Tuesday, October 31
Return To Blender Edition
Top Story
- Apple's M3 MacBook Pro is here. Or will be in a week. (Ars Technica)
It's about 15% faster than the M2 model.
A 14" MacBook Pro with 64GB of RAM and 4TB of SSD costs $4900.
My 14" HP Pavilion with 64GB of RAM and 4TB of SSD cost me around $900.
So... Yeah.
- The new M3 iMac has also been announced. (Ars Technica)
As well as the speed boost, the maximum memory has been increased from 16GB to 24GB.
That extra 8GB of RAM will cost you $200.
It retails for $20.
So... Yeah.
Tech News
- Why does my code run slower on a 5950X running Linux than a 5625U laptop running Windows 11 and WSL?
Probably cache latency.
The 5950X has two CPU chiplets that share cache over a high-speed interconnect, while the 5625U is a single chip. The chiplet design makes large CPUs cheaper to build but can do weird things to cache latency.
This is a real question that is happening with some code I wrote recently; it's 50% slower on what should be a much faster CPU.
Guess when I (eventually) build my new system I might want to go for the 7800X3D rather than the 7900.
- So, how do those Qualcomm benchmarks for their upcoming Snapdragon X Elite hold up under independent testing?
Just fine, actually. (AnandTech)
Ryan Smith at AnandTech got to try out two pre-production laptops under both Windows and Linux, and the benchmark scores he got matched Qualcomm's numbers and are genuinely faster than equivalent Apple M2 laptops.
Of course, Apple just announced the M3 which if everyone's claims are accurate will give them back the performance lead in Arm laptops... By 2.5%.
Looking very good for Qualcomm at this point, except for the six month wait for these systems to arrive.
- Most of the claims in an artist lawsuit against AI art companies have been dismissed for being, well, crap. (Reuters) (archive site)
The Reuters article doesn't really give you the facts of the case, the claims of the plaintiffs, the arguments of the defendants, or the relevant laws.
This Twitter thread does a much better job and is well worth reading if you're interested in this kind of thing.
Essentially, while generative AI is new, to claim that it is infringing on your copyrighted works - under current law - would require that the work it produces to be substantively the same, not merely influenced by your own work.
And that's simply not what generative AI does.
- The Biden Administration has issued its executive order governing the development and use of AI systems. (WhiteHouse.gov)
As you would expect, it's a mix of useless bullshit, impractical bullshit, unconstitutional bullshit, and just, well, bullshit.
- Speeding up Python by 17,000,000%. (Sidsite)
The article takes a real-world problem with analysing correlations in survey results (people who answered A in question 5 were most likely to answer D in question 15) and then tightens up the code until it squeaks.
While the end result is pretty hairy (though much less so than some of the code I have to maintain) the first two optimisations are straightforward and make it run 50x faster.
Posted by: Pixy Misa at
04:20 PM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 543 words, total size 5 kb.
Monday, October 30
In The Pool Edition
Top Story
- Small Hadron Collider: Scientists have for the first time fired up the world's tiniest particle accelerator. (Space)
While early accelerators were much much smaller than today's giants like the Large Hadron Collider - which is an underground ring five miles across - this new device, called a nanophotonic electron accelerator, or NEA, reverses the trend entirely.
It's the size of a dime.
The electrons it produces have a millionth of the energy of the particles in the LHC, and there are far fewer of them, so the overall beam energy is tiny.
What it does allow you to do though is produce a small and perfectly focused beam of radiation exactly where you want it. Targeting cancer? The beam is sharper than any scalpel.
While it has been proven to work, practical uses - like performing surgery without having to cut the patient open - are likely still years off.
Tech News
- Apple's M3 CPU is expected to be announced today. (WCCFtech)
It probably won't be a lot faster than the M2, but if it at least supports more memory it will be a big improvement.
