of modern gynecology, who literally experimented on enslaved women and girls, often without anesthesia. So in that case, the fruit of that atrocity has literally saved millions of lives, and yet most gynecologists must grapple with that dilemma at the root of their profession, and, you know… these days we don't experiment on people without ethical consent, because we know where it ends.
In a world where Joe Average can't repair his own tractor because that "would infringe", in a world where an university has to pay a publisher to download a publication it produced itself, it apparently is OK to steal what is not yours, repackage it with an authoritative tone and sell it again.
Which is why I am somewhat leery of the whole “intellectual property“ argument against Generative AI, because you've hit the nail on the head on what the problem is — it's a matter of labour and exploitation, of having one's efforts be stymied by the needs of a corporation to make-a-de-monie.
It's a shitty deal. Seeing someone make big bucks of work that you originated while open source developers struggle to make the money is already an issue in the FLOSS space — LLMs ramp it up to 11. Why should anyone do any work if all that happens is that some douchebag slurps up the code into a token-predicting synthetic text extruder, and then gets paid billions for… saying, well actually, this is the fruits of my genius. On top of that, we'll need to build a data centre next to your house and drive up your power bills and pollute your water.
But, incidentally… LLMs are useful, you cannot deny. Who gives a flying fuck?
>” Okay. So what does that mean?”
From the various articles, I suggest currently Linus is saying AI has a role in the analysis of code for bugs.
I sus-ext there has been no change to his stance on code contributions to the kernel: these have to be made by named individual.So as far is Linus is concerned, you are welcome to use an AI to produce code for submission, but it is your name a reputation that is on the line.
If the tooling is just to use it for documentation and general boiler plate autocompletes. Then I will buy it. But for actually code generation that one could trust I would never trust it. I would not even trust the code analysis report seriously. Recently I got called on a report that found SQL injection vulnerability on SQL client web based tool. The finding was that SQL queries were not parametrized. That parametrization was being asked for generic SQL language query. Neither the LLM not the people who managed the reports would acknowledge it is false finding. So I can imagine a thousands similiar crazy finding and statements form AI. That for sure is frustrating for a lot of developers who do perfectionist work in what they do and need to deal with endless nonsense reports. Now imagine that same flaw also is used for actually generating code like a kernel that can end up in critical environment like medical care. I think people that really trust AI they should go for it but do not ask people to intervine or approve whatever is spewing the AI. If they think AI has better version of code then the manager who trust AI should just sign off and promote it. There is no point putting humans under the stress of reviewing and owning the risks of AI generated code.
sadly in the comments everyone has to have black and white at-extreme stance - either completely anti-AI or ad-absurdum generating slop with ClaudeCode without seeing any middle ground that Linus is talking about (ie. glorifies "auto complete on steroids")...
one (well, two actually) big problem with AI is that either the output is not read / reviewed properly - either because of time constraints or general slack or because it's used by incompetent people that can't sensibly verify the output...
Reading and reviewing code is hard. Really hard. Personally I think anyone who says it isn't is kidding themselves. It can also be crushingly dull at times, if not most of the time.
Code review is hard when that code is written by a human, simply because given a problem (especially a code related one), two people can probably think of at least three ways of solving it. I've seen code review descend into "well I wouldn't do it that way" arguments rather than trying to assess whether the code in front of you is a valid and clean approach.
I tip my hat to those that will scoff, but this is why I've always been keen on pair programming rather than code review as a way of getting two pairs of eyes, and more importantly two brains, solving a problem. Both people are working within the same context, are wading through the thicket of (often vague) requirements and share an understanding of the solution, plus its strengths and weaknesses.
Can you pair with an LLM? Given the stochastic parrot nature of the tool, my opinion is that it's highly unlikely. I also worry about the real hallucination, which is as Cory Doctorow recently said not the output of the LLM, but being lured into the belief that there's actual, human-like intelligence at work.
This process is hardly unique to code. People have been using the syntax of written language for millenia. Any critique of language funds the potential for myopic perspectives from people who forget they are subjective entities. It's hardly the review process and usually the individual that is the issue
I've been sampling AI (local, open and subscription models) and it is interesting how it can now, with relative success, one shot basic things like:
"refactor this codebase to push two protocols over the same port forwarded udp socket. Extract the socket management into a base class which emits an event for each received socket, and write two wrappers. Each wrapper filters for its protocol and forwards the relevant message to it's current message handling".
That's not the prompt I used verbatim, but it was around that length. I used actual class/interface names in some places.
5 minutes later the refactor is exactly as intended, all tests are updated, new ones are added to verify the new logic, and I then prompt one last time for an extra test to cover a testing gap. No duplication. Existing protocol detection/filtering code was moved around appropriately.
Probably would've taken be about 15-20 mins to complete that. I got it done in 5 plus a min or two of review. In the meantime i was figuring out my next step so I had it working on that after the cursory review and before the thorough review
They're very useful tools nowadays
Old tool here. Been using Claude for reworking and extending an existing application. It did a good job of analysing the existing codebase and that helped me to get my head around the application is a shorter time than had I walked the code and reviewed the available documentation. Some of the code it produced was again fairly good, with little or no rework/re-direction required. At other times, the effort has been poor. That's the problem - consistency.
