• Renesas (a Japanese company formed by the merger of the semiconductor operations of Hitachi and Mitsubishi) has announced multiplexers for third-generation MRDIMMs.  (WCCFTech)
  
  You'd be entirely justified in asking Wut? at this point, but this is a great technology landing at precisely the wrong time.
  
  MRDIMMs are registered memory taken to the next level.  Since the memory modules have interface chips - the registers - on them, you can have multiple banks of memory on each module, allowing 256GB DIMMs instead of just 64GB.
  
  What the M part stands for is multiplexed, meaning that when you access the memory it interleaves data from two banks at once, so that transfers are twice as fast.
  
  The new multiplexers run at 16GHz.
  
  Which means you get user upgradable high-capacity memory modules that run at the sort of speed normally associated with the memory on video cards.
  
  Which would be amazing and allow for some seriously powerful integrated graphics except that memory is so expensive right now that there is absolutely no point to it.

• LLMs are broken and can't ever be fixed.  (GitHub)
  
  We've touched on this before.  Computers have separate areas in memory reserved for instructions and data respectively, with strict controls over which is which.  When those controls break down, you get a common class of security vulnerabilities - buffer overflows and the like.
  
  LLMs don't have that.  Worse, they pass benchmarks running tests for that problem with flying colours - typically blocking 100% of attacks - but when faced with a real human attacker they fail 100% of the time.
  
  
  
• What we know of Nvidia's 5000 Super series.  (WCCFTech)
  
  Supposedly they are already in production but have been hidden away because nobody knows how to price them.
  
  
  
• Also, I bought an RX 6400 today.  The least powerful recent AMD graphics card and by a good margin, but also the least power hungry, and physically the smallest.  It's not much slower than the RX 580 (I have a couple of those) but is a tiny low profile card that only uses 50W of power.
  
  Also a lot cheaper and smaller than, say, an RTX 3050, while not being that much slower.  Not something you'd buy unless you need to fit the card in a low-profile slot, but the Hyte / Hololive limited edition cases (of which I now have five) only allow for a single full-height card, so I do.
  
  
  
• 2x, not 10x.  (OBryant)
  
  Sounds about right.

• The United States government just banned Roombas.  (The Verge)
  
  Well, it didn't, and wouldn't have at all if it hadn't previously blocked the acquisition of US company iRobot by US company Amazon leaving the market to Chinese competitors.  And they're only being blocked over well-founded fears of being used for surveillance just like every other thing in our lives these days with more functional purpose than a Chia Pet.
  
  If you like your Roomba you can keep your Roomba, and existing models of Roomba can continue to be imported, but new models of Roomba - and of any other imported robot - will need to be certified.

• Huawei's new MateBook Pro S weighs just 798 grams, which is vanishingly light for a 14" model.  (Notebook Check)
  
  Only problem is that it runs an in-house Kirin XE90 Arm CPU, which has decent multi-threaded performance thanks to its 10 dual-thread cores, but pretty lousy single-threaded performance.  Expect performance like a 2017-era laptop in that regard.
  
  
  
• Speaking of which, I just realised that the new Minisforum NAB9S has a 13900HK CPU rather than a 13900HX.  So while it is a current model that takes DDR4 SODIMMs (of which I have an abundance) and is a fair bit faster than my existing hardware that uses that memory, it is also a fair bit slower than the full HX model.
  
  It's also relatively inexpensive because it's available as a barebones model, and has a 2.5" bay for a SATA SSD if I should happen to find where mine have gone.
  
  
  
• How to run Gemma 26B on an 8GB Apple Silicon system.  (GitHub)
  
  Gemma 26B requires 52GB of VRAM to run as it ships with 16-bit values.  The process squimshes it down to 4-bit quantisation and then caches the core (using about 2GB of RAM in total) and streams the rest from SSD.
  
  Which is great because while you can certainly run Gemma 26B-A4B (4-bit quantised) on a 16GB graphics card without hacks, this sort of trick would allow you to run some 200B parameter models on the same hardware - the scale where AI models stop being stupid and useless.
  
  
  
• And if you have an Acemagic F9A, 1.6T parameters.  (WCCFTech)
  
  You'd want a fast SSD for that, though, because that would be a lot of data to fling about.
  
