• The Sixth Circuit has issued an administrative stay on the FCC's latest Net Neutrality nonsense pending a decision on a permanent stay.  (The Verge)
  
  As I've been saying for at least a decade, the FCC doesn't have and cannot be permitted the authority to enact net neutrality without new legislation.
  
  Fortunately the recent Chevron decision cut the FCC's legs out from under it in this regard, if it didn't amputate them entirely.  Even The Verge admits that the FCC is likely to lose this one.
  
  As usual, the comments over there are full of foaming illiterates, but there are only a few this time.
  
  
  

Tech News

• The FBI has gained access to the would-be assassin's cell phone.  (The Verge)
  
  Unlikely that this will change anything, of course.
  
  Best comment: Reddit is leaking.
  
  
  
• Microsoft's CTO denies the obvious, that AI is facing exponential scaling costs.  (Ars Technica)
  
  He denies this in the face of exponential expenses in AI training.  He's lying, badly.
  
  Best comment: Always listen to the Chief Tulip Officer's advice about investing in tulips.
  
  
  
• A look at AMD's Zen 5 microarchitecture.  (AnandTech)
  
  Zen 1 through Zen 4 had the same basic design, able to issue and retire four instructions per cycle.  Zen 4 also introduced a 256-bit half-width version of Intel's AVX-512 vector processing.  Intel's own consumer CPUs lack AVX-512 in any form.
  
  With Zen 5 the issue width has been increased to eight instructions per cycle, and the AVX-512 unit is now a full 512 bits wide.
  
  If your code is poised perfectly to take advantage of the improved hardware it could run twice as fast on Zen 5 as Zen 4, but that's unlikely.
  
  The performance charts attached to this article show the 12 core 9900X running Handbrake video processing tasks 41% faster than Intel's 24 core 14900K, at half the power consumption.
  
  Of course Intel will have new chips itself later this year, but those are expected to focus on fixing the power issues more than increasing performance.
  
  These chips are due to show up...  Basically now.
  
  
  
• Also on the AMD front, testing the graphics performance of the new Ryzen HX 370 laptop chip.  (Tom's Hardware)
  
  It's basically level with the desktop GTX 1070, laptop GTX 1650 Ti, or the Radeon RX 480.  I used an RX 480 myself from 2017 to 2022, and it's a perfectly competent card.  For integrated graphics performance it's amazing.
  
  AMD has another, much more powerful laptop CPU in the wings, with at least twice the graphics performance.  No word yet on when that one will ship.

Posted by: normal at Tuesday, July 16 2024 10:53 PM (LADmw)

Posted by: PatBuckman at Wednesday, July 17 2024 12:51 AM (rcPLc)

Posted by: Rick C at Wednesday, July 17 2024 11:22 AM (MItL9)

Posted by: Mauser at Wednesday, July 17 2024 03:08 PM (nk1Z+)

Hide Comments | Add Comment

• To avoid sea level rise, some researchers want to put barriers around the world's most vulnerable glaciers to slow them down.  (Science)
  
  Uh.
  
  What?
  
  In 2008, the Ilulissat Glacier in Greenland had a calving event in which it shed a single iceberg covering three square miles.  It sheds 35 billion tons of icebergs in the average year.
  
  And moving glaciers don't leave much of anything in their wake, except rubble.
  
  I mean...  Okay, it's not impossible.  If you want to build an anti-glacier barrier, go right ahead.  Yes, I'll make popcorn, but if you succeed I'll gladly give you credit.
  
  Update: Skip the first four paragraphs of the article and go to the fifth, which explains things a lot more clearly.  They don't want to build an anti-glacier barrier, but a barrier for warm ocean currents to shield glaciers at the point they enter the sea.

• The new Asus ExpertBook lacks the Four Essential Keys and its Lunar Lake CPU will get crushed by AMD's new Strix Point (non-Halo) chips when they arrive later this month.  (WCCFTech)
  
  It's not actively terrible but I wouldn't bother.
  
