To state this again [LWN.net]

To state this again

Posted Jul 3, 2026 8:10 UTC (Fri) by rweikusat2 (subscriber, #117920)
Parent article: Two LLM-assisted memory-management patch sets

I'll stop subscribing to LWN unless you stop this mindless LLM marketing. I'm not interested in being sold these "warez."

to post comments

To state this again

Posted Jul 3, 2026 10:49 UTC (Fri) by Karellen (subscriber, #67644) [Link]

Just to counter this, I am against LLMs in general, and perplexed at their uptake by Free Software projects. But I am still interesting in finding out which projects are using them, and why, and getting a well-researched outside view of what results those projects are actually seeing. Then, if I do end up in a discussion about LLMs my position can be well-grounded and doesn't necessarily come across as plain ranting. Although I can often end up doing that anyway, at least it will be well informed ranting :-) I appreciate the reporting.

To state this again

Posted Jul 3, 2026 12:21 UTC (Fri) by daroc (editor, #160859) [Link] (1 responses)

We're really not mindlessly pro-LLM.

But whether you approve of the use of LLMs or not, there is certainly a lot of LLM-related news happening right now. Open source projects dealing with debates over their development policies is definitely in our typical topic area, even when doing so involves discussing LLMs. And we continue to cover plenty of non-LLM-related news.

That said, of course you're entitled to stop subscribing if you no longer find value in our news coverage; we'll be sad to see you go, but we do very much appreciate the support that you've given us so far.

To state this again

Posted Jul 3, 2026 19:51 UTC (Fri) by kleptog (subscriber, #1183) [Link]

I think LLMs are the most interesting thing to happen in software development in a long time. I don't know how this will all turn out, but I do know that LWN will keep me informed of the impact. The comments here are (mostly) at least well informed, even if I don't agree with them.

Keep up the reporting.

To state this again

Posted Jul 3, 2026 13:41 UTC (Fri) by vbabka (subscriber, #91706) [Link]

Calling this well-researched article "mindless LLM marketing" is frankly an insult to Jon and the rest of the LWN team and completely uncalled for.

To state this again

Posted Jul 3, 2026 13:51 UTC (Fri) by ljsloz (subscriber, #158382) [Link]

I personally find the blanket 'LLMs are bad and evil and I don't want to hear any more' approach to be actively harmful.

By taking a position which is contradicted by reality on the ground (they are effective, there are suitable ways of using them, they do provide a productivity boost albeit far less than claimed), you lose any ability to do anything about the extremely negative aspects (slop, climate impact, imperiling of the entire world economy).

Because you essentially put yourself in a fantasy world rather than dealing with reality which is messy, doesn't clearly have 'good' and 'bad' and requires nuanced thought.

Naturally we all want to be emotionally satisfied by binary thinking. But that's just not how the world is.

I gently suggest that you pay attention to those actually doing the work and experiencing the realities and who are taking risks to push back on the negatives, rather than immediately jumping to an emotional position.

And to suggest that Jon or LWN or the team in any way blindly promotes AI is so at odds with the facts (including LWN being repeatedly DDoS'd by AI scrapers) it's silly.

You should keep your subscription because, as an actual kernel developer, I can tell you that LWN is the _only_ press I've seen be consistently accurate and insightful on linux, the kernel, and open source as a whole. It's not even close.

We should all remain very grateful for the work Jon + the team does here.

To state this again

Posted Jul 4, 2026 12:26 UTC (Sat) by aimannajjar (subscriber, #184277) [Link] (13 responses)

Seriously, LWN seems compromised. Well, Big Tech has been trying to change Linux philosophy for awhile, maybe they finally did it? Look at Ubuntu acting exactly like Microsoft. Now we've got a bunch of Satya Nadellas as maintainers. Too much pragmatism does have a cost

To state this again

Posted Jul 4, 2026 14:51 UTC (Sat) by corbet (editor, #1) [Link] (12 responses)

I have to admit to being entirely mystified by these comments.

These tools are being used. One of the points of this article was that they are being used to create code at the core of the kernel. Should we pretend that this is not happening? Somehow I don't think that would serve our readers well. LLMs will not go away just because we choose to ignore them.

To state this again

Posted Jul 4, 2026 15:19 UTC (Sat) by koverstreet (subscriber, #4296) [Link] (11 responses)

There's definitely been an increase in trolling and outright hysteria, it's something I've been having to take time to deal with too.