  
  

Musical Interlude

• So how are things going a year into the DRAM Apocalypse?  (Tom's Hardware)
  
  I'm glad you asked.
  
  Actually, no I'm not.
  
  But anyway, you remember when DDR5 memory was brand new and people preferred DDR4 because it was much cheaper?
  
  Yeah.  DDR5 costs twice as much now as it did back then when it was too expensive to be worthwhile.
  
  
  

Tech News

• Speaking of DDR4, how do Intel CPUs fare when using it?  (Tom's Hardware)
  
  Okay-ish.  The tricky part is when you see benchmark numbers for older AMD chips like the 5800X3D you know they are using DDR4, because that's all they support.  Intel's 12th, 13th, and 14th generation chips all support both DDR4 and DDR5, and using DDR4 can lop off 14% from published benchmarks.
  
  
  
• If you send Claud Opus the character sequence - - - it writes a novella about an unnamed architect.  (Austin's Nerdy Things)
  
  Sadly only via the API.  I tried it in the web interface and it just said "Huh?"
  
  
  
• AMD's new Radeon 9050 graphics card is half a 9060.  (WCCTech)
  
  It comes in 8GB and 4GB flavours, each with half the compute cores of the 9060, and the 4GB model having half the memory bandwidth as well.
  
  
  

• The best time to buy an Nvidia graphics card is never. The second best time is also never. (WCCFTech)
  
  Nvidia has just raised card prices by another 30%. A minor inconvenience for me, as I do have four perfectly adequate graphics cards and two kind of crappy ones already. If you were waiting for prices to come down, well, keep on waiting I guess.

• Anthropic CEO Dario Amodei says he doesn't support dropping space colonies on competing Chinese AI firms but if a colony were somehow to drop would that really be such a bad thing? (Tech Crunch)
  
  I mean, really?
  
  
  
• Speaking of Anthropic, shocking news from Claude where it turns out that sharing a chat session means... You share your chat session. (Notebook Check)
  
  And it ends up in Google.
  
  
  
• ChatGPT now refuses to write in the style of famous authors, even mostly dead ones like Charles Dickens. (Ars Technica)
  
  Since even college English majors these days struggle to understand Dickens, maybe shoot a little lower.
  
  
  
• Who will be the last man standing when the cookie crumbles, and why will it be Apple? (MacRumors)
  
  Because while Apple's Vision Pro headset was a resounding flop, it was at least a product you could buy. Make it smaller and lighter and cheaper and it might not fail so hard next time. And the entire project cost less than some AI companies are burning in a week.

• Whenever you copy this weird white text on a white background, a little white light lights up white to let you know you've done it. (Futurism)
  
  If only it were so simple.
  
  If you're copying and pasting your exam question into an AI to cheat, at least read the answer it spits out.
  
  Also, don't ever do that. Copy into a neutral environment like Notepad (or Notepad++) precisely so that you don't inadvertently include nasty Unicode stuff, or, in this case, an instruction to the AI to make a nonsensical reference to an island off East Africa in your final exam paper.
  
  
  

Tech News

• Apple is facing the worst possible future. (WCCFTech)
  
  By which I mean having to pay as much for RAM as everyone else.
  
  
  
• Minecraft system requirements have gone up. (Tom's Hardware)
  
  Apparently my work laptop now falls below not only the recommended specs, but the minimum specs. Yes, it's a bit laggy when running with ~200 mods installed, but it's perfectly fine unmodded.
  
  
  
• TSMC's 2nm production is in full swing. (WCCFTech)
  
  Not ahead of time either, given that it can take six months for a completed chip to roll off the production line.
  
  
  
• Research: Cosmic ray spallation, tufa versus travertine flowstone, positrons are evil, muon tomography, and the proper attitude for a cat tasked with editing fanfiction.

Noodling

Musical Interlude

Posted by: normal at Tuesday, July 28 2026 12:09 AM (LADmw)

Hide Comments | Add Comment

• China is breeding a clone army of terror yaks, probably.  (CNN)
  
  They say they're not, but they also say:

  These are Tibetan yaks, designed and cloned in secretive labs more than 4,000 meters above sea level, where researchers hope to create an "elite" herd of super-yaks, healthier and more fertile than their predecessors.