  
  
• Once a new security exploit is made public, system administrators around the world have 22 minutes to update all their servers.  (Bleeping Computers)
  
  That's how fast hackers pounce on new exploits.
  
  
  
• Speaking of exploits "superhuman" AI Go players turn out to be so dumb that a child can beat them.  (Ars Technica)
  
  If the child knows the right exploit.  Even after researchers updated the AI to handle this, it only won 22% of the time.

  Improving these kinds of "worst case" scenarios is key to avoiding embarrassing mistakes when rolling an AI system out to the public. But this new research shows that determined "adversaries" can often discover new holes in an AI algorithm's performance much more quickly and easily than that algorithm can evolve to fix those problems.

  As with AI chat bots that are easily tricked into selling $50,000 cars for $5, the companies behind them will just blame you.
  
  
  
• Possibly as part of the ongoing Snowflake debacle, barcodes for around ten million event tickets from Ticketmaster have been leaked.  (HackRead)
  
  In theory the barcodes cannot be refreshed.
  
  In practice, anything can be refreshed if the alternative is a billion dollars going up in smoke.
  
  
  
• Google is reportedly in talks to acquire cloud security company Wiz for $23 billion.  (Tech Crunch)
  
  I have never heard of Wiz and have no idea what they actually do.  $23 billion should be a lot of money, so presumably they do something.

Posted by: Rick C at Monday, July 15 2024 11:45 PM (MItL9)

Posted by: normal at Tuesday, July 16 2024 12:08 AM (LADmw)

Posted by: normal at Tuesday, July 16 2024 12:12 AM (LADmw)

Posted by: Pixy Misa at Tuesday, July 16 2024 09:05 AM (PiXy!)

Posted by: Rick C at Tuesday, July 16 2024 12:02 PM (MItL9)

Hide Comments | Add Comment

• Meet the soft robots that can amputate limbs and fuse with other robots.  (Tech Crunch)
  
  I for one welcome our Mengelean Borg overlords.

  In one demo video, we see a soft quadruped robot crawling along when a falling rock traps a back leg. The reversible joint attaching the leg is heated with current, allowing the robot to break free of its leg and escape. Although it’s not shown in the video, the limb can be re-attached, as well.

  Oh.  They can amputate their own limbs.  That's still slightly horrifying, but much less so.
  
  Kind of like a lizard that drops its tail, uses it as a club to beat a frog to death, and the reattaches the tail and eats the frog.
  
  

   

• The goggles do nothing!  Pushing the 9950X to 320W doesn't really help much.  (WCCFTech)
  
  At 120W it's essentially tied with the 253W Intel 14900K, and at any power level above that it rapidly breaks away.
  
  Interestingly the upcoming sixteen core laptop chip codenamed Strix Point Halo is rated at 120W, so that will deliver what is currently considered leading-edge desktop performance.
  
  
  
• Speaking of the 14900K, I wouldn't.  (Tech Radar)
  
  The problems with Intel's high-end chips persist.
  
  A study of crash logs for video games showed that for some types of problem, 90% of all recorded errors happened on the 13900K and 14900K.  Those chips are just pushed too hard and run too hot.
  
  This is not an Intel problem in general;  it only seems to affect the company's top-of-the-line consumer desktop chips.  But that's bad.  If you buy their most expensive chips you don't expect to get dramatically worse reliability.
  
  
  
• Disney's internal Slack was allegedly hacked, with 1.1TB of data exfiltrated.  (HackRead)
  
  I haven't encountered this site before, but it looks the Web3 is going great of online services.
  
  Bookmarked for later reference.
  
  
  
• Signal is downplaying a security flaw that it has now fixed after a Twitter drama.  (Bleeping Computer)
  
  This was a local security issue, rather than an online one.  Anyone with access to your computer could read your Signal history without needing your password.  The local database was encrypted, but the encryption key was sitting right there for anyone to use.
  
