I want to respect the guidelines for the good of the community, but at this point it isn't serving the community well for there to not be backlash against the rising flood of AI-generated garbage.
It was really truly bad enough when it was ~half the articles either being about AI directly or indirectly. Now it's that, plus half of it is written by Claude too.
What meaningful community is going to be left for these guidelines to protect?
Moderation needs to put their foot down in some cases, as a matter of necessity. Sometimes users need to put their foot down, too.
I'm all for banning AI slop articles. The HN guidelines were recently updated to address slop comments[1], but they have not put their foot down yet about slop articles.
I think HN is just fucked. A lot of people either genuinely don't see the problem with having a bunch of AI-generated slop garbage on the frontpage, or they are themselves posting it so they have a personal stake in not seeing anything wrong with it.
Don't be too surprised: there are literally comments that are just blatantly written by Claude on HN, which seem to be coming from human accounts that predate Claude. Which means that there are people here who, in trying to respond, actually ask Claude to basically do it for them. I find this utterly stunning and honestly, truly alarming. Even if the person behind the keyboard is technically alive, what exactly are they becoming? Are they even going to think for themselves, or will they just ask Claude what they're supposed to think from now on?
And as much as HN moderation has been genuinely pretty great at keeping the community under control with a relatively light touch, it's already too late. Dang and friends needed to do something much sooner, and they didn't. It literally doesn't matter what they do now, so there's no point in bugging them, not that I expect they would be interested in listening anyway.
I'm not going to make a lot of dramatic "I'm leaving Twitter" type comments, but I'm losing respect for HN's rules and guidelines the more I see this page overran with literal CRAP. And just so I can make my opinion clear, it's not crap because it's AI generated, it's crap because I can tell it's AI generated, full of fluff, cliches and a lack of substance.
It says a lot about the taste of the average person voting on HN that this is what we get now, and it fucking sucks because I don't really like any of the competing news aggregators either. I actually had to log in to post this comment because lately I've been staying logged out of HN and visiting less frequently now that I'm not sure what I get out of it.
At least I won't miss HN when the internet becomes an inaccessible hellscape in part due to AI crap outnumbering human posts 1000:1 and in part due to horrible legislation screaming ahead at breakneck speeds with literally no opposition from anybody.
Intelligence for HN posters is like boobs for strippers: everyone knows that bigger is better when it comes to the attention they seek, so if they are lacking, or feel inadequate in that department, they seek augmentation which anyone can tell is fake but seems to get the job done.
32 GiB of VRAM is possible to acquire for less than $1000 if you go for the Arc Pro B70. I have two of them. The tokens/sec is nowhere near AMD or NVIDIA high end, but its unexpectedly kind of decent to use. (I probably need to figure out vLLM though as it doesn't seem like llama.cpp is able to do them justice even seemingly with split mode = row. But still, 30t/s on Gemma 4 (on 26B MoE, not dense) is pretty usable, and you can do fit a full 256k context.)
When I get home today I totally look forward to trying the unsloth variants of this out (assuming I can get it working in anything.) I expect due to the limited active parameter count it should perform very well. It's obviously going to be a long time before you can run current frontier quality models at home for less than the price of a car, but it does seem like it is bound to happen. (As long as we don't allow general purpose computers to die or become inaccessible. Surely...)
New versions of llama.cpp have experimental split-tensor parallelism, but it really only helps with slow compute and a very fast interconnect, which doesn't describe many consumer-grade systems. For most users, pipeline parallelism will be their best bet for making use of multi-GPU setups.
Yeah, I was doing split tensor and it seemed like a wash. The Arc B70s are not huge on compute.
Right now I'm only able to run them in PCI-e 5.0 x8 which might not be sufficient. But, a cheap older Xeon or TR seems silly since PCI-e 4.0 x16 isn't theoretically more bandwidth than PCI-e 5.0 x8. So it seems like if that is really still bottlenecked, I'll just have to bite the bullet and set up a modern HEDT build. With RAM prices... I am not sure there is a world where it could ever be worth it. At that point, seems like you may as well go for an obscenely priced NVIDIA or AMD datacenter card instead and retrofit it with consumer friendly thermal solutions. So... I'm definitely a bit conflicted.