With all variants of Apple's CPUs the memory is soldered onto the CPU module. You can't upgrade it, ever. So if you don't configure enough RAM in your new Mac you've just bought some very expensive e-waste. And with current the M1 iMac you can't configure enough RAM. 16GB is all you can have.
(I have 64GB in my budget HP laptop.)
- Lenovo is set to release YAPALAT - yet another perfectly adequate large Android tablet. (Notebook Check)
The Tab M11 has a definitely mid-tier but not awful CPU with two A75 and six A55 cores, a roughly 1920x1200 screen, and up to 8GB of RAM and 128GB of storage.
Make that screen anywhere from 7" to 9" rather than 11" - keeping the same resolution - and I'll buy three of them.
- Apropos of nothing, I kind of had the impression that the block of land my house is on had no two sides parallel.
Looking at the subdivision plan again after eighteen months I was reminded that this is not true at all. The land is rectangular, but the house is at an angle.
This may or may not become relevant in the near future.
Posted by: Pixy Misa at
04:23 PM
| Comments (2)
| Add Comment
| Trackbacks (Suck)
Post contains 408 words, total size 3 kb.
Sunday, October 29
Bricking It Edition
Top Story
- The Hugo Awards are here again and the Best Novel apparently did not go to a woke trash diversity pick. (Gizmodo)
I haven't read T. Kingfisher's (not her real name) Nettle & Bone, which won this year, but I have read her Paladin series and Clocktaur books, and they're fine. Not the sort of genre-changing works we expect to see winning Hugos, but enjoyable reads.
Which, sadly, might count as genre-changing these days.
Tech News
- The Lexar NM790 is a low end (though not lowest-end) PCIe 4 SSD. But what does low-end mean in late 2023? (Serve the Home)
It means that under sustained heavy write loads - the weak point for DRAMless budget drives - it slows down to as little as 1GBps.
Which is... Fine. For most tasks, that's plenty.
Sequential reads are over 7GBps, which is as fast as PCIe 4 will go.
I wouldn't recommend it for a server, but for a desktop, particularly as a second disk, it seems like it would be great.
The 4TB model sells for $189.
- The new RTX 4060 Ti model from Asus takes the unused eight PCIe lanes and adds an M.2 slot with them. (Tom's Hardware)
A prototype had two M.2 slots - after all, they use four lanes each, and eight are available. The problem is, Intel desktop CPUs can't split the lanes up like that.
So rather than make a card that would only work fully in AMD systems, they took that second slot out.
- Supposed specs of the upcoming Nvidia 40x0 Super cards. (Tom's Hardware)
The 4080 Super would be a cut-down 4090, the 4070 Ti super a cut-down 4080, and the 4070 Super would be an irrelevant overpriced piece of junk.
- Should you buy a second-hand Nvidia 170HX from a bankrupt Mongolian crypto mining farm? Probably not. (niconiconi)
These cost $5000 new two years ago, and are now on the market second hand for around $500. The 170HX is a variant of Nvidia's A100, the top of the line, which cost over $10,000.
The problem is, Nvidia lobotomised it to stop it cannibalising sales of the more expensive card. It was good for mining Ethereum, but nobody mines Ethereum anymore. Literally nobody, since there no longer is such a thing as Ethereum mining.
On other tasks it's all over the map, performing anywhere from 3060 levels to the equivalent of the 3080. Unless you know exactly what you want to use it for and have benchmarks to hand, best to avoid.
Posted by: Pixy Misa at
04:32 PM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 437 words, total size 4 kb.
Saturday, October 28
Age Of Empires Edition
Top Story
- Tech layoffs are back with a vengeance. (Tech Crunch)
Not just in the "big tech" firms that are 90% diversity hires at this point, but real companies building real products like Nokia, SiFive, and Solidigm (the name of Intel's SSD unit now that it's owned by SK Hynix).
Wait, you say, didn't Tech Crunch tell us tech layoffs were a thing of the past?