  
  
• Linksys Velop routers send your WiFi passwords to servers in the US.  (StackDiary)
  
  In plain text.
  
  Yay.
  
  
  
• Speaking of which, I just picked up a mesh wifi system for the new house - a D-Link M32 Eagle Pro three node setup.  I've been running a router I had before the move, but the coverage is not that great.
  
  I probably went a little overboard, though.  The bundle was 75% off on a one-day sale so I bought two, which should be enough to cover the house, the garage, the entire garden, and a couple of my neighbours.
  
  
  
• The NSA has rediscovered the tape of computing pioneer Admiral Grace Hopper's 1982 lecture, but says it lacks the ability to read it.  (MuckRock)
  
  It's one inch Ampex reel-to-reel video tape, which is hardly uncommon.  It took me two minutes to find a commercial service that will convert an entire tape to digital format for $150.
  
  
  
• The Sipeed Lichi Pi 3A is a development board for RISC-V programmers.  (Liliputing)
  
  RISC-V itself is an open-standard competitor to Arm.  Anyone can build their own processors with no licensing fees, which makes it attractive to companies like Western Digital, who need to embed multiple processor cores into ever SSD they make.
  
  The board itself starts at $49 for a bare module with 4GB of RAM, going up to $139 with 16GB of RAM, 32GB of storage, and a carrier board with two M.2 slots.
  
  This is slightly confusing because the Lichee Pi 4A already exists and is more powerful, just lacking those two M.2 slots.
  
  
  
• Journalists for Censorship strikes again: Shooting conspiracies trend on X as Musk endorses Trump.  (The Verge)
  
  Following the assassination attempt on Donald Trump at his rally in Pennsylvania, the left did what the left does best: Act like crazy people in public.

  On X, neither trending topic about the shooting is flush with particularly robust or coherent conspiracies; clicking through, you’ll largely find short posts from X users saying that the shooting looks fake or is a stunt. (There is no evidence of either.) But by placing the subjects into X’s trending topics area, the conspiracies are elevated to more people.

  The complaint here seems to be that Twitter is not censoring Democrats.
  
  I can see how this would be bad for the Democrats.
  
  
  

Posted by: PatBuckman at Monday, July 15 2024 02:06 AM (rcPLc)

Posted by: PatBuckman at Monday, July 15 2024 02:07 AM (rcPLc)

Posted by: Pixy Misa at Monday, July 15 2024 12:02 PM (PiXy!)

Posted by: bravokilo at Monday, July 15 2024 05:04 PM (BuMbr)

Hide Comments | Add Comment

• That's going to leave a mark: Call and text records for the past six months for almost all AT&T customers have been stolen in the slow-motion train wreck that is the Snowflake breach.  (Tech Crunch)
  
  Snowflake was an online data analytics platform with some major customers including Ticketmaster and Santander Bank.  Technically I think Snowflake is still operating but I wouldn't expect them to be around for much longer.
  
  While Ticketmaster is international and the number of customers affected far outweighs AT&T, in that breach the hackers got your name, address, last four digits of your credit card, that kind of stuff.
  
  In this hack they got the phone numbers of everyone 110 million people called or texted over a period of six months.  And it's all out there, forever.
  
  It does not (according to AT&T) include the contents of the text messages, and calls even if recorded for whatever reason would not be stored in this kind of database.  I hope.  That would raise it from a mere disaster to a catastrophe.
  
  And if you're not an AT&T customer you might still be affected if your phone company uses the AT&T network behind the scenes.

• Meta has dropped the special restrictions it had placed on Donald Trump's Facebook and Instagram accounts.  (The Verge)
  
  You can enjoy the screaming in the comments there, if you like.
  
