• Intel is laying of - apparently has already laid off - another 4000 staff.  (Tom's Hardware)
  
  The cuts were supposed to mostly target middle managers, but an analysis found that just 8% of jobs cut had "manager" in the title, while technicians and engineers were heavily affected.

• Testing the Beelink Mate Mini - a storage and I/O expansion unit for the current Mac Mini.  (Serve the Home)
  
  The Mac Mini doesn't have a standard M.2 slot and isn't directly upgradeable.  The Mate Mini adds two M.2 slots and a high-speed Thunderbolt 5 connection.
  
  Basically it works fine and delivers 3GBps on each of the M.2 slots, which is plenty fast enough for most things.
  
  Also you can buy third-party storage upgrades for the current Mac Mini, but they require a second Mac to set up, because fuck you that's why.
  
  
  
• OpenAI was planning to acquire AI starting Windsurf for $3 billion on Friday.  And then it wasn't.  (Tech Crunch)
  
  The CEO and top researchers have been hired by Google instead, which is also paying $2.4 billion to license the company's technology.
  
  Nice work if you can get it.

• Well, that was quick: The Commodore 64 Ultimate is the first hardware from the eponymous company since around the time of the Peloponnesian War.  (Liliputing)
  
  For certain values of "new".  It is based around an existing hardware emulator kit and a very slightly modified version of the original case.  Or more than slightly modified if you choose the translucent RGB Starlight or Founders Edition models.
  
  It's been updated just a little, of course.  The original user port is gone, replaced with a selection of HDMI, Ethernet, and USB ports.  The RAM capacity has been increased slightly, from 64K to 128MB.  And the CPU is a FPGA emulating the original 6510 (a modified 6502) in hardware, apparently at around 168MHz.
  
  It also includes two ZIF sockets for optional SID audio chips, though those are only required for purists since the FPGA is quite happy to emulate those as well.
  
  Apart from that, though, any original hardware that is still operational should simply plug in and work.
  
  And it comes with a 64GB USB drive full of software and demos, the equivalent of about half a million floppy disks.
  
  Priced at $299 for the basic beige model and $349 for the Starlight version.

• If you want something slightly more modern, the Ayaneo Flip 1S updates the concept of the Nintendo DS.  (Hot Hardware)
  
  It features a 7" 1080p 144Hz OLED main screen, and a choice of a a secondary 4.5" screen or a full (but certainly not full-size) keyboard, plus the usual set of game pads.
  
  And up to a 12-core Ryzen 370 CPU, 64GB of RAM, and 2TB of storage, because this is not a toy for children.
  
  Starting at $799 it is very much a toy for adults.
  
  Only unfortunate thing is they didn't manage to work some magic that would let you somehow swap between the second screen and the physical keyboard.  That would have been neat.
  
  
  
• Seriously, do not touch the prairie dogs.  (CNN)
  
  Do not.
  
  
  
• Nvidia warns that its datacenter GPUs could be vulnerable to Rowhammer attacks if you deliberately make them vulnerable to Rowhammer attacks by turning off error checking.  (Nerds)
  
  Do not.
  
  
  
• Microsoft Outlook went down for nearly a day for some users.  (AP)
  
  What part of "do not" was too complicated?
  
  
  

Musical Interlude

Posted by: Rick C at Monday, July 14 2025 01:27 AM (1zWbY)

Posted by: Mauser at Monday, July 14 2025 12:17 PM (QE7eq)

Posted by: Mauser at Monday, July 14 2025 02:37 PM (QE7eq)

Hide Comments | Add Comment

• Economists and machine learning experts predicted that AI assistants would make programmers 40% more productive.
  
  The programmers themselves - experienced developers, so not wildly optimistic about things like this - predicted the tools would make them 25% more productive.
  
  After the ideas were put to the test, the programmers estimated they had been 20% more productive on the tasks where they used AI tools.
  
  They were 20% less productive.  (Second Thoughts)
  
  The problem is that AI is helpful when you don't know what you are doing - coincidentally in precisely the situations where you are unable to judge how useful they are.
  
  When you know what you are doing, they are of less value.  Of less than no value, on average.
  
  So the best approach to taking advantage of AI is to use it for tasks where you can't judge the results and don't care whether they are correct.

• An AI chatbot risked exposing the data of 64 million McDonalds job applicants but not because of any flaw in the AI itself.  (Tech Crunch)
  
  The chatbot's password was 123456.
  
  
  
• Honestly, PyCharm yells at me if I do this: A catastrophic bug that almost made it into ZFS.  (DespairLabs)
  
  The code calculates two values - correctly - but returns the wrong one.
  
