Oh, lighten up
I'm sure they, being Eastern European, can take a joke.
C.
3533 publicly visible posts • joined 21 Sep 2011
Yeah, normally we try not to let infosec companies hype up problems they conveniently have the solution to, and we sat on this article for a few days thinking about whether it was worth doing.
But we figured it would be worth bringing io_uring to a wider attention, it appears to be a contentious feature, and this might be a nail in the coffin for it.
We publish a lot of work; some of it is less Earth shattering and more just whatever we felt people would find mildly interesting.
C.
Well, heh, we did say it can be switched off, and if you can find the command in one Google search, one wonders if it's worth including here.
We also don't like including commands we haven't tested and are not confident about. But seeing as it is pretty straight forward, and easy to run, we've now included it in the article.
C.
They're just names of benchmarks - you're not expected to know what they stand for, just that they are individual tests.
"A Massive Multi-discipline Multimodal Understanding" doesn't tell you a whole lot more about it; it's just a benchmark called MMMU.
It's not like we're using something like LoRA and not spelling it out. What I have done is link through to the things, which gives you more idea of what they do.
C.
That makes us sad because we try to make things clear. Sophgo is central to the story, so it's not noise.
Here are the main claims: TSMC was given a chip design to fabricate (TSMC makes chips for others, it doesn't design them) by a Chinese company, Sophgo.
TSMC thought that chip looked a lot like one of Huawei's and that Huawei had given its design to another company to give to TSMC to manufacture, with the resulting chips being funneled back to Huawei. US sanctions prevent TSMC from making chips for Huawei.
TSMC let the US know about it. In response, the US wants to penalize TSMC.
Does that make sense? I've revisited the article to clear it up best I can. If TSMC being punished seems confusing, well, that's the situation.
Also if something doesn't make sense, feel free to email corrections@ with your feedback so we can jump on it right away.
C.
The translation is correct - we don't just use machine translation, we have Mandarin speakers on the team.
It's claimed to have INT8 performance of 8 TOPS, that's eight trillion 8-bit operations a second. As GJC said, INT8 is good enough for AI inference and chips can scream through 8-bit ops efficiently.
C.
Nah, we're not that organized or cynical. It's as simple as this: I wrote boffins as the headline at the office. I got home, went to the gym, had dinner, relaxed a little, and looked at the site to make sure things were ticking over OK.
And it just felt on reflection that teachers and professors would work better there, because we already have a revenge of the nerds reference, and teachers (and a lot of professors) aren't boffins. Plus it's educators v the White House. That seemed more striking to me than wrapping it all up in boffins.
I wouldn't read into tweaks and editorial choices too much; we're regular joes who are lucky enough to get paid to poke fun at the tech world, get people to lighten up, and explain stuff along the way. We make decisions and change our minds on gut feeling a lot.
C.
Hi - it's pretty simple. We love the word boffins and we use it every week, but sometimes we over-use it and it loses its impact. We try to reserve it to specifically to scientists, particularly those in traditional sciences such as physics and chemistry as well as strong computer science.
Also this is a story that ought to appeal to a wider tech-interested audience, and hitting them with our boffin slang right out of the gate along with a Revenge of the Nerds reference seemed like it would be off-putting for some for such an important issue.
We bite the hand that feeds us, but we don't hack our own wings off either. Sometimes we have to balance humor with audience accessibility in order to maintain or grow our reach, influence and impact. That means headline tweaks from time to time on reflection. This is online, it's not print. It's also all a long battle, not short skirmishes.
More analogies as they come.
C.
It's because you're looking for a list-like structure, I guess, but list is just our flippant way of describing it. It's (as the article states) a set of so-called specification decisions, which is typical obfuscated Euro-jargon. Hence why media is calling it a list.
I've added a direct link to the specs but we did summarize them in the piece, and the linked announcements also point to them.
C.
It can help explain how to solve those sorts of equations; it can also suggest formula to use for a given situation. It can teach you trigonometry if you trust its output, for instance.
But if it can't even do 420*69 or something simple like that by itself, I'm not expecting it to come up with answers to differential equations. That said, it can walk you through the steps of solving these sorts of equations by yourself as it's ingested so many textbooks on the subject, etc.
LLMs are by their very nature lossy information retrieval and prediction, not mathematical packages.
C.
First, I always feel bad when I pipe a script like that into my shell, but I'm never quite sure if I'm right to feel that way. What's the alternative, I download the tar ball and run make install without even looking through it? Am I really going to read through all the build and installation code?
If you're the kind of person who does read it all, then yes, we provide a link to the manual installation instructions. If you're the kind to git clone && cd project && make all then curl | sh isn't very far off from that.
I would say, work to your own threat model. I wish the QubesOS approach was the normal these days, in that a lot of what you do is confined to individual VMs that can be thrown away.
