I don’t know, what was stopping us from solving world hunger and housing before LLMs? There’s always some fresh new way for assholes to make life difficult for all of us, for the benefit of the wealthy few. I don’t really see what LLMs have to do with that, other than being in vogue at the moment among the grifter class. It’s like seeing a pyramid scheme and deciding that the solution is to ban pyramids. The focus should be on the actual malicious behavior, not whatever trinket they’re using to disguise it nowadays.

I think you’re missing the trees for the forest. Yes, all those downsides make the current crop of LLM implementations pretty shitty for humanity, on balance, but none of them are intrinsic to LLMs, they are intrinsic to the current business practices that are attempting to profit from LLMs. LLMs are just math. Often computationally expensive math, sure, though they needn’t be. But I don’t think it helps anyone to say “LLMs are intrinsically bad and no one should use them” any more than it makes sense to say that blockchains are intrinsically bad and no one should use them (a common sentiment about a decade ago). I think a more realistic approach is to find out what real benefit can be had by using LLMs, and see if that benefit can be realized without all the downsides you mentioned.

To take the case at hand, I think open source code review is probably a pretty good candidate domain for exploration, because there’s a large corpus of available text that could be obtained ethically, if LLM trainers bothered to put in the work. It seems like Torvalds isn’t concerned with that, and it looks like sashiko only supports LLMs that are known to have been trained on stolen data, which is shitty and I think Torvalds and the Linux maintainers should be taken to task for that. I think LLM makers should get permission from project maintainers before training on their codebases, just as natural-language rightsholders should be asked for permission before using their works for training. I think LLMs trained on ethically sourced data should be permissively licensed and, if they collect revenue, should be expected to kick some of that revenue upstream to the projects they benefit from. I also think such projects could be powered ethically: I have some plans for my own small solar powered server that will run batch jobs when production is higher than my storage capacity. Right now I’m mostly targeting Folding at Home, but I might well run my own LLM trainer if I get enough capacity, and I could see a similar distributed processing network for high-cost jobs from trusted open source LLM projects.

I don’t think you’re wrong about the scope of the downsides, but I also don’t think it makes sense to take them all as a singular block and judge all possible LLM tools by all possible downsides. I think they are problems that can be engineered around with technical and social guidelines for use. And I think it will be down to the open source community to set those guidelines, because I don’t know of any other group that has the expertise and the motivation to do so.

I think asbestos is a good analogy, but I think your claims that the inherent downsides of LLMs outweigh any possible upsides are as-yet unfounded. I also think it’s kind of strange that you assume anyone who thinks otherwise is being disingenuous.

Maybe they are too dangerous for broad use, and we need to regulate them like asbestos, or uranium. Maybe they shouldn’t be used outside of a laboratory setting or by anyone who doesn’t have extensive training with how to interact with them safely. It seems pretty clear that Torvalds has decided they’re worth the downside, and while I don’t know if that’s a good call, I don’t think he’s operating in bad faith to the detriment of the kernel project. That doesn’t sound like something he would do.

I feel confident that I don’t have the expertise to say for certain one way or the other, though the experience I do have with software tools makes me think there’s probably an application for them where the downsides can be mitigated to the point where they become worthwhile. I think there’s probably some single-purpose or tailored application of LLMs related to textual analysis that don’t require the theft of the whole internet, and don’t require insane amounts of energy to run. I don’t think we have discovered them yet (at least I haven’t), but saying “this kind of software is only bad and can have no ethical uses, ever” seems premature.

Yeah, you put your finger on maybe the biggest problem that I don’t have a good answer to yet. The ability to interact with LLMs in natural language comes from an analysis of huge tracts of contemporary human writing, and there’s not an ethical corpus that can compare in scope to the corpus of just snatching the whole internet. Maybe they don’t need to be strictly “natural” language interactions? Maybe a sufficient ethical corpus could be compiled and maintained somehow? I don’t know, this is kinda where I get out of my depth.

I think if we, as a community, really put our heads together, we could figure out how to define and make useful user-respecting tools that incorporate LLMs. There are a lot of hard problems, like the massive power consumption, and the ethical use of data, that I don’t really know how to solve.

I think a start would be to focus on making smaller, lower-complexity models that are built for purpose, rather than trying to make a jillion-parameter jack-of-all-trades trades model. I think it would also make sense to focus initially on areas where there are already large corpuses of freely available text, like all the writings in the public domain. But I don’t really have a good idea of what these tools would be used for, exactly, which is where I’m stuck.

He doesn’t really specify use cases beyond Sashiko (I believe that’s the project they’re discussing), which, if I understand correctly, uses an LLM to review kernel patches using a pre-set list of prompts. The output of those prompts presumably gets added to the PR for anyone who cares to read it.

These outputs sometimes have mistakes in them, so the maintainers must read them with a critical eye, but I think Linus is arguing here that it’s sufficiently… “good enough” that including them is helpful to the maintainers, that it catches enough real stuff without wasting too much time on fake stuff. There are always tradeoffs when adding a tool to the stack, and sufficiently rare misfires are probably worth the times it catches real problems that humans miss.

The solution is to make sure those LLM tools help maintainers instead of just causing them pain. There’s no question on that side.
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.
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.
The kernel project has been and will continue to be about the technology.
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.
This is NOT some kind of “social warrior” project, never has been, and never will be.
In the kernel community we do open source because it results in better technology, not because of religious reasons.
And so we make decisions primarily based on technical merit. Not fear of new tools.