  I do not trust them.

• Lenovo's new ThinkCentre X system comes with up to 256GB of RAM.  (Notebook Check)
  
  If you expect that it would also come with a price tag previously reserved to professional workstations with top-of-the-line CPUs and graphics cards, you would be correct.
  
  
  
• So here's a 28.9 million parameter small LLM running on a $8 microcontroller.  (GitHub)
  
  It'll do.
  
  
  
• The US has accused a citizen re-entering the country of giving the CBP a "duress" password that wiped his phone.  (Tech Crunch)
  
  Do not do this.  Whatever you think about border agents scrolling through your sick memes, anything you don't want on your phone you should simply not have on your phone when you cross the border.
  
  
  
• Research: How to prevent runaway salinisation in an Australian irrigation megaproject (secret: add salt), there's more than one way to force an ice age, and why there is a shoe shop in Bali named Wallace.

• Anthropic's Opus 5 model is here - nearly as smart as Fable for half the price, and less annoying.  (Engadget)
  
  It has been trained similarly but with much less effort towards security, so it isn't as clever at hacking systems or at preventing your systems from being hacked, but it hasn't been forced into I'M SORRY DAVE I'M AFRAID I CAN'T DO THAT insanity that flips on the moment you stray into anything even the tiniest bit questionable.
  
  I'll give it a try since I'm on the $20 Claude plan making Opus 5 free of additional charge while Fable 5 uses my limited credits.  (Paying accounts received $100 worth of credits free, but I want to keep those for actual security scans.)
  
  Also worth noting: If Fable 5 runs into trouble or doubt, it doesn't just run in circles, scream and shout, it can silently drop back to Opus.  This can happen even if your question is entirely orthogonal to security issues.

• Astronomers have discovered the first THAT'S NO MOON, THAT'S A CATEGORISATION ERROR outside of our solar system.  (Space)
  
  That is, it's a solid body orbiting another celestial body orbiting a star - clearly a moon - except that the body in the middle is a brown dwarf, a "failed star".
  
  
  
• Zen 7 Epyc "Florence" will succeed the just announced Zen 6 Epyc "Venice" family in 2028 - using the current sockets.  (Tom's Hardware)
  
  Wait.  Yes, it says right on the AMD presentation that Zen 7 uses the same server socket as Zen 6, but...  Oh.  The other slide doesn't explicitly say DDR6 - which would require a socket change - but a vaguer "next gen memory".
  
  
  
• The Geekom A9 Max is a Ryzen 470 mini PC with one stick of RAM because they don't want to force you to throw away a perfectly good dual-channel kit in order to upgrade.  (Tom's Hardware)
  
  Which gives us a perfect chance to see how much difference single-channel memory makes to performance, because there's a whole list of benchmark results from similar AMD systems and competing Intel ones.
  
  And the answer is, it makes almost no difference to CPU performance, but cuts the performance of the integrated graphics nearly in half.
  
  
  
• If coding has been solved, why does software keep getting worse?  (Piotr)
  
  Because everybody is using AI tools to push changes faster, with the same level of care as before to finding and fixing bugs.
  
  
  
• Marimo notebooks can now be embedded in PyCharm.  (Marimo)
  
  Bringing modern development environments the interactivity we 29 year olds enjoyed in the prior millennium.
  
  
  
• SpaceX's latest test launch was mostly successful, with the only failure being with the booster's engines not fully restarting for the test splashdown.  (Tech Crunch)
  
  Which turned into a splashboom.
  
  No such issues with the Starship V3 second stage which bobbed happily undamaged in the ocean after its own splashdown.
  
  
  
• Lenovo has a new all-in-one desktop PC with an interesting 2560x2880 27" display and an Intel Panther Lake CPU with up to and including Arc B390 integrated graphics.  (Notebook Check)
  
  It is fucking expensive.
  
  
  
• Research: Today I ran a Shinkansen line through Eromanga.  I didn't plan it that way; that was the inevitable outcome of the parameters I had set.

