I understand the spirit of what he means, but I still don’t know how an LLM can be used efficiently as a precise tool like this. If I can describe a problem narrowly enough to guide an LLM to give a useful output, I’ve already solved it myself and may as well just type it out.
If there’s anybody I trust to use LLMs in a responsible way, (to whatever degree that’s even possible) it would be Linus Torvalds. He has some of the highest standards for software development in the world, and the same can be said for the core Linux kernel development team in general.
The general problem though, is that the people who are the least qualified to do software development are very often the people who are the most likely to use LLMs to try to slop their way to the top.
I think that the combination of cost increases for inference, inherent limitations in model integrity, and the economic destruction that will ensue when these fraudster CEOs have finally squeezed every last drop of blood from every last stone they can find, will leave “AI” in a similar place as blockchain/crypto/NFTs.
There will be a relatively niche market for software engineers and related tech workers; local LLM boxes, some subscription-based models that companies pay for with their licensing packages, hobbyists, and some minimal integrations into software suites for tools like background removal, lighting enhancement, etc.
There will be no Singularity, no AGI, no all-knowing Machine God.
I wish more would listen.
Many seem to interpret the headline as “Torvalds is an AI Bro now,” when its more like him yelling “If you all don’t treat these things as tools, shit is going to hit the fan.”
After seeing what happened to windows because of this “tool” I think many of us are rightfully skeptical of what will happen to this community.
It happened before, why are we okay with it happening again?
Microsoft doesn’t treat it as a tool; that’s the problem.
…Which isn’t surprising. Not only because Microsoft is Microsoft, but they own like a third of OpenAI, so they have a vested interest in perpetuating the AI hype.
Microsoft has far more people relying on their products to work, so you’d think they’d exercise far more diligence in making things work…
Well… Microsoft is Microsoft.
This is what they do. They simply cannot help themselves.
It honestly makes me distrust the whole move because you would genuinely think that if the world depends on your stuff to work you would make it work, AI or not.
We’re still talking about Microsoft, right?
They have a looooong history of enshittification and footgunning that predates “AI” by decades, no matter how much the world has depended on their stuff. So no, I suppose I wouldn’t think that, as that would break decades of precedence for MS. But I suppose I never trusted them in the first place.
It’s strange to me because well, I was a 20+ years windows veteran who managed to never have issues on any of my windows installs prior to the last couple years when the ai thing went full swing. maybe part of that was me, because I was inattentive and reckless, but it really just became too much to ignore. i have a genuinely hard time with the idea of essentially retconning the way things always used to Just Work for me in my Windows days. maybe it was bad when I was there and I just didn’t notice. but blegh.
Far more people compared to what?
Linux which is still stuck at around 3% market share
Windows basically rules the world on desktop PCs
Desktop PCs are really small part of the whole, far more people rely on working linux systems to work, like orders of magnitude more.
far more people rely on working linux systems to work
Far more people work on Windows machines to do work
I mean the fact that one of my biggest sources of productivity back when I was there has been essentially nuked since I left, does make me feel limited even though everything else works really well so far. Hopefully that changes but I’m having a hard time migrating that last piece of the puzzle over.
It does make me feel like we need more people here, to make that work.
@inari@piefed.zip , may I ask why do you change the title so awfully it does not only not state the same as the article but misleads it?
So it looks like the source originally used that awful title misrepresenting the article it self. In other words the source needs to be blacklisted for spreading bullshit. Dose this community have a blacklist yet?
I know blacklisting sources is a steep slope to censorship, but at a certain point if the source feels they need to click bait people with tabloid headlines they are not worth our time.
It’s moderation, automating a local rule about reliable sources.
I just think it should fail early, so the user doesn’t type a whole post only for it to be blocked.
The URL they posted has the same title as the post here so it’s the source that changed their title after the fact.