I do like the Arc Pro B70 so far. Its not a performance monster, but it's quiet and relatively low power, and I haven't run into any instability. (The AMDGPU drivers have made amazing strides, but... The stability is not legendary. :)
I'll have to do a bit of analysis and make sure there really is an interconnect bottleneck first, versus a PEBKAC. Could be dropping more lanes than expected for one reason or another too.
You could fit your HEDT with minimum RAM and a combination of Optane storage (for swapping system RAM with minimum wear) and fast NAND (for offloading large read-only data). If you have abundant physical PCIe slots it ought to be feasible.
Unfortunately it really is running this slow with Llama.cpp, but of course that's with Vulkan mode. The VRAM capacity is definitely where it shines, rather than compute power. I am pretty sure that this isn't really optimal use of the cards, especially since I believe we should be able to get decent, if still sublinear, scaling with multiple cards. I am not really a machine learning expert, I'm curious to see if I can manage to trace down some performance issues. (I've already seen a couple issues get squashed since I first started testing this.)
I've heard that vLLM performs much better, scaling particularly better in the multi GPU case. The 4x B70 setup may actually be decent for the money given that, but probably worth waiting on it to see how the situation progresses rather than buying on a promise of potential.
A cursory Google search does seem to indicate that in my particular case interconnect bandwidth shouldn't actually be a constraint, so I doubt tensor level parallelism is working as expected.
But on my 3x 1080 Ti 1x TITAN V getto machine I learned that multi gpu takes a lot of tuning no matter what. With the B70, where Vulkan has the CPU copy problem, and SYCL doesn't have a sponsor or enough volunteers, it will probably take a bit of profiling on your part.
There are a lot of variables, but PCIe bus speed doesn't matter that much for inference, but the internal memory bandwidth does, and you won't match that with PCIe ever.
To be clear, multicard Vulkan and absolutely SYCL have a lot of optimizations that could happen, but the only time two GPUs are really faster for inference is when one doesn't have enough ram to fit the entire model.
A 3090 has 936.2 GB/s of (low latency) internal bandwidth, while 16xPCIe5 only has 504.12, may have to be copied through the CPU, have locks, atomic operations etc...
For LLM inference, the bottleneck just usually going to be memory bandwidth which is why my 3090 is so close to the 5070ti above.
LLM next token prediction is just a form of autoregressive decoding and will primarily be memory bound.
As I haven't used the larger intel GPUs I can't comment on what still needs to be optimized, but just don't expect multiple GPUs to increase performance without some nvlink style RDMA support _unless_ your process is compute and not memory bound.
I don't care anymore, if it happens to violate HN guidelines: Please, authors. Please write your own damn articles. We can absolutely tell that you're using Claude, I promise. (I mean, it might not be Claude specifically this time, but frankly I'd be willing to bet on it.) The AI writing is like nails on a chalkboard to me.
The worst part is the phrases don't actually mean anything. It's the LLM equivalent of flowery prose. The author admitted below that the article was Claude. So there you go.
I tried to set up a partner account for driver signing last year (as a business entity) and it already seemed basically impossible. I think they're getting ready to just simply not allow it at all.
This is stupid. If Microsoft wants people to stop writing kernel drivers, that's potentially doable (we just need sufficient user mode driver equivalents...) but not doing that and also shortening the list of who can sign kernel drivers down to some elite group of grandfathered companies and individuals is the worst possible outcome.
But at this point I almost wish they didn't fix it, just to drive home the point harder to users how little they really own their computer and OS anymore.
> Coloring book will be available only on Copilot+ PCs. To use Coloring book, you will need to sign in with your Microsoft account.
Can someone please explain why these two things are ever simultaneously true? You buy the stupid Copilot+ PC that has "AI" NPU hardware, right? So the AI features should be able to run locally. But if you have to sign in with a Microsoft account, then surely, it doesn't run locally, which begs the question, why does it require a Copilot+ PC at all?
Not even going to bother asking "does anybody want this to begin with" because at this point there is no real need to bother asking that.
I feel the deepest existential pain in my heart that despite companies being 'all in' on AI, they can't integrate anything meaningful that would make sense to the end use, but would require even a braincells worth of mental effort of ML expertise or actual requirements people have.
My two favourite 'AI' tools in image editing have been ones that can replace tedious work.
One such example are segmentation models that can be used for smart cutouts, removing backgrounds etc.
