Even in quantized form, reducing that size by 2 times or 4 times, a Raspberry Pi 5 could not run it otherwise than by reading the weights from an SSD.
Even thus, I do not believe that a Raspberry Pi 5 would be fast enough to be able to run inference on such a big model at the speed at which it can read from the SSD.
On the other hand, there are many mini-PCs with Intel or AMD CPUs that have both a PCIe 5.0 SSD and a PCIe 4.0 SSD, which may be read in parallel, achieving thus a reading throughput of up to 20 Gbyte/s.
Such miniPCs have fast enough CPUs/GPUs, so that they might be able to reach the inference speed limited by a 20 Gbyte/s weight reading throughput, which for a so big model like GLM-5.2 would be of one output token every few seconds (only a fraction of the weights must be read for one output token). The ratio between output tokens per second and the weight reading throughput can be improved by various methods, like multi-token prediction or batching multiple tasks.
Optimizing inference speed in such conditions is an active research subject, due to the high current memory prices.
All very achievable, I have a setup with a wireless keyboard to the htpc, and a script that wakes up my desktop with wake on lan, ssh's in and starts sunshine if I start moonlight.
Booting the htpc can be a pain; personally my best solution has been wake on lan via phone. I've also used universal remotes before cec was reliable, and I had to control the screen separately.
Do you want to help stop that, or do you just want to feel smarter than other people? If you happened to have seen my comments in other threads, you'd know I'm against all this. But telling people they are idiots isn't going to win them over. This is not just a criticism of your comment, it's widespread in these discussions.
Opposing this requires:
* Linking to specific harms, which the public can emotionally resonate with. For example, scanning billions of photos for suspicions of child abuse will result in false positives that cause innocent people's kids to be taken away.
* Not seeming like an overheated conspiracy theorist. Feeling angry about this is legitimate, but it's not necessary to communicate in the same emotional register that you are feeling, even if it feels inauthentic not to. The public are saturated with people being publicly angry, much of which is purely performance. Deep concern may work better.
* Have plausible reasons for why this is happening. Yes, a few individuals like Thiel want to create a digital Stasi, but this would still be happening without them. Mostly this is driven by companies that want to make money, and officials who have a bias towards centralised processes, and are tunnel-visioned with respect to some issue. And people who are genuinely concerned about kids and haven't been given another convincing solution.
* Get facts straight. (Eg, rent/job IDs aren't a future threat. They are here)
> On one end £9 of labour cost for a plate of asparagus seems deeply inefficient and unrealistic, particularly when the cost of ingredients that also include (hard) labour is £2.
Presumably the staffing cost is the front of house staff as well as the actual cooking and then the cost of employing someone to wash dishes, clean the restraunt and so on. Then compared to growing asparagus which seems to largely come from countries with substantially lower wages. Restraunts have always been infamously low margin businesses though.
After reading this. I realize how different Asian and Western consciousness really are.
My entire technology stack was built on Microsoft's ecosystem, not on open source. This was Microsoft's attempt to expand their base for the corporate hiring market and OS market share.
Conversely, open source was a huge barrier for me. When I have a product I've built, I have to get past open source, but accessing open source comes with the barrier of English. And once you get past the English barrier, you hit the cultural barrier.
My hobby projects do integrate with open source, but all the technology that actually makes me money depends entirely on the Microsoft ecosystem. Most of the Asian developers around me are also tied to specific vendors. On the other hand, the Korean companies that do have a culture of contributing to open source are large corporations, and entry is determined by academic pedigree.
Because the entire context of open source is in English, and learning English reliably is expensive in itself. So to properly work as a developer in Korea, you actually need to be vendor dependent. The corporate ecosystem is not oppression; it is the only viable path to education and survival. If you want to grasp the latest trends, you ultimately need curation from a specific company. Some people say Hangul is a great writing system, but to me, this is where it becomes a curse and a shackle.
So when I read Hacker News, I feel just how large the gap in thinking is between the West and the East. The Japanese developers I have talked to mostly talk about coding within corporate environments rather than open source, and Chinese developers are also shaped by their corporate environments. But the posts on HN talk about their 'gardens' being ruined and absorbed by corporations, and they resist that. But since I was raised in a corporate environment from the start, I cannot imagine a different one, so this resistance tends to feel like an aristocratic hobby to me.
On the flip side, HN might see corporations as predators. Technology should be a commons, and developers should be free, not tenant farmers of a platform.
But the irony I personally feel is that to protect this 'garden commons,' they end up creating centralized, non-public coordination mechanisms with the very corporations that plunder the commons. That feels contradictory to me.
For security vulnerability response, non-public coordination may be necessary. If a vulnerability is disclosed before a patch is ready, attackers can create exploits. But the principle of open source is transparency and open discussion, while the Akrites-style security principle is non-public coordination and a single point of contact.