Thank you, but no, it does not. The URL may mean literally anything, where the “slug” may not include question marks, any punctuation, or have a partial content even:
torvalds-ai-is-the-new-compiler-not-the-new-programmerWhat if the nonsense/sloppy slug was generated like?:
Torvalds: AI is the new compiler, and not the new programmer? Torvalds. AI. The new compiler is not the new programmer.Therefore, the URL may not express the actual title, and may also be malformed.
Found the vibe coder
Have you considered rechecking your “sensors” lately?
Oh, and fudge you too, moron.
Yeah, what if? No way lol
fuckin gross.
Never seen a non-deterministic compiler though
Or one that pretends compilation was a success when it wasn’t.
Technically, LLMs (and most ML models) are deterministic with the same input and same seed.
I get what you mean though.
So does ChatGPT intentionally change seed at every interaction so it always spits two different outputs given the same input?
In short: yes. You can tune these values when self hosting - it basically changes the b chance which tokens will be used under which circumstances.
Like Scipitie said, sort of.
In the ChatGPT app, there’s tons of shuffling in the background, like context being injected, maybe sampling changed, agenic action, quantization… its “far” from the actual LLM, opaque by design, and as a result certainly not deterministic.
You can sometimes get deterministic output with the OpenAI API, but it’s also dependent on nothing changing on their end. And their end changes a lot.
But self hosting or using a more consistent provider will give you deterministic output.
C/C++ compilers are non deterministic due to support of super macros that change run to run, non-deterministic optimisation strategies or ordering due to parallelism, and linkers often produce different outputs every time they are run where subtle bugs can cause crashes when addresses don’t line up how you expect. And that’s without mentioning projects that use a configuration step.
It’s actually a big problem in producing reproducible builds for security.
Can’t most serious compilers produce reproducible builds these days given the same build environment. I know there has been a drive towards reproducible builds in general for security verification purposes.
Yeah, but it does take a lot of work to coerce them to do it.
GCC is honestly obnoxious, and you have to do a bunch of unintuitive things to get it. The compile stage needs the built-in RNG seeded, parent file paths stripped, you need to ensure that the date/time macros are not used anywhere, and you need to ensure that all command line flags are passed in the exact same order every time.
I just went through this with GCC16 on a new project.
Not by default usually but yes, you need to do a lot of work to set all necessary configurations and sometimes provide your own RNG seed for things which insist on random looking values.
Not only can they but certified toolchains exist in the safety critical space. Medical, aerospace, nuclear projects, and the like are often required to use and procure such toolchains as part of their validation.
I’m not very good with C/C++ so please correct me, isn’t that what’s called a “race condition”? Parallelism can cause non-determinism but not in the same sense LLMs generate non-deterministic output. Compilers are not statistical machines.
You don’t need parallelism to have a race condition, just not handling an event with expected timing can cause one - like when two keys are pressed within one polling cycle and you depend on one being pressed before the other for some logic like up and right arrow for a diagonal but they register as right and up so the diagonal movement doesn’t trigger
Compiler optimisation strategies sometimes use statistical machines and link time optimisation does use random number generators for producing output
You can have single threaded race conditions appear simply from inputs appearing in an unintended ordering. Or on a single core CPU, you can have one task meant to be done first take unusually long time, so the CPU gives time to another thread which finishes first but expected to finish last.
From my understanding, which is very limited, race conditions are more an issue with concurrent programming. Parallel computing uses separate processor cores for each task so there’s less reliance on stack machines. But I guess each core still shares the memory, so maybe it still happens. Like I said, my understanding is limited. Just use rust.
An LLM is not inherently non-deterministic though - if you don’t randomly sample and instead have a fixed rule (which is what the recent fingerprint embedding approach does), if applied in all cases the output is deterministic, as the neural net at its core is deterministic function. A lot of the randomness beyond that is due to optimizations [source].
LLMs are however unreliable at ‘compiling’. Whether or not it will be able to complete the requested task (translate human language into code) correctly it not guaranteed - at least nowhere near the compilers we use.
While they can be deterministic in runtime they can be compared to cryptographic hashes in that they come preloaded with pseudorandomness which will behave unpredictably
There is a parameter in llms called temperature. If you reduce it down to zero it will become deterministic. And probably even worse.