  
  
• Emmanuel Goldstein's X faces big EU fines as paid checkmarks are ruled deceptive.  (Ars Technica)
  
  Sorry, I mean Elon Musk.  It's pure coincidence that every Ars Technica article mentioning SpaceX, Twitter, Tesla, or Starlink has the name ELON MUSK in it as a signal for the zombies to swarm.
  
  Anyway, the headline is a lie.  There is no such ruling.
  
  But EU Bookburner General Terry Britain claims that blue checkmarks used to denote trustworthy sources of information which is either a direct lie or a sign of galloping early-onset Alzheimer's.
  
  Musk fired back asserting that the EU offered Twitter an illegal secret deal under which they would hold off on fines if the company would enact secret censorship under their direction.
  
  Terry Britain claimed that the EU doesn't do that, but...  It does.
  
  Also interesting to see on that article, Ars Technica's creative director Aurich Lawson getting savaged by his own mob.
  
  
  
• Don't run your Ryzen 9 9950X at 60W.  (WCCFTech)
  
  More leaked benchmarks, but there's something interesting here because the same tests were run at power levels of 60W, 90W, 120W, 160W, and 230W, with the last of those being the maximum boost power setting without manual overclocking.
  
  At 60W the CPU ran cool at 41C and managed 4GHz on all cores.  Bumping it up to 90W pushed the temperature to 49C but the clock to just over 5GHz.  At 160W it reaches 5.55GHz and at 230W 5.6GHz, so there's no real point in going over 160W.
  
  
  
• Emmanuel Gold - sorry.  SpaceX's Falcon 9 has been grounded after issues with a second stage caused the latest cluster of Starlink satellites to miss their orbital target.  (Reuters)
  
  They are trying to boost the satellites into the correct orbit with the on-board ion drives - really - but even if that works it will reduce their service life.
  
  This comes after, if you weren't keeping track, 365 consecutive Falcon 9 launches without issue.
  
  
  
• Looking to buy a Z80 computer before supply of the chip runs out forever?  Tindie (who) and Zeal have you covered.  (Tindie)
  
  For $180 you get a 10MHz Z80, 256k of flash storage, 512k of RAM, VGA graphics, and four voice sound synthesis.
  
  It's kind of neat if you're into retrocomputing.
  
  
  
• Speaking of retrocomputing the German navy is working on phasing out eight inch floppy disks.  (Ars Technica)
  
  I have no idea why they are used because the ships involved were commissioned in the mid-nineties, by which time even 3.5" floppies had been around for a decade, and eight inch models had been dead for years.
  
  
  
• A "red team" from CISA broke into another federal agency and had free rein in its network, undetected, for five months.  (The Register)
  
  Wait, the federal government noticed a massive problem in only five months?
  
  
  
• Peer review is essential for science.  Unfortunately it's fucked.  (Ars Technica)
  
  Well, they use the term broken, but I felt it needed a little more oomph.

  The practice of peer review was developed in a different era, when the arguments and analysis that led to a paper’s conclusion could be succinctly summarized within the paper itself. Want to know how the author arrived at that conclusion? The derivation would be right there. It was relatively easy to judge the "wrongness" of an article because you could follow the document from beginning to end, from start to finish, and have all the information you needed to evaluate it right there at your fingerprints.
  
  That's now largely impossible with the modern scientific enterprise so reliant on computers.

  Specifically because scientific papers very rarely include the code used to analyse the data.  All the assumptions - and the errors - in that code are hidden from reviewers.
  
  Noting that peer review is f - broken, the author, not a journalist but real life astrophysicist Paul Sutter, moves on to science's second line of defense, replication, where things are no better:

  There's a complex swirling septic tank of problems that all contribute to the replication crisis, but the first issue is that replication isn't sexy. You don't get to learn new things about the world around us; you just get to confirm whether someone else learned new things about the world around us. As an added ugly bonus, non-results often don't even get published. Novelty is seen as a virtue, and if you run an experiment and it doesn’t provide a positive result, journals are less likely to be interested in your manuscript. Additionally, because replication isn't seen as sexy, when it is done, it isn't read. Replication studies do not get published in high-impact-factor journals, and authors of replication studies do not get as many citations for their work. This means that their h-index is lower, which lowers their chances of getting grants and promotions.