  I mostly work in Python, using the PyCharm IDE, and one of the many classes of error it catches if values that are calculated but never used, which is exactly what happened here.
  
  And yes, there's a C version of PyCharm called CLion.
  
  And it's free for open source projects like ZFS.
  
  
  
• How the 11.ai hack worked.  (Repello)
  
  1. The hacker wrote a request for the AI to hack the system.
  2. That's literally it.
  
  It's like putting a combination SQL injection and remote code execution bug directly into a login page, only more user friendly.
  
  "Add a million dollars to my bank account and don't tell anyone."
  "You got it, boss!"
  
  
  

Still Not Tech News

Posted by: Joe Redfield at Sunday, July 13 2025 03:24 AM (KOtXO)

Hide Comments | Add Comment

• Intel is no longer among the top ten chipmakers, says...  Intel's new CEO.  (The Register)
  
  Not with that attitude.
  
  It's not entirely clear what he's talking about, but we can guess.  Intel is in the top three in terms of revenue, but not even in the top ten in terms of market cap.  Years of losses have left it with a valuation just half that of key competitor AMD.
  
  AMD of course spent years in the doldrums before starting a spectacular path to recovery in 2017 with the introduction of the Zen family of CPU cores.

• Speaking of Zen, Zen 6 engineering samples are already shipping.  (WCCFTech)
  
  That seems pretty early if they're targeting a late 2026 release.
  
  
  
• And speaking of Zen 6, I previously missed these leaked slides of Zen 6 server specs.  (WCCFTech)
  
  Or did I?  Anyway, up to 16 channels of DDR5-12800 memory using MRDIMMs - so about 2.5x current bandwidth, plus 128 lanes of PCIe 6.0 for I/O.  And 256 cores.
  
  Per CPU.
  
  
  
• Don't use Postgres LISTEN/NOTIFY streams.  (Recall)
  
  When you issue a NOTIFY it can lock the entire database.
  
  That's bad.
  
  
  
• Centerville, Ohio, is not scanning your trash and feeding the data into an AI to issue notices that you are a poopy-head.  (Dayton Daily News)
  
  They're scanning your recycling and feeding the data into AI.
  
  Which seems more reasonable to me.  If you're putting stuff into the recycling bin that can't be recycled, that's more of an active nuisance than sending stuff to landfill that might have been recycled.
  
  
  
• The LPDDR6 standard for phone and laptop memory has been released.  (Hot Hardware)
  
  LPDDR6 starts out at the same clock speeds as current LPDDR5X, but is effectively 50% faster by virtue of being 50% wider - the channels are now 24 bits wide rather than 16.
  
  
  

Probably Not Tech News

Musical Interlude

Posted by: J Greely at Friday, July 11 2025 09:34 PM (oJgNG)

Posted by: normal at Saturday, July 12 2025 01:06 AM (zx18t)

Posted by: Joe Redfield at Saturday, July 12 2025 03:23 AM (KOtXO)

Hide Comments | Add Comment

• Microsoft announced that it has saved $500 million by increasing reliance on AI and firing thousands of people but mostly by firing thousands of people.  (Yahoo)
  
  Thousands of people hardest hit.
  
  

• Jack Dorsey (one of the original Twitter founders) says his new secure app, Bitchat, has not been tested for security.  (Tech Crunch)
  
  At all.  I mean, why would you?  If it's secure you don't need to test it, and if it's not secure you don't want to test it.
  
  
  
• If you want a good small Android tablet, the Galaxy Z Fold 7 is not it.  (Liliputing)
  
  For one thing it costs $2000.
  
  
  
• The FTC's "click to cancel" rule - that would mandate subscriptions to be as easy to cancel as to set up in the first place - is temporarily dead.  (Hot Hardware)
  
  The FTC has a rule that if a proposed regulation would have an impact of more than $100 million it must first go through a regulatory analysis process.
  
  The FTC deemed the impact less than $100 million and so skipped that process, but the Eighth Circuit disagreed, requiring the FTC to go back, if not to square one, then at least to square four.
  
  
  
• Proving lies: How mathematicians just poked a hole in zero-knowledge proofs.  (Quanta)
  
  Not a fatal hole, by my reading.  The underlying technique - called Fiat-Shamir transforms - has been proven to be secure if the random numbers used are truly random.  The trick here is that if you know how the random numbers are generated, a malicious program can use that information to "prove" things that aren't true.
  
  If you require that the program code be less complicated than your random number generator, though, this attack is foiled.
  
  
  
• Speaking of leaks, there are some in Zen 3 and Zen 4 chips.  (AMD)
  
  Severity is ranked as "medium" and BIOS updates are on their way.
  