Now we have both 'segmentation' and 'AI' in paint - but the segmentation uses the exact same shitty flood fill with tolerance that's probably existed in the first paint program at Xerox PARC, while the 'AI' feature is another by-the-numbers crappy stable diffusion model that's strictly worse than anything you could get with your first Google search.
I read somewhere that even for the "Copilot" things that supposedly run locally, windows needs to send a request to microsoft to confirm if your input is allowed by their rules.
Microsoft really want to force you to log in with an MS account, as well as slurp all your documents into the spy cloud, and they keep pushing back on the various ways round this people have found.
(I found an odd one: for some reason I can't log into my PC with my MS account, which let me create the local account I actually wanted. System broken in my favor.)
The best way is to set up samba on a Linux machine, even a raspberry pi, and create a domain. Then you can create group policy to turn off a lot of nonsense and set up your computer by connecting to the domain. No MS account required, although you can associate one of you like.
Windows feels like it has a lot of attrition from home users now and perhaps it is only a matter of time before it's no longer worth writing exclusive software for it.
The reason we're getting this AI gumbo is that obviously the product people at M$ we're told: "Make money by selling AI features!!!". Which flipped their minds from their usual "I am Steve Jobs" fantasies, which tell them to _consider the User experience first_, to _Consider the companie$ experience first_, and they can't keep the two concepts in their little heads at the same time because they are, after all, just product people.
Unfortunately, some packages which are free software can only be found in nonguix, due to Guix wanting to be able to build everything from source. So things like Gradle that require a huge bootstrap chain that no one yet has bothered to do are instead packaged in nonguix as a binary download.
The Google Pixel 10 can give you notifications when your location is tracked in this manner as well. I turned it on and have been notified a few times.
It is interesting that we let this happen. Modern phones are very useful devices, but they're not really mandatory for the vast majority of people to actually carry around everywhere they go, in many cases they merely add some convenience or entertainment, and act to consolidate various other kinds of personal devices into just one. If you wanted, you could more often than not avoid needing one. Yet, we pretty much all carry one around anyways, intentionally, and this fact is somewhat abused because it's convenient.
Having watched a fair bit of police interrogations videos recently (don't knock it, it can be addicting) I realized that police have come to rely on cell phone signals pretty heavily to place people near the scene of a crime. This is doubly interesting. For one, because criminals should really know better: phones have been doing this for a long time, and privacy issues with mobile phones are pretty well trodden by this point. But for another, it's just interesting because it works. It's very effective at screwing up the alibi of a criminal.
I've realized that serious privacy violations which actually do work to prevent crime are probably the most dangerous of all, because it's easy to say that because these features can help put criminals behind bars, we should disregard the insane surveillance state we've already built. It's easy to justify the risks this poses to a free society. It's easy to downplay the importance of personal freedoms and privacy.
Once these things become sufficiently normal, it will become very hard to go back, even after the system starts to be abused, and that's what I think about any time I see measures like chat control. We're building our own future hell to help catch a few more scumbags. Whoever thinks it's still worth it... I'd love to check back in in another decade.
The entire point of the free software movement is to promote free software principles and software rights. What I think many Linux distributions would prefer is a model where companies who do benefit from selling software and hardware are funding them indirectly, so they can focus on continuing to promote free software in a more neutral way, without the pressures and potentially misaligned incentives that come from running a store front can bring.
There are distributions like elementary OS which are happy to sell you things with this model, though, but I just don't think it's surprising many distributions would actively prefer to not be in this position even if it leaves money on the table. This sort of principled approach is exactly why a lot of us really like Linux.
It's really unfortunate the term "free software" took off rather than e.g. "libre software", since it muddies discussions like this. The point of "free software" is not "you don't have to pay," it's that you have freedom in terms of what you do with the code running on your own machine. Selling free software is not incompatible with free software: it's free as in freedom, not as in free beer.
Nobody in this comments thread appears to be confused by or misusing the term "free software". We're talking about free software vs (commercial) proprietary software.
> I am still surprised most Linux Distros haven't changed their package managers to allow for selling of proprietary solutions directly
Free packages remain unaffected, but now there are optional commercial options you can pay for which fund the free (as in free money) infrastructure you already take advantage of so that these projects are fully sustainable. I imagine some open source projects could even set themselves up to receiving donations directly via package managers.