Second, these instructions aren't aimed at newbies. They're for intermediate or higher users who want to get started with LLMs locally.
C.
Everyone knows Rupert's eleventy billion years old*. In fact, he just celebrated his birthday on March 11. Many happy returns, chief.
C.
* 94.
(I confess, I've worked for Viscount Rothermere. I've worked for Sly Bailey. But I've never worked for Murdoch, though I was quoted in the final NOTW as a friend was doing news desk shifts for them and needed to fill some space, and I had a story in The Sun ages ago about university security guards gawking at students through their bedroom windows using well-placed CCTV cameras.)
AIUI, if the intermediate CA expires, the chain of trust is broken. Software -- standard cryptography libraries used by today's apps -- looks up the chain and sees the dead signing CA and rejects the signed device cert.
Per-device key-pair certificate --> signed by intermediate CA --> signed by a trusted root CA.
If the intermediate CA expires, the chain's toast, the device cert is invalid, and the device isn't trusted by the client app. It's not that the cryptography suddenly magically stops working mathematically, it's a decision by the client apps (specifically the libraries they use) to reject certificates that have an invalid chain of trust.
Hence why one of the workarounds is to push out apps patched to overlook the certificate expiry or hard code an acceptance of the last-known good certs.
C.
The issue, from the point of view of Suhy, is that v3.5 was released under a license that is basically the AGPLv3 with the CC tacked on. The AGPLv3, it's argued, says you can remove added restrictions, so that's what Suhy did. Took off the CC, leaving the AGPLv3, giving him, in his mind, the right to distribute a fork and make money from it.
Neo4j disagrees.
C.
Calendar years and fiscal years aren't necessarily aligned; Nvidia's isn't aligned to the calendar.
As we said, "the fourth quarter of its 2025 fiscal year, the 12 months to January 26." Nvidia's fiscal 2025 runs from January to January a year head of the calendar year. So it's already in fiscal 2026.
C.
Ah, we're kinda damned if we do, damned if we don't here - when we do (say) video interviews, we're asked for full transcripts; and when we provide full transcripts, it's a pain for some people.
We've done our best to make it readable but sometimes, when some of our vultures have a lot of other work on, we have to ship what we've got - and in this case, we shipped this as a full-length interview. There is a summary at the start if you can't read it all.
Also if there is anything confusing in it, feel free to drop us a note to corrections@theregister.com and we'll fix it up best we can. Cheers,
C.
The files you link to are GPLv2 and/or BSD licensed. They are open source.
Your complaint really is that hardware has to be specifically programmed to work, whether that's with magic values in IO registers or microcode, but that actually means you're upset that the designs of the electronics, whether it be a network card or a microprocessor, aren't open - fine, take that up with the hardware designers.
The kernel code that operates the hardware is open source. It's open source. I'm done here.
C.
We're not saying they are necessarily connected. If they were connected, we would have said so. We're saying that around the time the Linux kernel project was opening the door to Rust, the Azure CTO was remarking how good the language is for security and reliability.
From that you can infer Rust was at the time getting more recognition and support.
C.
I like Rust. But I like Zig more because Zig just lets me code without Rust's obstinance, and with some safety, but not as much as Rust's.
Rust's compile-time safety is absolutely non-magical thinking. Have you encountered the borrow checker? No doubt there are ways to trick your way past it, or just turn it off with unsafe{}. But by and large, it's going to stop you doing silly things at build time - and it cannot be bargained with.
The underlying hardware is irrelevant IMHO. The masochism is very relevant.
C.
I'm not sure how rewriting a C kernel in a C-like language with strict compile-time safety would make it suck. It might suck in that porting/converting millions of lines of kernel code is going to take some time, and in that time, the new kernel will be incomplete.
But like low-level C, Rust compiles down into GC-less machine code. It might even run faster in places as some run-time safety checks/mechanisms can be performed at build time.
And I suppose we should just call Rust what it is. Ada fan fiction.
C.
It's a bit nuanced but we've added that as a break-out box. USAID was being probed by its own watchdog (its IG) regarding...
"how (1) the Government of Ukraine used the USAID-provided Starlink terminals, and (2) USAID monitored the Government of Ukraine’s use of USAID-provided Starlink terminals."
To me that speaks more to USAID's management of the situation, but the conflict of interest is great enough in any case that IMHO Elon should have stayed out of the agency's pending demise.
C.
I don't know how on Earth you could get that conclusion from our coverage. We're just pointing out the history of the thing. No, they're not paying us. I don't know why people go for the bribes claims right away.
Also, I'll make the team aware of the other point above that Fujitsu owned the majority of ICL before its takeover; we'll mention that going forwards.
C.