They did indeed, last month. And this article actually calls out their own previous reporting rather than hiding it under the rug:Last month, Alex wrote that tech layoffs were pretty much a thing of the past. Shouldn’t have said that, buddy, you jinxed it.
Blame Alex.
- My notebook unpooped itself. Not entirely sure what happened there, but I did get a new notebook set up last night just in case. (One I bought last year but haven't used much yet.)
Tech News
- Asus has a big new Threadripper motherboard. (Tom's Hardware)
Five PCIe slots (three of them PCIe 5), four ECC Registered memory slots, three M.2 slots, optional IPMI for remote management, and... Stuff. They don't have the full tech specs up yet since release isn't until next month.
Only having four memory slots is a bit meh; in practical terms that means you can have 256GB of RAM where you can easily get 192GB on a regular motherboard. There are probably larger modules out there but they're not easy to find.
Price TBA. Chance of me getting one: Slim.
- The Threadripper Pro is getting benchmarked and is impressive. (WCCFTech)
The 64 core and 96 core models now hold the top two positions on Passmark, 60% faster than Intel's fastest chip.
And the 24 core 7965WX outruns the 64 core 3995WX, which is... Well, probably a quirk of the benchmark, because the individual cores are not twice as fast.
- Intel's new 7529 pin CPU socket doubles as a waffle iron. (Serve the Home)
You'll want to put it into low-power mode though or you risk burning your waffles.
- Sam Bankman-Fraud says he didn't steal customer funds. (Tech Crunch)
Rather, Alameda (which he controlled) borrowed the funds from FTX (which he controlled) without authorisation by the customers and with no plan to pay them back.
Totally different!
- Almost what I want: The Pimoroni Picovision has two RP2040s: One as the CPU and one as the video chip. (Tom's Hardware)
The RP2040 is the chip used in the Raspberry Pi Pico. It has no dedicated video hardware, but the chip is so well-designed that not only can it generate video without external hardware, it can generate an encoded HDMI stream without external hardware. And the chip costs a dollar.
This device is a main board with an RP2040 and an HDMI slot, plus a connector for a regular Pi Pico, plus 16MB of external RAM because if you're doing video the Pico's internal 256k fills up pretty fast.
I'd like to see a single board with the two RP2040s and all the necessary connectors, but in the meantime this will do fine.
Posted by: Pixy Misa at
03:34 PM
| Comments (1)
| Add Comment
| Trackbacks (Suck)
Post contains 521 words, total size 5 kb.
Friday, October 27
Parks And Wrecks Edition
Top Story
- Humanity is at risk from an AI "race to the bottom". (The Guardian)
What's the risk?A handful of tech companies are jeopardising humanity’s future through unrestrained AI development and must stop their "race to the bottom", according to the scientist behind an influential letter calling for a pause in building powerful systems.
What's the risk?"We're witnessing a race to the bottom that must be stopped," Tegmark told the Guardian. "We urgently need AI safety standards, so that this transforms into a race to the top. AI promises many incredible benefits, but the reckless and unchecked development of increasingly powerful systems, with no oversight, puts our economy, our society, and our lives at risk. Regulation is critical to safe innovation, so that a handful of AI corporations don't jeopardise our shared future."
What's the risk?In a policy document published this week, 23 AI experts, including two modern "godfathers" of the technology, said governments must be allowed to halt development of exceptionally powerful models.
What's the risk?The paper, whose authors include Geoffrey Hinton and Yoshua Bengio – two winners of the ACM Turing award, the "Nobel prize for computing" – argues that powerful models must be licensed by governments and, if necessary, have their development halted.
What's the risk?The unrestrained development of artificial general intelligence, the term for a system that can carry out a wide range of tasks at or above human levels of intelligence, is a key concern among those calling for tighter regulation.
None of these companies are working on AGI. All of them are dumping huge amounts of money into glorified typeahead systems that understand nothing.