If you do that it will basically produce the most average possible output, given the context. I don’t know if that’s going to be useful in a compiler context…
Yeah but the output would be crap. Just use the same prng seed and you’ll get reproducible output.
Exactly this. If you made an LLM that had deterministic code output then I’m all up for saying “AI is a compiler for human language”. But until then AI is most definitely not a compiler.
Setting the temperature parameter to 0.00 makes an LLM deterministic.
Yeah, that sounds like a disaster waiting to happen…
Did the headline change? The current headline is Torvalds: “100% of Their Code is Written by Compilers” Too
The title of this post makes for a dumb headline, at any rate. Linus, in response to claims in the industry that projects are now “99% written by AI”, drew an analogy to compilers by saying that all machine code for projects are written by compilers, yet no one says their project was 100% written by a compiler. It’s just talking about stupid hype, which is valid.
That being said, it’s still quite unfortunate that Linus has fallen into the trap of LLMs.
I don’t get the impression he’s fallen in the trap at all.
More so, he sees it for what it is. It’s a tool that can help, but needs smart people that can validate its output. It can’t design entire systems nor effectively solve real world business problems on its own. It’s not going anywhere, so may as well adapt to our new reality.
Source did change the title if you look at the URL matches what OP posted
Even Linus is conflating “doing more” with “being more productive”. A car factory could, if they wanted, pump 10x more cars out the door. But qualiry would decrease these would be full of defects, so in the end you weren’t more productive
I don’t care if there’s more PRs. It’s just and illusion of productivity
Even Linus is conflating “doing more” with “being more productive”.
No he isn’t:
Torvalds’ broader argument is that generating code has become easy, while maintaining a system over years or decades remains hard — and AI does not close that gap on its own. Without a working understanding of the underlying architecture, he suggested, AI assistance can help someone make mistakes faster rather than build something durable.
Productivity = Doing more things Efficiency = Doing the right things
I wrote a whole thesis on the subject AMA
Wouldnt productivity have two knobs? Efficiency and energy/effort input?
Also efficiency sometimes goes against “right”? If you need resiliancy you then “right” migh include redundancy. In software, that could also be guardrails like memory safety that necissarily take some extra abstraction to implement and a processing penalty?
Efficiency is a bit more about also not doing the wrong things :)
Har du en länk kanske?
Where can I read your thesis?
Haha its “out there” but that would also mean that I need to out myself… Let me have a think on this one. Its also in Swedish but mb not a problem(?)
Depends on what your metric for “productive” is. If you’re producing more cars that have to all be recalled and fixed, you’re not making more money than you would had you made higher quality cars from the start, and you then also lose face and your brand is viewed as inferior, so sales drop, and the spiral goes brr.
That doesn’t matter as long as it doesn’t happen in this quarter.
too fuckin’ true
I understand what you’re trying to say, and even agree with it, but the car factory example is just wrong. Those have a design speed that is generally limited by cycle times for individual steps. Like a robot has to place say 10 spot welds. You can’t speed that up. To stick with this example: in theory you could add a robot so both place 5 spots, but that rework takes probably a year and costs millions. Plus you don’t have one bottleneck like this, but a lot of them with similar speed (hence the “design speed” of the whole factory).
You can slow the whole factory down to run under capacity (or not run it during night shift), and then you have room to speed it up. But if you’re running at capacity, there is very little you can do.
It’s physically impossible to speed up the factory to 1.5x, let alone 10x.
Most of these robots are purposely slowed down for human safety and other factors. So I do believe you can speed up a factory a lot, perhaps not10x, but still.
No they aren’t slowed down for “human factors” or safety. They are in cages or behind barricades (at least in Europe). If you open them (like a door or cross a laser barrier), the whole line or at least the cell cell just stops and all of the red lights come on. You can believe what you want, but you can’t speed anything up significantly without massive investment or work. Or they would be running faster, cause money.
Source: I work in the industry and have been in running car plants.