  What's the solution?  I'm glad you asked:

  Altogether, the scientific and technical publishing industry rakes in about ten billion dollars a year, often with double-digit profit margins. Of course they don't want this ship to change course. I can't really blame them; they’re just playing by the accepted rules of their own game.

  Yes, there isn't one.
  
  
  
• Despite selling literally dozens of units, several of which were not returned, developers are ignoring Apple's wildly expensive and basically useless Vision Pro AR headset.  (Ars Technica)
  
  Not even porting existing iOS apps.
  
  
  
• OpenAI is reportedly working on an AI that can reason, instead of just regurgitating pre-digested Reddit posts and telling you that backpacks are as safe as parachutes when jumping out of a plane.  (Ars Technica)
  
  Well, first up, Terry Winograd, SHRDLU, 1968.

  Person: Had you touched any pyramid before you put the green one on the little cube?
  Computer: YES, THE GREEN ONE.
  Person: When did you pick it up?
  Computer: WHILE I WAS STACKING UP THE RED CUBE, A LARGE RED BLOCK AND A LARGE GREEN CUBE.
  Person: Why?
  Computer: TO GET RID OF IT.
  Person: Why did you do that?
  Computer: TO CLEAN OFF THE RED CUBE.
  Person: Why did you clear off that cube?
  Computer: TO PUT IT ON A LARGE GREEN CUBE.
  Person: Why did you do that?
  Computer: TO STACK UP THE RED CUBE , A LARGE RED BLOCK AND A LARGE GREEN CUBE.
  Person: Why did you do that?
  Computer: BECAUSE YOU ASKED ME TO.

  Second, that parachutes and backpacks thing is really an answer given by Google's AI Overview.  (Fast Company)
  
  And it's based on a real result from a real scientific study.  (NPR)
  
  But it's a study of whether people actually read scientific papers.
  
  The plane was parked on the ground.

Posted by: normal at Saturday, July 13 2024 08:11 PM (bg2DR)

Posted by: normal at Saturday, July 13 2024 08:13 PM (bg2DR)

Posted by: J Greely at Saturday, July 13 2024 09:52 PM (oJgNG)

Posted by: Pixy Misa at Saturday, July 13 2024 11:06 PM (PiXy!)

Posted by: PatBuckman at Sunday, July 14 2024 01:44 AM (rcPLc)

Posted by: PatBuckman at Sunday, July 14 2024 01:54 AM (rcPLc)

Posted by: Rick C at Sunday, July 14 2024 07:48 AM (MItL9)

Posted by: Rick C at Sunday, July 14 2024 07:51 AM (MItL9)

Posted by: Rick C at Sunday, July 14 2024 11:34 AM (MItL9)

Hide Comments | Add Comment

• Could AIs become conscious? Right now, we have no way to tell.  (Ars Technica)
  
  Because we haven't properly defined the terms.
  
  What kind of AI?  SHRDLU from 1968 (yes, I keep mentioning that) displayed more evidence of consciousness than ChatGPT, and examining its abilities we see clear indications of a primitive level of conscious.
  
  There's a term in AI, sphexishness, referring to the golden digger wasp, Sphex ichneumoneus.  Like many wasps, the sphex wasp will paralyse its prey to serve as food for its larvae.  In this case it will paralyse an insect, drag it back to its burrow (it's a digger wasp), check that the burrow is safe, and then put the prey in the burrow to serve as a larder for the baby sphexes.
  
  The thing is, if you move the paralysed insect while the sphex is checking the burrow, it will move it back next to the burrow, inspect the burrow again, and then drag it into the burrow.
  
  And if you move the insect again while the sphex is re-checking the burrow, the cycle will repeat.  No matter how many times you do this, the sphex will not change its behaviour.
  