On top of that, corporations used open source as free infrastructure, and now that the risk has grown, they are building corporate-led governance systems based on that risk. That feels ambiguous to me. Of course, open source sponsorship has always had some tension, but if that was buying a craftsman's work, this looks more like buying the craftsman's workshop.
I wonder how Westerners would read this. I am curious. To me, this looks like a political struggle to take control of governance over the commons. Do Westerners see it as the Avengers? The difference in mindset is sometimes painful.
> but how do you deal with the fact that university, in practice, is not "for the joy of learning"?
I think this strongly depends on the subject you study, most people in mechanical engineering will have a very different attitude to people studying world literature or astrophysics.
yeah open source is cool and all but can we talk about how literally everything is written in javascript now. even your toaster probably runs on node. its an infection.
I don't see how they could ban them in the US. Code is speech, and the first amendment still mostly holds. They might try, but I don't see the courts upholding it.
True. But automation also pushes up wealth. Starbucks drivethrus aren't a thing because we need iced sweet nominally coffee stuff. It exists because we have the disposable income to pay someone else to make the coffee.
Yes a doorman is a cost, and a greater cost than previously, but we've also got more money to waste on such fripperies.
270 years sitting in a museum and ML cracks it in a few years. Makes you wonder how many other 'unreadable' artifacts are just waiting for the right model."
I think OPs point wasn't specifically about how to preserve and recover the books, but about how something unimaginable, can happen and can our sci-fi writers come up with such unimaginable now, but possible in the future plot.
I did that for someone (after jumping through QUITE some hoops) and apparently the next days some popup made the person click the upgrade button anyway.
So yeah, probably just dark pattern + non-technical user but still.
Great, but that's only a part of operational costs. A craftsman's revenue may exceed the electricity bill for the power drill, doesn't mean the business is sustainable.
Followed more than 3k experts and practioners in ai and assembled their links and conversations into an analysis of what mattered to the industry during this quarter. Interesting findings about the pushback and criticism that is also starting to appear from within the industry
That's one of the reasons why I opted for multiple citizenships. I was fortunate to have the option to rehash parts of my name, and I am using my birthplace country's passport as a throwaway for everything, and my other citizenships docs only the governments and legal system know about.
Plus, like others mentioned, I am preparing my future self to be in a total offline/unplugged world and be a semi-ghost digitally, and even now in 2026 I found a bunch opportunities already doing it in a semi-agorist way (e.g. second hand, cash on hand, offline remote physical work, hunting/fishing, farming/a little farm, etc).
Lately I’ve been finding most restaurant dishes “low quality”: in particular, less meat and tastes overcooked compared to what I make at home, though grains and vegetables are also blander.
I suspect this is more me being a harsher critic than restaurants enshittifying. I’ve been improving my cooking. I do get premium ingredients, that sometimes cost much more than the cheapest alternative, but still always much less than even low-end restaurants.
So my conclusion is, if you like good food you should cook yourself. Maybe if you’re rich enough to always eat at especially expensive restaurants, but even then I think you’d prefer a private chef.
> If anyone feels negative, I recommend seeing a psychologist.
Telling everyone that disagrees with you to see a psychologist is certainly something. The delusions of grandeur may make your statement quite ironic, heh.
I am now facing my second major wall. The first was when attaching the proposal as a PDF caused the AI to stop responding — I was able to continue by converting it to a .txt file. Of course, since I am using the free version, I am consuming significant computing resources at no cost, and the responsibility for that burden on the AI companies is entirely mine.
Due to family circumstances, I am currently unable to pay for a subscription. I feel genuinely sorry toward the AI companies, and I am deeply grateful for everything they have provided.
Now, even attaching the proposal as a .txt file causes the session to stop, or returns only surface-level responses. DeepL has also entered a usage restriction period.
My next plan is to attempt a German version inspired by Viktor Frankl, whom I deeply admire. I intend to push through this wall through trial and error.
I would love to hear your thoughts. What do you think?
Even in quantized form, reducing that size by 2 times or 4 times, a Raspberry Pi 5 could not run it otherwise than by reading the weights from an SSD.
Even thus, I do not believe that a Raspberry Pi 5 would be fast enough to be able to run inference on such a big model at the speed at which it can read from the SSD.
On the other hand, there are many mini-PCs with Intel or AMD CPUs that have both a PCIe 5.0 SSD and a PCIe 4.0 SSD, which may be read in parallel, achieving thus a reading throughput of up to 20 Gbyte/s.
Such miniPCs have fast enough CPUs/GPUs, so that they might be able to reach the inference speed limited by a 20 Gbyte/s weight reading throughput, which for a so big model like GLM-5.2 would be of one output token every few seconds (only a fraction of the weights must be read for one output token). The ratio between output tokens per second and the weight reading throughput can be improved by various methods, like multi-token prediction or batching multiple tasks.
Optimizing inference speed in such conditions is an active research subject, due to the high current memory prices.