I think people's feelings on the AI companies at the forefront are entirely understandable (and I share a lot of those!); I wonder if it would help to highlight what's been going on with open weights LLMs. The latest open weights LLMs are starting to match or maybe even surpass the closed ones, the open software stacks are becoming much more capable, and the hardware required is going to quickly drop in price by an order of magnitude - the DRAM bubble will pop, there's a ton of fabs coming online now or very soon, and it's early days for the hardware architectures still. The only remaining piece we're dependent on giant corporations for is training, and not for long on that one, I think.

Back in the early, early days of open source it was hard to get access to hardware - you needed to be affiliated with a university, or a company - or even a good compiler. It took time for stuff to get cheaper, and to democratize (and then turn into a different sort of lovely chaos). The development of AI is a bit reminiscent of that...

The world is always changing, the current situation is never a permanent state of affairs :)

To state this again

Posted Jul 4, 2026 16:50 UTC (Sat) by aimannajjar (subscriber, #184277) [Link] (10 responses)

I apologize if I came across too harsh. I understand you are just reporting on LLM use in Linux kernel. While I think my comment was uncalled for, I think Linux maintainers might be missing some crucial aspect regarding the "public perception of AI" and how that might impact Linux image, especially when people are migrating their desktops to it. Microsoft, and to lesser extent Microsoft, made these mistakes and the backlash was strong.

I believe much of the toxicity in AI discourse is due to the persistent "human-replacement" talking point championed by Big Tech CEOs, and their recurrent doubling-down on creating false hype by misrepresenting technical "achievements".

All experienced programmers know that all LLM agents will inevitably (and repeatedly) choke in any serious non-localized code change requests, we also know there is no serious "national security risks" with Mythos or Fable, we also know that LLMs will not "self-improve themselves" in any meaningful way using this technology, not when they cannot even guarantee production databases safety without humans in the loop.

Additionally all serious financial experts know that tech layoffs were not due to "AI productivity gains" but rather due to the irresponsible spending adventures and COVID-era fall outs.

Yet despite all that, Big Tech, with assistance from media and US government, is insisting on inflicting lasting damage on 1) real humans, 2) societal stability and 3) even small open-source projects by creating an extremely false narrative that does not reflect true capabilities and benefits of LLM agents.

When a large trusted open-source project is seen helping (in any way) those Big Tech's clearly-dark and psychotic ambitions, I think many people will feel disillusioned by such a free helping hand.

I agree local models will feel like a good converging point of this messy journey, even cloud LLM services becoming a commodity service and not the hyped "gold mine" being portrayed is a satisfying end result to see. But I don't think the toxicity will go away that easily, not when Big Tech, US media and government are still acting like a psychotic propaganda machine - and certainly not when hundred of thousands of people are still suffering from economic downfalls of that irresponsible behavior by those elites.

This is all to say, even if we wanted to be pragmatic about LLM use in Linux kernel, the toxicity that will come with that should be taken into account, especially when for the first time ever, the Year of the Linux Desktop is within reach.

To state this again

Posted Jul 4, 2026 16:53 UTC (Sat) by aimannajjar (subscriber, #184277) [Link]

* Sorry, meant to say "Microsoft, and to a lesser extent Canonical, ..

To state this again

Posted Jul 4, 2026 17:21 UTC (Sat) by koverstreet (subscriber, #4296) [Link]

100% agree with everything you have to say about Big Tech, and I could add a lot more of my own :)

But I was interested in AI long before the AI companies came along, because understanding how consciousness and our own minds work is even more fascinating to me than understanding how computers work; it builds on a lot of mathematics and computability theory going back 100 years or more, and takes it to the next level, and I find it absolutely beautiful.

And now we've got AIs that are legitimately capable of doing useful work, and I'm getting to see the realization of some of my own old ideas, and doing my own research (with the little time I have left over after bcachefs) into things that I wanted to be working on 20 years ago but never thought we'd crack. And in my own work, things that a year ago clearly needed to happen but looked like 10+ year projects (converting bcachefs to Rust, so I can keep making it more reliable and bulletproof, then formal verification) are now... completely tractable.

This is bloody amazing!

Yes, big tech absolutely sucks, and it's gotten worse over the course of my career, I'm seeing that too. But don't let it dominate your worldview and get you down.

The basic spirit of open source and free software was that individuals can make a difference, freely collaborating, just trying to make things better, and that hasn't changed. We'll always still be here, as long as we keep it alive.