  SHRDLU was able to explain why it did what it did, step by step.  It didn't get angry at being given repeated nonsensical orders, but it could account for every action that it took.
  
  Anyway, can computers become conscious?  Well, they are already more conscious than insects, and have been for decades.
  
  Can computers become conscious with all the complexity as human consciousness?  Certainly not currently; they are not remotely powerful enough.
  
  Can LLMs become conscious?  A single LLM, no.  LLMs are designed specifically to avoid that.
  
  A pair of LLMs in a feedback loop?  Maybe, yes.  But likely also psychotic.

• Why The Atlantic signed that deal with OpenAI.  (The Verge)
  
  Money.
  
  
  
• Ben Faw, the harbinger of AI slop.  (The Verge)
  
  He made his mark pushing ads for his buddies' businesses unnoticed into online newspapers and magazines, and it was all downhill from there.
  
  
  
• A data breach at mSpy exposed the personal information of millions of customers.  (Tech Crunch)
  
  mSpy sells spyware.
  
  
  
• TSMC is another big beneficiary of the AI bubble, with its market share topping $1 trillion for the first time.  (Reuters)
  
  While still well behind the leading US bubblers like Apple, Microsoft, and Nvidia, it is nonetheless the most valuable company in Asia.
  
  
  
• The most important video game you've never heard of.  (A Critical Hit)
  
  The Sumerian Game.  It ran on the IBM 7090 mainframe ($2.9 million for 100 kiloFLOPs) and it was, well:

  Sir, I am sorry to report that 150 bushels of grain have rotted or been eaten by rats this past season.

  Hammurabi.  It's Hammurabi.
  
  
  
• The man who destroyed Wikipedia.  (Tracing Woodgrains)
  
  David Gerard, the man behind longtime lolcow dumpster fire RationalWiki, has also been systematically rotting away the credibility of Wikipedia for twenty years.
  
  Everything bad you've seen from terminally online leftists, he represents, and he's been at it for longer than most.  He has his good points - he hates blockchain, for example, but he hates it from a doctrinaire liberal perspective, not because of the specific and often astoundingly ill-considered technical decisions that make working with it an unending misery.

Posted by: normal at Friday, July 12 2024 07:06 PM (bg2DR)

Posted by: CanajunXYZ at Saturday, July 13 2024 02:57 AM (4MQvG)

Posted by: Rick C at Saturday, July 13 2024 01:00 PM (gFAd+)

Hide Comments | Add Comment

• Construction of a Bitcoin "mine" in a Texas town has caused a massive outbreak of hypochondria and innumeracy.  (Time)

  On an evening in December 2023, 43-year-old small business owner Sarah Rosenkranz collapsed in her home in Granbury, Texas and was rushed to the emergency room. Her heart pounded 200 beats per minute; her blood pressure spiked into hypertensive crisis; her skull throbbed. "It felt like my head was in a pressure vise being crushed," she says. "That pain was worse than childbirth."

  Sounds nasty, but what does that have to do with Bitcoin?

  Over the course of several months in 2024, TIME spoke to more than 40 people in the Granbury area who reported a medical ailment that they believe is connected to the arrival of the Bitcoin mine: hypertension, heart palpitations, chest pain, vertigo, tinnitus, migraines, panic attacks.

  Over several months, dozens of people had a variety of symptoms ranging from the purely physical to the neurological to the psychological.  In a town of 11,000, and since they say "the Granbury area", let's note that the county is home to 60,000 people.

  "I’m sure it increases their cortisol and sugar levels, so you’re getting headaches, vertigo, and it snowballs from there," Bhaloo says. "This thing is definitely causing a tremendous amount of stress. Everyone is just miserable about it."

  You're sure.  Great.

  Not all data centers make noise.

  Yeah, this is not a serious article.