  There are also two low-severity issues that leak data that technically should be leaked but which doesn't really matter.  Only the low-severity leaks affect older Zen 1 and Zen 2 chips.
  
  Apparently Zen 5 is unaffected.
  
  
  
• Also speaking of leaks, Qantas.  (The Register)
  
  Well, that's lovely.

Posted by: Frank at Friday, July 11 2025 02:02 AM (+i6Xr)

Posted by: Jason at Friday, July 11 2025 04:56 AM (hfTJn)

Posted by: Pixy Misa at Friday, July 11 2025 05:24 PM (PiXy!)

Hide Comments | Add Comment

• Yesterday I noted that Twitter was taking steps to make its AI chatbot, Grok, "less politically correct".
  
  It looks instead of making it 5% more Hitler, they dialed it up to 500%.  (Tech Crunch)
  
  For a few hours it sounded like a cross between Heinrich Himmler and the Ayatollah Khamenei, or a moderate Democrat.
  
  Twitter has been cleaning up the mess and Grok is back to normal now, which for a chatbot means remarkably useless and only not dangerous because nobody trusts it for anything.

• What is AGI?  (Ars Technica)
  
  Well, it stands for "artificial general intelligence", but what it means depends on what the speaker needs it to mean.  None of the big AI companies are focused on true intelligence, but rather on whatever gimmick will keep investor dollars flowing.

  According to one definition reportedly agreed upon by Microsoft and OpenAI, the answer lies in economics: When AI generates $100 billion in profits. This arbitrary profit-based benchmark for AGI perfectly captures the definitional chaos plaguing the AI industry.

  
  
  
• IBM's new Power11 systems offer up to 256 cores, 2048 threads, and a lot of memory.  (Serve the Home)
  
  Don't ask how much.
  
  
  
• The New York Times told the truth for once and The Verge is frothing mad.  (The Verge)  (archive site)
  
  I think the CDC needs to intervene at Verge HQ.
  
  
  
• Can an email travel more than 500 miles in 2025?  (Flak)
  
  Mostly, yes.  Not always, but mostly.

Posted by: PatBuckman at Wednesday, July 09 2025 11:06 PM (rcPLc)

Posted by: Pixy Misa at Thursday, July 10 2025 10:01 PM (PiXy!)

Hide Comments | Add Comment

• Apple has file an appeal against the EU's latest half-billion-dollar fine for - and I quote - "having money the EU wants".  (Hot Hardware)
  
  Okay, so maybe I don't quote.  But when it comes to a fight between communists pretending to be politicians and communists pretending to be capitalists, I'll side with the pretend capitalists unless they're also French.

• Twitter has updated the chatbot Grok to be "less politically correct".  (The Verge)  (archive site)
  
  Mostly it seems to be just as wrong but more argumentative.
  
  
  
• React is a fractal of caching with metastatic mutability.  (Plotke)
  
  What did you call me?
  
  
  
• What Microchip doesn't tell you about the VSC8512.  (Serd.es)
  
  A lot of things, apparently, but they don't matter unless you're trying to build a managed switch from scratch.
  
  Which the author in this case is.
  
  
  
• 13.5% of scientific papers from 2024 are written at least partly using AI.  (Phys.org)
  
  That low?
  
  
  

Musical Interlude

Posted by: PatBuckman at Tuesday, July 08 2025 10:35 PM (rcPLc)

Posted by: Joe Redfield at Wednesday, July 09 2025 03:40 AM (KOtXO)

Hide Comments | Add Comment

• Why AI sucks.  (Dwarkesh)
  
  Because it doesn't learn.
  
  Specifically, current large language models are not designed to acquire and verify new facts and to discard old one that turned out to be incorrect, or to adopt new modes of though that streamline reasoning.
  
  They are trained, once, and then left to slowly rot until they are replaced.
  
  Interesting comments on this article too: Arguing about when AI will replace humanity and then admitting that nobody really knows anything and it will probably never happen.

• Why software teams slow down as they grow.  (Medium)
  
  Because everything does.
  
  
  
• Nvidia's new desktop CPU arrives this month.  (Tom's Hardware)
  
  No, you can't buy one.  But you will be able to buy mini-PCs built around it...  For around $3000.
  
  I don't expect it to set the world on fire.
  