To state this again

Posted Jul 4, 2026 22:16 UTC (Sat) by Cyberax (✭ supporter ✭, #52523) [Link] (7 responses)

> All experienced programmers know that all LLM agents will inevitably (and repeatedly) choke in any serious non-localized code change requests, we also know there is no serious "national security risks" with Mythos or Fable, we also know that LLMs will not "self-improve themselves" in any meaningful way using this technology, not when they cannot even guarantee production databases safety without humans in the loop.

First, we don't actually know that LLMs can't self-improve. Fifteen years ago we also "knew" that things like natural-language queries and machines passing the Turing test is just a crazy talk.

Second, even _with_ the current limitations the LLMs are still powerful tools. If nothing else changes and the LLM development stays frozen forever, existing models _already_ require significant adjustments in the development process.

So yeah, LWN providing overview of AI effects on the kernel development is very interesting and important.

To state this again

Posted Jul 4, 2026 23:44 UTC (Sat) by aimannajjar (subscriber, #184277) [Link] (6 responses)

LLMs can't self-improve because they're shown to be highly unreliable, i.e they hallucinate with high degree of confidence. Being both non-deterministic and highly unreliable is clear evidence they lack "intelligence" that is required to achieve self improvement.

Many pioneer researchers have said LLM as a technology is a dead-end for the stated goal of those AI companies, including things like AGI, and the so-called self-improvement.

LLMs are useful as tools, correct - too bad they didn't market them as such, I guess?

To state this again

Posted Jul 5, 2026 0:57 UTC (Sun) by koverstreet (subscriber, #4296) [Link] (5 responses)

Have you met the average human? :)

Humans guess, confabulate and outright hallucinate all the time. Tracking how well we know the things we think we know and why is a skill that takes effort to develop, and many never develop at all.

The big obstacle with LLM self-learning is a thing called the credits problem.

To state this again

Posted Jul 5, 2026 2:28 UTC (Sun) by aimannajjar (subscriber, #184277) [Link] (1 responses)

Yeah but humans can recognize when they have made a mistake and self-correct. The ones who are stuck in a hallucination cycle typically require medical intervention or end up in a self-destructive loop if they refuse help :)

LLMs can happily continue iterating with compounding hallucinations, not only that, LLMs that are trained on other LLM output degenerate and done repeatedly it leads to model collapse.

To state this again

Posted Jul 5, 2026 5:40 UTC (Sun) by koverstreet (subscriber, #4296) [Link]

> Yeah but humans can recognize when they have made a mistake and self-correct. The ones who are stuck in a hallucination cycle typically require medical intervention or end up in a self-destructive loop if they refuse help :)

You've never seen an LLM realize it made a mistake and self-correct?

> LLMs can happily continue iterating with compounding hallucinations, not only that, LLMs that are trained on other LLM output degenerate and done repeatedly it leads to model collapse.

That's the credits problem I was referring to.

To state this again

Posted Jul 6, 2026 6:41 UTC (Mon) by anselm (subscriber, #2796) [Link] (1 responses)

Calling what LLMs do “hallucination” is misleading because in humans, hallucinations are issues with spurious input processing whereas with LLMs, “hallucinations” are an output problem. A more appropriate technical term would be “bullshitting” (as per Harry G. Frankfurt) – the model just produces output regardless of its veracity, because LLMs have no knowledge, or concept of “correct” or “wrong”; all they're designed to do is to come up with a probable continuation of their output stream given their weights and context.

Of course bullshitting is something that some people do, too, but it's not generally considered a desirable trait in one's friends.

To state this again

Posted Jul 6, 2026 13:13 UTC (Mon) by aimannajjar (subscriber, #184277) [Link]

That's an interesting perspective and I agree with it. Generally speaking, anthropomorphizing those models is the most annoying thing those companies have chosen to do, Anthropics literally is now talking about models ability to "dream".

I've seen research papers by AI labs on "How to detect when AI lie", I asked the researcher presenting the paper "how is lying different from hallucinating?" - he answered (paraphrasing): "in the first the model is aware of the correct answer but is choosing to give the wrong answer". Those companies are suffering from an advanced form of psychosis when they think models have "choices" or "intentions". What's wrong with just saying "failure" or "error" like any other software?

I think it reveals something deeper about their view of the world, and their visions for it. It also explains the reckless spending chasing AGI and the toxic AI fear-mongering, which of course all is backfiring now.

To state this again

Posted Jul 6, 2026 14:56 UTC (Mon) by ballombe (subscriber, #9523) [Link]

Humans have more or less consistent personality. If you find that someone to be technically inept you will ignore them.
They have stake in their professional reputation.

On the other hand, LLM are totally unpredictable.