Tech News
- OpenAI is forming a team to study catastrophic risks relating to AI. (Tech Crunch)
What is the risk?... including nuclear threats.
I have a surefire mitigation for that.
If you even think of giving AI access to nuclear weapons, we will shoot you.
- Writing a simple virtual machine in less than 125 lines of C. (Andre Inc)
Which isn't bad, because my current test framework for opcode performance is double that.
But it's a really simple virtual machine. The only mathematical operation available is addition, for example.
Posted by: Pixy Misa at
05:46 PM
| Comments (4)
| Add Comment
| Trackbacks (Suck)
Post contains 363 words, total size 3 kb.
Thursday, October 26
First Rule Of Rules Edition
Top Story
- Pixy's First Rule of Rare Firmware Bugs That Take out All Your CPUs One by One: They only happen between 2AM and 4AM.
- Twitter competitor Pebble is shutting down on Wednesday. (ZDNet)
It had 20,000 users.
Which used to be a lot.
In the early 70s.
Tech News
- The Biden Administration is set to unveil a "sweeping" executive order on AI next week. (Washington Post) (archive site)
As an executive order, it can only affect the federal government, and the problem I see is that the order is unlikely to go far enough in curbing government use of AI bullshit.
- In which academics reap what they have sown. (The Verge)
The story tries to pin the blame on conservatives, even though nobody in the story is to the right of Mao. A typical AWFL college student goes off the deep end and files Title IX complaints - and this is her mistake - against every lecturer she ever had contact with.
No. Single out the weakest animal from the pack and take it down. If you charge in they're going to see you coming and stomp you.
- Boeing has now lost more than $2 billion - which used to be a lot - on two new Air Force jets. (CNN)
As in, two planes, not two new jet fighter programs.
Air Force One A and One B - I guess - were supposed to total $3.9 billion. Since the contract was signed during the Trump Administration, it's fixed-price, not cost-plus, so Boeing gets to eat the loss.
Which of course means that they pass that loss on to customers, rather than the government passing the overrun on to taxpayers in the more traditional way.
- Where ae all the laid-off workers from Big Tech going? (Dev Interrupted)
They're writing RPG IV for Mutual of Omaha.
I mean, not all of them, but there were a lot of boring companies that actually do stuff that were looking for programmers, and if you move from California to Nebraska you can take a hell of a pay cut and still have more left over at the end of the month.
- Team's Cardea Z540 is a 12GBps PCIe 5 SSD. (Tom's Hardware)
While PCIe 5 is in the "I still don't need it" category, it's only about twice the price of PCIe 3 storage and more than three times as fast. It might even make a noticeable difference in some things.
Maybe.
Posted by: Pixy Misa at
05:41 PM
| Comments (2)
| Add Comment
| Trackbacks (Suck)
Post contains 434 words, total size 4 kb.
Wednesday, October 25
Breathing Space Edition
Top Story
- Qualcomm has shown off its new Snapdragon X Elite - an Arm CPU for laptops that might not suck. (AnandTech)
To be specific, an Arm CPU for laptops that might not suck and doesn't come from Apple.
The benchmarks on offer aren't very detailed, but they do show it beating Apple's M2 in single and multi-threaded workloads - the single-threaded win being significant because previously Apple had the fastest Arm core around.
It also beats Intel's 13800H on multi-threaded tests while using much less power, and beats AMD's integrated graphics by 80% - and AMD's integrated graphics are pretty good.
Products are expected "mid-2024".
- Apple is announcing its M3 chips next week, which is a bit of a spoiler. But they'll need to increase single-threaded performance by around 12% to match Qualcomm.
Tech News
- A look at the Beelink SER7, a smaller, cheaper version of Beelink's GTR7. (Serve the Home)
That is, a mini-PC with a Ryzen 7840HS. All the bits are where they should be, though the custom power connector is a definite minus. (It can also be powered via USB-C, so it's not fatal.)