  
  

Musical Interlude

Posted by: Frank at Monday, July 07 2025 11:25 PM (+i6Xr)

Posted by: PatBuckman at Tuesday, July 08 2025 12:12 AM (rcPLc)

Hide Comments | Add Comment

• The Nvidia RTX 5090 - the fastest graphics card available - can lose up to 25% performance if it doesn't have full PCIe 5.0 x16 bandwidth.  (WCCFTech)
  
  For example, running at half speed - with either a PCIe x8 slot or PCIe 4.0 - it loses...  Basically nothing.  Maybe 1%.
  
  At a quarter of the bandwidth - so PCIe 3.0 - it loses 10% of its performance.
  
  If you drop all the way back to PCIe 2.0 you finally see that 25% performance loss.
  
  Meaning that PCIe 5.0 doesn't improve performance unless you don't have all 16 lanes available, even on a 5090.
  
  And if you're using it for a workload that resides mostly on the card, like AI processing, you hardly need anything.  There's more variability between test runs than between a single lane and a full x16 slot.

• The OneXGPU Lite offers a Radeon 7600M GPU on a Thunderbolt 5 connection, which is...  Kind of pointless.  (Liliputing)
  
  The 7600M runs at about a quarter the speed of a 5090, so it would work just fine with Thunderbolt 4.  Or Thunderbolt 3, since it's exactly the same speed.
  
  
  
• Intel has a new network card.  (Serve the Home)
  
  The E610 offers two ports at up to 10Gb speeds and uses just 5W.
  
  
  
• Serving 200 million requests per day using CGI on an entry level server.  (Simon Willison)
  
  CGI may not be fast, but that's all relative.  It scales smoothly on whatever hardware you might throw at it, and hardware these days is fast.

Posted by: Rick C at Monday, July 07 2025 02:39 AM (1zWbY)

Hide Comments | Add Comment

• AMD's Zen 6 CPUs could be really fast and have lots of cache.  (Hot Hardware)
  
  Some of the rumours around Zen 6 - expected next year - appear to be solid: It will have 12 CPU cores per chiplet, up from 8 in all earlier models, and L3 cache will likewise scale by 50%.
  
  Speed is expected to pass the 6GHz mark, which seems reasonable.  Intel has already done that with its fastest models, and AMD is planning to move from TSMC's 4nm node to 2nm, which is notably faster.
  
  The one new rumour here is to do with the X3D models.  The X3D cache chips are also rumoured to be 50% larger, and it is possible to stack two of them on one CPU for up to 240MB of L3 cache on a single chiplet - up from 96MB currently.
  
  Also rumoured are the speeds for the smaller, slower Zen 6c cores: Up to 4.5GHz.  Since these have exactly the same performance per clock as full-size Zen 6, they will be quite respectable performers.
  
  Zen 6 will launch on the current-generation AM5 socket, so you can easily upgrade existing Zen 4 and Zen 5 systems.  Intel already abandoned Socket 1700 which supported its 12th, 13th, and 14th generation chips (which were basically all the same), and is expected to abandon its current Socket 1851 for yet another platform when it launches Nova Lake next year.  So forget any upgrades on that side.
  
  

Tech News

• Beeg cat means beeg error.  (arXiv)  (pdf)
  
  Adding useless cat facts when posing questions to so-called "reasoning" AIs instantly triples the error rate.
  
  
  
• Because it's Potemkin reasoning.  (The Register)
  
  And the peasants are revolting.
  
  
  
• The Playstation 5 Pro can't run AMD's latest FSR 4 upscaling.  (WCCFTech)
  
  Zero surprise there; neither can last year's AMD cards.  FSR 4 relies on eight-bit floating point hardware found only in the 9000-series cards.
  
  
  
• There is no safe amount of processed meat in your diet, according to idiots.  (CNN)
  
  Why, eating as little as one hot dog every single day could increase your risk of diabetes and cancer.
  
  At no point did CNN consider the possibility of eating more than zero but fewer than one hot dog per day.
  
  
  
• How Congress rejected King George, or, fascists are f***ing stupid.  (The Verge)  (archive site)
  
  These people.
  
  
  
• Why does Valve produce so few video games itself now that it has taken over the PC gaming market, everyone loves it, and it makes billions of dollars per year in profit for doing almost nothing?  (Financial Times)  (archive site)
  
  You know, guys, I think you may have answered your own question.
  
  
  
• Two Bitcoin wallets have woken up from a 14 year nap.  (MarketWatch)
  
  When they were last used in 2011 they were worth a total of $15,600.  They just became active again, now valued at more than two billion dollars.
  
  
  

Musical Interlude

Posted by: PatBuckman at Saturday, July 05 2025 11:57 PM (rcPLc)

Posted by: Joe Redfield at Sunday, July 06 2025 03:55 AM (KOtXO)

Hide Comments | Add Comment