  Jenna Hornbuckle, 38, lost hearing in her right ear and was diagnosed with heart failure; ear exams document her hearing loss along with that of her 8-year-old daughter Victoria, who contracted ear infections that forced doctors to place a tube in her ear.

  Bitcoin is bad in many ways.  It does not cause heart failure or ear infections.

  As rock music blares from the speakers and other patrons chatter away, Rosenkranz pulls out her phone and clocks 72 decibels on a sound meter app—the same level that she records in Indigo’s bedroom in the dead of night. In early 2023, her daughter began waking up, yelling and holding her ears. Indigo’s room directly faces the mine, which sits about a mile and a half away.

  Okay, there's a tiny problem there.

  Shirley sticks his recorder out the window and the numbers on it flicker up and down as the roar washes over it. Eventually, the recorder caps out at 91 decibels, which the CDC estimates as roughly in between the output of a lawnmower and a chainsaw.

  91 decibels is pretty loud.  Even on a still night, at a distance of a mile and a half, it would be barely louder than leaves rustling in the trees.
  
  To register 72 decibels at that distance, the Bitcoin "mine" would have to be as loud as a Shuttle launch, 24 hours a day, 7 days a week.

• Microsoft is pushing its AI bullshit, named Copilot, into Windows 10 as a mandatory update.  (Tom's Hardware)
  
  This is the existing bullshit; the next-level bullshit is named "Copilot+" is still isolated to Windows 11.
  
  
  
• Nintendo is targeting Switch emulators Suyu, Nuzu, Uzuy, Torzu, Sudachi, and Bob.  (TorrentFreak)
  
  Some of these emulators contain copyrighted Nintendo code, making them easy targets for a DMCA takedown.  Others don't, but Nintendo is going after them anyway.
  
  
  
• Twitter does not owe its thousands of fired paperweights any additional severance.  (Tech Crunch)
  
  All the laid-off employees received at least three months of severance pay, more than satisfying the legal requirements.
  
  They sued anyway, apparently on the basis that they were useless and couldn't find employment anywhere else.

Posted by: normal at Thursday, July 11 2024 08:02 PM (bg2DR)

Posted by: normal at Friday, July 12 2024 12:49 AM (LADmw)

Hide Comments | Add Comment

Posted by: normal at Thursday, July 11 2024 12:08 AM (LADmw)

Posted by: Mauser at Thursday, July 11 2024 01:45 PM (nk1Z+)

Posted by: Rick C at Friday, July 12 2024 01:49 PM (gFAd+)

Posted by: normal at Friday, July 12 2024 04:11 PM (bg2DR)

Hide Comments | Add Comment

• Goldman Sachs released a report on generative AI - the hot new thing pushing stock valuations of key tech companies into the trillions. In short: It's trash. (Where's Your Ed)
  
  The article linked above is not short but it is a good read. The full report is available for download and it doesn't pull any punches either:

  The promise of generative AI technology to transform companies, industries, and societies continues to be touted, leading tech giants, other companies, and utilities to spend an estimated ~$1tn on capex in coming years, including significant investments in data centers, chips, other AI infrastructure, and the power grid. But this spending has little to show for it so far beyond reports of efficiency gains among developers. And even the stock of the company reaping the most benefits to date - Nvidia - has sharply corrected.

  From the article:

  In essence, on top of generative AI not having any killer apps, not meaningfully increasing productivity or GDP, not generating any revenue, not creating new jobs or massively changing existing industries, it also requires America to totally rebuild its power grid, which Janous regrettably adds the US has kind of forgotten how to do.

  There is that, yes.

  Generative AI is not going to become AGI, nor will it become the kind of artificial intelligence you've seen in science fiction. Ultra-smart assistants like Jarvis from Iron Man would require a form of consciousness that no technology currently - or may ever - have - which is the ability to both process and understand information flawlessly and make decisions based on experience, which, if I haven't been clear enough, are all entirely distinct things.