- Okta's little security incident has wiped $2 billion of its market cap since Friday. (CNBC)
Okta has not been having a good year. If you put all your eggs in one basket, that had better be a pretty robust basket.
- Oregon State University warned students and staff to avoid all robots. (NBC)
Good advice generally, though this was related to a bomb threat - that turned out to be a "prank".
The sort of prank that is likely to get you bed and breakfast at a federal facility for an extended duration.
- Microsoft truly seems to believe that the reason people don't love them is because they're just not annoying enough. (Neowin)
They now try to force you to take a quiz before you can install Chrome on a new Windows system. Not sure if that applies to other browsers like Brave or Vivaldi.
Curiously, the response Just fuck off already is not among the listed options.
Posted by: Pixy Misa at
05:21 PM
| Comments (2)
| Add Comment
| Trackbacks (Suck)
Post contains 362 words, total size 3 kb.
Tuesday, October 24
Poisoning Poisoners Edition
Top Story
- Researchers at the University of Chicago have come up with two filters - named Glaze and Nightshade - that prevent AI systems from learning from human art for at least, oh, a day. (MIT)
The filters change the images in ways invisible to the human eye that trick the AI into producing incongruous results.
The problem with this should be obvious: If the filter is invisible to the human eye, you can run the infected filter through an second filter than removes invisible nonsense and makes the image safe to plagiarise again.
It's copy protection, and it doesn't work.
Tech News
- Graphics card maker Colorful is preparing a single-slot version of the 4060 Ti. (WCCFTech)
This is welcome, given that most graphics cards - even low-end ones - now take up three PCIe slots. In terms of width, that is; they only plug in to one slot on the motherboard.
- Apple could be releasing a refreshed iMac next week. Or not. (9to5Mac)
The current iMac looks great, is reasonably fast, and has a brilliant screen, but is limited to just 16GB of RAM. If they improve on that, great. If not, it will be DOA again as far as I'm concerned.
- 1Password had a "security incident" related to the recent Okta breach. (Bleeping Computer)
Which is exactly what you'd expect to happen when a password manager outsources its customer support login management to a third party.
However, apart from that 1Password seems to be run by people who know what they are doing; the hackers were detected before they could do anything, and it is impossible - at least in theory - for them to have accessed password data anyway. since 1Password doesn't have your password.
- COBOL on Wheelchair is a framework for writing web applications in COBOL. (GitHub)
I mean, sure, why not.
Posted by: Pixy Misa at
05:28 PM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 317 words, total size 3 kb.
Monday, October 23
Mushroom Mushroom Edition
Top Story
- Meta (Facebook) has revealed the details of its internal serverless platform - similar to Amazon's Lambda - that "processes trillions of function calls per day on more than 100,000 servers spread across tens of datacenter regions." (Engineer Codex)
Does anyone see the problem with this? Anyone? Yes, the girl at the back?
Correct. That's fucking stupid.
Tech News
- Nvidia might be bringing out a 4080 Super model sooner rather than later now that the 4090 has been banned from export to China. (Tom's Hardware)
Hang on. Wait a minute. Where exactly are the 4090s - the boards not the chips - being manufactured?
- Intel's real 14th generation chips have shown up in benchmark databases and the results are good, bad, and irrelevant. (Tom's Hardware)
In single-core tests it's 6% faster than the 13th gen equivalent, which is... Meh.
In multi-core tests it's 25% slower, despite having two extra cores. Which is terrible, but is certain to be fixed before launch date so it doesn't matter.
Looks like an unexciting release on the CPU side; any interesting news will come from the integrated graphics. Or not at all.
- Everything new is old again: Ruffle is a Flash emulator written in Rust - and compiled to Web Assembly. (Ruffle)
Which means you can add it to a web page and it will Just Work (TM).
Except that it doesn't have full support for all Flash functionality yet, so while Ruffle will work your Flash game might not.