I promise you, everybody understands the general idea, but adding a built-in store to your operating system is far from a neutral action that has no second- or third-order effects. It isn't that it somehow affects "free" packages. Incoming text wall, because I am not very good at being terse.
- It creates perverse incentives for the promotion of free software.
If development of the operating system is now funded by purchases of proprietary commercial software in the app store, it naturally incentivizes them to sell more software via the app store. This naturally gives an incentive to promote commercial software over free software, contrary to the very mission of free software. They can still try to avoid this, but I think the incentive gets worse due to the next part (because running a proper software store is much more expensive.)
Free software can be sold, too, but in most cases it just doesn't make very much sense. If you try to coerce people into paying for free software that can be obtained free of charge, it basically puts it on the same level as any commercial proprietary software. If said commercial software is "freemium", it basically incentivizes you to just go with the freemium proprietary option instead that is not just free software, but also often arguably outright manipulative to the user. I don't really think free software OS vendors want to encourage this kind of thing.
- It might break the balance that makes free software package repositories work.
Software that is free as in beer will naturally compete favorably against software that costs money, as the difference between $0 and $1 is the biggest leap. Instead of selling software you can own, many (most?) commercial software vendors have shifted to "freemium" models where users pay for subscriptions or "upsells" inside of apps.
In commercial app stores, strict rules and even unfair/likely to be outlawed practices are used to force vendors to go through a standardized IAP system. This has many downsides for competition, but it does act as a (weak) balance against abusive vendors who would institute even worse practices if left to their own devices. Worse, though, is that proprietary software is hard to vet; the most scalable way to analyze it is via blackbox analysis, which is easily defeated by a vendor who desires to do so. Android and iOS rely on a combination of OS-level sandboxing and authorization as well as many automated and ostensibly human tests too.
I am not trying to say that what commercial app stores do is actually effective or works well, but actually that only serves to help my point here. Free software app stores are not guaranteed to be free of malware more than anything else is, but they have a pretty decent track record, and part of the reason why is because the packaging is done by people who are essentially volunteers to work on the OS, and very often are third parties to the software itself. The packages themselves are often reviewed by multiple people to uphold standards, and many OSes take the opportunity to limit or disable unwanted anti-features like telemetry. Because the software is free, it is possible to look at the actual changes that go into each release if you so please, and in fact, I often do look at the commit logs and diffs from release to release when reviewing package updates in Nixpkgs, especially since it's a good way to catch new things that might need to be updated in the package that aren't immediately apparent (e.g.: in NixOS, a new dlopen dependency in a new feature wouldn't show up anywhere obvious.)
Proprietary software is a totally different ball game. Maintainers can't see what's going on, and more often than not, it is simply illegal for them to attempt to do so in any comprehensive way, depending on where they live.
If the distributions suddenly become app store vendors, they will wind up needing to employ more people full time to work on security and auditing. Volunteers doing stuff for free won't scale well to a proper, real software store. Which further means that they need to make sure they're actually getting enough revenue for it to be self-sustaining, which again pushes perverse incentives to sell software.
What they wanted to do is build a community-driven OS built on free software by volunteers and possibly non-profit employees, and what they got was a startup business. Does that not make the problem apparent yet?
- It makes the OS no longer neutral to software stores.
Today, Flatpak and Steam are totally neutral and have roughly equal footing to any other software store; they may be installed by default in some cases, but they are strictly vendor neutral (except for obviously in SteamOS). If the OS itself ships one, it lives in a privileged position that other software store doesn't. This winds up with the exact same sorts of problems that occur with Windows, macOS, iOS and Android. You can, of course, try to behave in a benevolent manner, but what's even better than trying to behave in a benevolent manner is trying to put yourself in as few situations as possible to where you need to in order to maintain the health of an ecosystem. :)
--
I think you could probably find some retorts to this if you wanted. It's not impossible to make this model work, and some distributions do make this model work, at least insofar as they have gotten now. But with that having been said, I will state again my strongly held belief that it isn't that projects like Debian or Arch Linux couldn't figure out how to sell software or don't know that they can.
It was really truly bad enough when it was ~half the articles either being about AI directly or indirectly. Now it's that, plus half of it is written by Claude too.
What meaningful community is going to be left for these guidelines to protect?
Moderation needs to put their foot down in some cases, as a matter of necessity. Sometimes users need to put their foot down, too.
reply