We have a 3-line whip from the senior management to actively use AI - and I have found cases where it has definitely saved me time - for example compilation issues, unless I spot an obvious simple issue, I found getting AI to look into it saves time. I had a particularly nasty issue which it homed in very quickly compared to the time I had already spent.
But you need an experienced person as the final arbiter - this is where in junior/less experienced hands, it has potential to do harm, especially if there are peer reviews being done by equally less experienced hands or those who treat peer-review it as a rubber-stamping exercise, and I have seen some dubious code getting through to production past peer reviews, without any AI coding. As soon we defer to "the computer knows best", then, the game is up/why bother at all
This was highlighted today here too... while 'AI' can do 90% of 'the thing' (whatever 'the thing' is), the human can then spend the time saved on those 90% on the 10% that 'the AI' didn't do, or couldn't do, which is the valuable work that only the human is able to.
Getting an LLM to summarise a collection of 40 inter-related RFCs and extract specific things from them is a lot more productive than you doing everything manually. Sure, you could write a script to do it too, and that's fine, but with the tokenised reasoning that exists in an LLM, you could do this on a local machine, point it at the RFC files and say "go do the thing with every single one of these files" and save yourself time and effort.
I'm definitely *not* a fan of the "launch a room full of agents and get them to write the full works of Shakespeare" (see what I did there) method of vibe coding stuff, but, as Torvalds points out, if you have specific things that can be automated, then automate them (that's what build systems do too, you know... they're just... well... the drawn-stick-man to the fleshed-out comic of AI).
Like someone else pointed out, pragmatism is a good approach to have to this. I would definitely prefer a packaged LLM that I can run on my own hardware (even if it's slower), but it's *my* electricity that's being used, and it's *my* domain knowledge that it continues to improve itself with.
Linus seems to have inadvertently made an argument against himself
"........... Because it’s not like natural intelligence is always all that great either.”
Guess which "natural intelligence" created AI?
So here's the current sad state of affairs as I see it anyway: flawed "natural intelligences" ie. human beings creating flawed "intelligences" ie. AI, used by flawed intelligences ie. hoomans, to create flawed software that is then consumed by flawed intelligences ie hoomans again.. What a mess!
AI is not the everything to everyman solution it is being sold as
but it has its uses in limited cases
it can be good at finding bugs / vulnerabilites in code, (so long as a human verifies the finding)
it can be helpful writing code, (so long as the human using the tool understands and verifies the code generated)
But it is only as good as the data the model is trained on, so the old adage of shit in = shit out still stands
when more and more companies are blocking the scrapers from the AI houses the pool of quality training data gets smaller and the percentage of AI generated slop in the training data increases and consequence is more shit in, so therefore more of the output is going to be shit and its only going to get worse over time
the best models I have seen are the ones trained on a small set of private data for a specific task, the big public LLMs tend to be the worst performers as they fall foul of trying to be a jack of all trades and master of none
This post has been deleted by its author
...wait for the rant where he complains that the herbal is being over run with slop code.
Then again, maybe he's protected, because unlike a lot of Open Source projects that are drowning under slop bug reports and terrible updates , Linux Kernel has a huge budget and team behind it.
...my problem with it is that people seem to think it's some kind of magic bullet that means you don't have to understand or be good at something to be able to do it.
This is not true.
While I can get AI to spit out bash scripts and simple python programs at a speed I could never possibly produce them myself, the only reason I know they're any good - and oftentimes the first few revisions really aren't any good - is because I could, in theory, have just sat down and written them myself. The difference is that I can get the AI to do it over a dozen iterations and gentle corrections in a couple of hours, where as I'd have had to spend several days typing it all out, writing tests, doing the tests, refactoring... end result is I'm somewhat more productive than I was before I had it in terms of getting things done on time.
What I'm not is a better programmer.
What I'm also not is enjoying the code reviews, because clearly there are plenty of people who are happy to have the AI write their code for them and submit merge requests without having read it themselves. The number of things I have to reject because - yes, they pass the pipeline - but dear god look at the code man, can you not see how dangerous that would be if you let it lose on unsanitized user input rather than the well understood environment of our CI?
So... Yeah. I guess I agree with Linus. AI has a place. It can be useful. But like most useful tools it has downsides - and that's before we get anywhere near the ethical or financial arguments, both of which I think are certainly due consideration and I don't think I've really hammered out my own complete opinion on either yet.
Rather than chatbots writing the code while humans review it, it's possible for the humans to write the code and then have the chatbots review the code written by the humans. Since writing code is fun and code review is not, this avoids the historical problem of people skipping the code review because it's difficult and boring.
AI powered code review might be where most of the 600 Windows patches and 400 Edge patches came from this month.
I don't hate "AI". It just think that LLM technology in the field of Artificial Intelligence is a red herring and a long-term dead end for many of the things they're still too naive to realize it is very ill suited for. Why do I say that? Intuition. I trust my own intuition about the limitations of statistical event processing and handling general case real world situations.
I do not believe for an instant that it is possible to make LLM technology predictable enough to be safe enough to use for fields like self-driving vehicles, for example, much less relying on it for health care! I expect all the current-generation would-be providers to lose their deployment authorizations from the licensing bureaus around the world soon enough.