Posted by: Mauser at Sunday, July 26 2026 04:37 PM (XWgGM)

Hide Comments | Add Comment

• AMD has announced its new Zen 6 Epyc server CPUs and we know, if anything, less about them than we did before the announcement.  (WCCFTech)
  
  Here's the problem: Zen 6 is strongly rumoured to increase the core count per chiplet from 8 to 12, and increase the L3 cache and X3D cache per chiplet by 50% to match.  Which means that the raw numbers of cores and cache do nothing to confirm or deny the 12 core rumour.  It doesn't help that the announced mainstream products focused on 96 core models which would have the same numbers whether that's twelve 8-core chiplets or eight 12-core chiplets.
  
  Until you get all the way down to the bottom of the chart for Socket SP8 processors and find the 8-core Epyc with 48MB of cache which rather hints that 12 core chiplets are not a lie.
  
  Oh, also supported (officially) are PCIe Gen 6 - probably for servers only, and 12.8GHz MRDIMMs - also probably for servers only.  And up to 16 memory channels per socket, which at 12.8GHz delivers close to three times the bandwidth of Epyc Zen 5.

• Ryzen 7 has been confirmed on track for 2028, bringing with it support for DDR6 memory and the new ACE instruction set for matrix operations.  (WCCFTech)
  
  There is no spoon.  As far as I can see some bugger has nicked the entire cutlery draw.
  
  
  
• There's also the "new" X199 embedded CPU which is AMD's Ryzen AI Max+ 395 with a much more convenient name and a power budget of 55W.  (Tom's Hardware)
  
  It's competing with Intel's Panther Lake laptops chips and is unsurprisingly faster, although I'm not sure it's possible to manufacture competitively.
  
  
  
• Patreon has laid off 20% of its workforce.  (Tech Crunch)
  
  CEO Jack Conte blamed AI and not his own malicious incompetence.

Musical Interlude

• Across the social sciences, half of research doesn't replicate.  (Science)
  
  And only 34% of it is confirmed by multiple reviewers.
  
  But it gets worse.
  
  In political science and economics replicability rises to more than 60%.
  
  For education, the number is...  Zero.

• Apple is launching a new deal with Klarna where you can rent an iPad.  (ComputerWorld)
  
  You'll own nothing and we really don't care whether you like it.
  
  
  
• Code in the iOS 27 beta suggests that Apple will be able to switch your rented iPad off remotely if you don't pay your bill.  (9to5Mac)
  
  Because of course.
  
  
  
• Samsung's new Z Fold 8 is here.  (CNet)
  
  You can get a pretty decent phone for the price of a high-end gaming PC.
  
  
  
• The Linux kernel team just published 432 security issues in two days.  (The Register)
  
  Asdfghjkl.
  
  
  
• The Hisense A10 is an e-ink phone with a detachable LCD screen.  (Liliputing)
  
  It clips onto the back of the phone with magnets.

Posted by: Rick C at Friday, July 24 2026 12:58 AM (Hi0NU)

Posted by: Joe Redfield at Friday, July 24 2026 02:47 AM (KOtXO)

Posted by: normal at Saturday, July 25 2026 06:53 AM (WDfTs)

Hide Comments | Add Comment

• LLM company OpenAI says that LLM archive site Hugging Face was breached by hackers using OpenAI's LLMs to gain access to models hosted at Hugging Face in order to breach...  I dunno.  (Tech Crunch)
  
  Seems legit.
  
  
  

Tech News

• EU courts have ruled that if your country is invaded by VPNs that's your own problem.  (Tech Radar)
  
  This decision actually seems to be sane.
  
  
  
• How to link two Dell Pro Max GB10 AI systems together with the built-in 200Gb Ethernet to make a single system.  (Tom's Hardware)
  
  Step One: Plug them in.
  
  
  
• China is considering banning Chinese AI companies from manufacturing chips at Taiwan's TSMC.  (Tom's Hardware)
  
  And also from reversing entropy.
  
  
  
• AI companies are buying up old books - physical books printed before 2022 - because they are free of AI slop.  (404 Media)
  
  It's a real problem, because if you feed AI slop into a slop machine things go from bad to worse, fast.