Still if you happen to have 160,000 Flash games just lying around it offers a new way to share them.
Posted by: Pixy Misa at
05:35 PM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 288 words, total size 3 kb.
Sunday, October 22
Future So Loud Gotta Wear Shades Edition
Top Story
- Thanks to AI the future of programming may involve yelling in ALL CAPS. (Ars Tchnica)
Or more specifically, involve "AIs" (which aren't) yelling at each other in ALL CAPS.
In this case, ChatGPT (a chatbot) yelling at DALL-E (an image generator).
And we know this because DALL-E is stupid (though good at producing images) and repeated the instructions ChatGPT yelled at it. It's an idiot savant to ChatGPT's idiot.
Tech News
- Scaling to 15,000 functions and beyond. (OpenFAAS)
15,000 functions is a few megabytes of code. Sure, this would have been tricky on a CP/M system; even with a hard disk you'd have to be careful optimising your overlays so you weren't spending all your time swapping code in and out. For it to be a problem in 2023 has to mean you're doing something incredibly stupid.
So, first thing: We're not talking about regular functions in a piece of code here; we're talking about "serverless" functions running in the cloud. These do have their place, though they are horribly inefficient.
Second thing:I started off by looking to hardware that I already owned. My workstation runs Linux and has an AMD Ryzen 9 5950x with 16C/32T with 128GB of RAM. Then, behind me sits the Ampere Developer Platform with 64C and 64 RAM. I paid an additional 500 USD to upgrade the Ampere machine to 128GB RAM in order to recreate the customer issue.
It took me a moment to unpack this. Every single one of those functions needs to run on its own container - a lightweight virtual server. 15,000 functions means 15,000 virtual servers. That's insanity.
The container limit of 110 per Kubernetes node means that even if you have a bare-metal machine like this, it’s largely wasted, unless you are running a few very large Pods.
The rest of the article discusses the struggles to get 15,000 virtual servers deployed, which is only interesting if you enjoy watching train wreck videos.
So we have:
1. Serverless functions, which are useful in a limited role.
2. Some crazy people who want 15,000 separate serverless functions, which is insane.
3. Some crazy people who deploy every single serverless function as a separate virtual server.
4. The poor guy who has to make all that shit work.
Every company has some of this nonsense going on: "Yes, I know we did this to ourselves, but we have to make it work somehow." But this example is truly spectacular.
- Unison is a language that is supposed to make this problem go away. (Unison)
You define your functions, deploy them somewhere, and then leave it up to Unison to manage where the functions are running and send the calls to the appropriate servers.
Sounds good.
The website is terrible, though; half of it is reminders to join the slack channel.
Also, this:tour/main> find : [a] -> [a]
When a user asks for a list of the functions that take and return a sequence value, you should not sort deprecated functions to the top.
1. lib.base.data.deprecated.Heap.sortDescending : [a] -> [a]
- Why could reviewers not run Geekbench on the Google Pixel 8? Because Google blocked it. (Notebook Check)
And why did Google do that? One has to assume, because the scores weren't very good.
And checking some reviews, that seems to be the case. The Mediatek chip I mentioned yesterday is significantly faster than Google's powerhouse - 40% single core and nearly 100% multi-core.
- The CEO of Hashicorp predicts that Silicon Valley will abandon open source unless open source stops, uh, being open source. (The Stack)
If you didn't want to give your software away, mate, maybe you shouldn't have given your software away. It's not my fault that you're an idiot.
Disclaimer: It's not, is it? I didn't flood the world with idiot juice by accident at some point... Though if I had, that would explain a lot.
Posted by: Pixy Misa at
03:23 PM
| Comments (5)
| Add Comment
| Trackbacks (Suck)
Post contains 655 words, total size 5 kb.
60 queries taking 0.162 seconds, 379 records returned.
Powered by Minx 1.1.6c-pink.