You can just skip a few welds, don’t measure lubricant exactly but do it quickly ±20%, or don’t bother smoothly apply sealant. Reduce time to let paint dry. Don’t bother with exact alignment. Sure, the car might rattle, rust, break in a small accident and the engine doesn’t last long but you’re cranking them out fast.
But you still have to compile
When the next thing is invented, then we’ll have to do three things.
So funny thing, a colleague of mine was arguing that there’s no reason to do a compiler because he got Claude to generate ELF binary directly for a hello world…
Some of the AI bros are very extremely bullish on it too the point of saying compilers are a waste of time…
I meaaan i guess. I don’t think the training set is good enough for anything complex eyes
AI is the new grifter buzzword 😐
Linus seems to be taking a perfectly pragmatic approach, given that AI is not going away short of WW3.
I can imagine some individual sub-system maintainers introducing various AI-roadblocks though.
I also expect this to be increasingly addressed (in general) with model & tooling improvements, giving more weight to higher quality reports and MRs, and more respect for project rules and processes. A mix of soft and hard gates, CLAs, improved early automated bug report & patch reviews and other CI gates.
model & tooling improvements, giving more weight to higher quality reports and MRs, and more respect for project rules and processes. A mix of soft and hard gates, CLAs, improved early automated bug report & patch reviews and other CI gates
This is exactly what I see happening inside the handful of large software companies into which I have visibility.
We have to deal with the fact that slop or not we all have 100x more security bugs than usual thanks to Mythos and friends, and some of them are real ones that will bit us in the ass if we don’t fix like, now
I hope this new adoption didn’t contributed to recent AMD GPU driver bugs
Pretty unlikely.
The way the kernel devs seem to be using LLMs isn’t really going to lead to any bugs that wound’t have snuck in anyways.
I’m not a tech bro or AI apologist. I’m choosing to skeptically trust Linus in this case. Based on my understanding of how it’s being used, I don’t think we have a lot to worry about.
That being said I wish he would STFU about it because the “Linus is pro-AI now guys” headlines are not doing the world any favors.
Using AI with Rust compiler is a combination that’s worked well for me many times already. You can check out https://slingshotspeeders.com/ if you want to experience compiled Rust in the browser using Bevy 0.19 for an orbital mechanics simulation racing game for free with no account easily.
Brain damage.
Torvalds was careful to frame this as a correction of language rather than a rejection of the technology itself.
Ok…
A maintainer burden few are talking about
Alongside the productivity gains, Torvalds described a less-discussed cost: a rising flood of low-effort, AI-generated bug and vulnerability reports hitting open source projects, including the kernel’s own security channels. He said it has become common for someone to run an AI tool against a codebase, file a report flagging a “possible” issue, and then vanish when a maintainer follows up asking for more detail or a fix.
Everyone is talking about that and it is surprising to me that he doesn’t recognize that as two side of the same coin. Why are you rejecting the low-effort, Ai generated bug reports, Linus, isn’t AI just a productivity tool? What’s wrong? Is it not being productive?
I say “surprising”, it’s not that surprising. It’s a “guns don’t kill people, people kill people” attitude. It’s willfully excluding the impact that technology has on people, and framing the outcome as coming purely from people and the existence of the technology not affecting the outcome.
Why are you rejecting the low-effort, Ai generated bug reports, Linus, isn’t AI just a productivity tool?
Because they are low effort. Good software engineering is never low effort, whether you use AI or not.
If you let AI replace people, and let them mess with software or report issues without understanding what’s going on, without putting in the effort to understand what’s going on, you’re not adding any value; just noise.
But if you use AI to find issues, then investigate the issue, reproduce it, document it, etc, then you’ve got something meaningful that someone can do something with.
But if you use AI to find issues, then investigate the issue, reproduce it, document it, etc, then you’ve got something meaningful that someone can do something with.
In principle that is sound, but LLMs can’t do that, yet, if ever. Sounds like anthropomorphism. The ‘hallucination’ thing is intrinsic. What you can do is point it at easily verifiable problems like searching for security flaws, and then verify them. Whether that’s cost effective in wasted energy and human time is still an open question, especially as token cost ramps up pre IPO.