IMO it’s a pretty clear eyed take. LLMs are tools and, like all tools, are more fit for some purposes than others. And, like many tools, they’re hard to understand, easy to misuse, and dangerous when they are misused (intentionally or otherwise). And, like any tool, they can be used for evil if the user has a mind to do so. Just because they are currently being used to do great harm doesn’t mean they can’t also be put to good use.

I think the FOSS community, in particular the Fediverse community, does itself a disservice by refusing to engage with them at all. They aren’t evil in and of themselves, they’re just new and selfish people will try to find an angle with them, same as it’s always been. I think it’s up to us to understand them and try to find a way to use them ethically. Show the world that it’s not that we invented scary Artificial Intelligence, it’s just that we created a cool, if wonky, search interface for text. Wonky in what way? Well, I think it’d be nice if there were robust FOSS tools and documentation I could use and/or contribute to, and learn about them with a community of my peers working for the common good, instead of having to either take the megacorps at their word or just renounce anything “AI branded” altogether, which doesn’t feel like a good long term strategy.

I think it’s a network problem. Most people use Chrome, therefore most people’s friends and co-workers use Chrome, so there’s a strong incentive to go with the pack. Google is obviously exploiting this to get their ecosystem claws in deeper, but I think the main thing is that Google effectively dethroned Microsoft as the default, and Mozilla has just kind of “been around” the whole time as the weird alt browser that the IT guy uses, even being propped up by Google in exchange for Firefox users’ search data.

What should Mozilla do? IDK. I don’t think “more AI” is the right answer, but I also don’t know what I would do in their position. It’s a tight spot.

They’re called data centers because they… do actually have lots of meaningful data. Tons of it. A lot of it is stolen, most of it has no business being in these corporations’ hands, and none of it is available to you unless you can cut a big enough check, but it is real data. Why do you think palantir has been worming its way into every government agency? Its because they want valuable data, and they process that data in data centers.

My town does have a website but it doesn’t do much other than list phone numbers and office hours for a few departments. what services are available on it are contracted out to corporate partners. I would not be surprised if the website itself is managed by a corporation as well. It is mostly useless and there is very little motivation to improve it.

But I’m not really talking about a municipal datacenter, more like a community center or library branch having a digital commons for the neighborhood, with some useful tools, access to reliable data, and maybe some recreational software. Something like what Nextdoor should have been but not enshittified to death by VC.

I guess it depends on how much equipment is needed before something becomes a “datacenter”. I don’t really see a community hub being a massive supercomputer, maybe a small office or a room in a local library. I’d argue that a single server could be called a datacenter if it centralizes data, but I don’t think that is the common understanding of the word. Maybe the word needs to be reclaimed.

I would love a (solar-powered) community datacenter that hosts services for the local population. Community bulletin board or forum to share event notices, lost pets, road closures etc, simple messaging and filesharing utilities for those not technical enough to host their own, maybe some simple games like chess or cards.

The problem with the current explosion of datacenters is that they don’t benefit the community at all, they’re just digital oil rigs that drain the community of resources while also actively poisoning the area they’re in. Small wonder communities are against them.

can’t wait to flee a pack of these on my ATV after robbing a megafarm for my scrappy resistance group.

Right, i mean if you made the context window enormous, such that you can include the entire set of embeddings and a set of memories (or maybe, an index of memories that can be “recalled” with keywords) you’ve got a self-observing loop that can learn and remember facts about itself. I’m not saying that’s AGI, but I find it somewhat unsettling that we don’t have an agreed-upon definition. If a for-profit corporation made an AI that could be considered a person with rights, I imagine they’d be reluctant to be convincing about it.

Yeah I think for it to be a proper strange loop (if that is indeed a useful proxy for consciousness-- I think there’s room for debate on that) it would need to be able to take it’s entire “self” i.e. the whole model, weights, and all memories, as input in order to iterate on itself. I agree that it probably wouldn’t work for the current commercial applications of LLMs, but it not what being what commercial LLMs do, doesn’t mean it couldn’t be done for research purposes.

There’s no reason an LLM couldn’t be hooked up to a database, where it can save outputs and then retrieve them again to “think” further about them. In fact, any LLM that can answer questions about previous prompts/responses has to be able to do this. If you prompted an LLM to review all of it’s database entries, generate a new response based on that data, then save that output to the database and repeat at regular intervals, I could see calling that a kind of thinking. If you do the same process but with the whole model and all the DB entries, that’s in the region of what I’d call a strange loop. Is that AGI? I don’t think so, but I also don’t know how I would define AGI, or if I’d recognize it if someone built it.

That’s true, maybe “Yet Another Discourse Alternative”? Discussion Alternative? I just like the idea of a chat platform whose acronym is YADA.

IME, code projects either die or live long enough that you think of a better name for them long after a name change becomes not worth the effort. Naming things is hard 🤷‍♂️

Yet Another Discord Alternative would be a better name than Harmony.

This is in some ways an easier problem than classifying LLM vs non-LLM authorship. That only has two possible outcomes, and it’s pretty noisy because LLMs are trained to emulate the average human. Here, you can generate an agreement score based on language features per comment, and cluster the comments by how they disagree with the model. Comments that disagree in particular ways (never uses semicolons, claims to live in Canada, calls interlocutors “buddy”, writes run-on sentences, etc.) would be clustered together more tightly. The more comments two profiles have in the same cluster(s), the more confident the match becomes. I’m not saying this attack is novel or couldn’t be accomplished without an LLM, but it seems like a good fit for what LLMs actually do.