That’s why you need to investigate and reproduce what the AI found: to verify it’s a real issue. Blindly submitting the issue is lazy noise.
They can find issues now though. Will they also find non issues due to missing context or not understanding the whole operational picture, well yes, but you can usually filter those ones out quickly. While i wish we had QA back and it be a norm again in software engineering as a whole, without it I’m happy to have something look for ways my code will break in unexpected manners that testers previously handled. Is it the same as having someone sit their and try and break your shit for the sole purpose of finding a spot where something will go completely off the rails, no its not but with a lack of true QA and rushed timelines that leave little room for quality QA yourself. I dont mind having something else that will try to gage my code from all angels and suggest areas that might have bugs I did not intend for or did not catch when I was reviewing my code. Especially considering I can do this without making users of my code the guinea pigs in lieu of an actual QA department. Sometimes you have to make the best of a shitty situation and honestly its an improvement over no QA which has been a norm for a while now so I’ll take the small wins where I can
Tools are just tools; there are good ways and bad ways to use them. “Grok, go find bugs in the Linux kernel” isn’t a good one. This isn’t hard to understand…
There were many, partially well-intentioned, mostly fame-seeking, one-off ‘projects’ ‘scanning’ public repos and flooding them with low-quality low-effort PRs and issues; things that no human ever read until the a maintainer did.
“They made a poison and sold us a cure” is absolutely a valid criticism; that’s two sides of the same coin. But pretending that he’s being hypocritical or that LLMs aren’t useful is just being wrong for the sake of hating AI
Tools are just tools
I don’t think that’s true, but we simply disagree on that.
I don’t think he’s hypocritical. He simply doesn’t have to deal with that negative impact he “warns” about. He is in support of using it as a tool and thinks the criticism is invalid. That’s a consistent perspective, even if I don’t agree with it.
If he used LLMs to create bug reports, but also complained about LLM written bug reports he gets, that would be hypocritical.
AI is useful as tool like he said, unfortunately these days in many organisations it is used as a replacement rather than a tool.
His conclusion was that as AI tools get more capable, the people who actually understand how systems work become more valuable, not less — because someone still has to review, maintain, and take responsibility for what the tools produce.
Humans still need to learn how to code.
… What?
Lazy people using tools in a shitty manner has been an age old problem.
This just lets it scale, which is the problem. There used to be a barrier to entry for those less knowledgeable or experienced, now there isn’t and that’s the result.
It is a productivity tool, and it’s being used as one, by both people that know how and people who don’t.
bingo. and unfortunately the lazy and inexperienced are using it more than others, or using it at a much higher degree. Linus even says it himself. they file a report and then “vanish” because they can’t answer follow up questions or clarifications because they don’t know how. if you deal with LLMs and the people that use them as much as I do for my job you see this constantly. daily even. they “build” something and when it’s reviewed and questioned the answers are just no where to be found because the “coder” can’t answer them.
Like the other day there’s this guy on Youtube who is a huge advocate for Omarchy for whatever reason. older guy but he’s always going on about how great Omarchy is. he posts a video about how “we” built some application for Omarchy but the entire video is him putting prompts into Claude Code and then walking away. never touches it, never writes a line of code. but there were glaring bugs with the build so I reached out to him, asked him how he was going to address these things and hoping he wouldn’t put this app out there because it’s clearly broken. works on a surface level but it WILL break. he essentially told me to “fuck off”.
you’re right an LLM is a tool but one that should be in the hands of those that know what they’re doing. you wouldn’t hand a hammer to a child and tell them to build a house. you wouldn’t hand a gun to a junkie and tell them they’re now a cop. I wouldn’t hand Vim to my mother and tell her to build a website. I feel LLMs should require human training just like any other tools we use. If you want to allow LLMs into your open source project, sure, have at it but it should be YOUR responsibility to vet and interview each individual that wishes to contribute utilizing said tools.