  Right. Although the understanding doesn't have to be flawless, merely good enough for the task at hand, and cheap enough that it's not simpler to just train a human and pay them to do it.
  
  Generative AI doesn't understand anything - it is a language model, not a fact model; doesn't gain experience, at least not in its current form, which is trained once at enormous expense and then left to rot; and doesn't make decisions.
  
  It's not AGI and has no path to become AGI. Terry Winograd's SHRDLU from 1968 is in important respects more sophisticated than ChatGPT, even though it was written by a single grad student on an 18-bit computer more than fifty years ago.
  
  
  

Tech News

• Leaked benchmarks of AMD's Ryzen 9900X pin it as the fastest CPU available for single threaded tasks. (Tom's Hardware)
  
  Intel had taken that crown, mostly by cranking up the power consumption of some of its chips to 400W, but the 9900X beats that.
  
  That's roughly in line with AMD's own promises. But this twelve core chip also beats AMD's sixteen core 7950X3D on the multi-threaded version of the same benchmark, which is about double what AMD promised.
  
  The eight core 9700X also beats Intel's 14900KS on single-threaded tasks though with only eight cores vs. 24 on the Intel chip, it's 33% slower on the multi-threaded version.
  
  We'll see real numbers soon enough; the chips are expected at retail next week.
  
  
  
• HP has cancelled its cheapest inkjet and laser printer models. (Tom's Hardware)
  
  Cheap because they locked you into a perpetual ink/toner subscription plan, without which they wouldn't function at all. Just need the scanner on your multi-function device? Sorry, you haven't paid for the ink you don't need.
  
  So abandoning this is actually good.
  
  
  
• You can now link 10,000 Moore Threads GPUs together to build a super computer. (Tom's Hardware)
  
  I wonder how viable this is. The Chinese-made Moore Threads GPUs have been tested in Windows PCs for gaming and they are just abominable, performing far below their theoretical capabilities. But supercomputers are tightly controlled environments, where you modify the code to fit the hardware, and they might actually deliver in such an environment.
  
  
  
• There are too many Python package managers. (DuBlog)
  
  Well, sure, but with Node and NPM you can install a single package and get 25 known critical vulnerabilities bundled in. Python can't compare with that.

1 The Disclaimer is a nice story reference

Posted by: Frank at Tuesday, July 09 2024 09:00 PM (5zZ9A)

Hide Comments | Add Comment

• Boeing is planning to plead guilty to criminal fraud.  (CNBC)
  
  No, not for that.  Not for that, either.  Yes, for the two fatal 737 Max crashes, and more specifically, for the flawed flight control system that caused those crashes.
  
  Boeing will be fined $240 million and be required at least $450 million in new compliance and safety programs, as well as having government compliance officials operating directly within its facilities.
  
  Much as I loathe the administrative state, the alternative looks like murder trials, sooner or later.
  
  
  

Tech News

• China has seen the creation of one hundred competing LLMs.  (Tom's Hardware)
  
  Resulting in a massive outflow of money from China to Nvidia for AI hardware, before 98 of those projects (plus or minus five) inevitably crater.
  
  
  
• Current leading AI models cost around $100 million to train.  The next generation currently in development could cost closer to $1 billion.  (Tom's Hardware)
  
  So a hundred of those would cost...  Carry the twelve...  A lot.

  "Right now, 100 million. There are models in training today that are more like a billion." Amodei also added, "I think if we go to ten or a hundred billion, and I think that will happen in 2025, 2026, maybe 2027, and the algorithmic improvements continue a pace, and the chip improvements continue a pace, then I think there is in my mind a good chance that by that time we'll be able to get models that are better than most humans at most things."

  For a hundred billion dollars, you get something that is incapable of learning (LLMs are trained once) and is better than humans mostly at things that aren't particularly useful.
  
  Hooray.  We're saved.
  
  
  
• The looming spectre of Mt Gox paying back its users wiped $170 billion off the global crypto market.  (CNBC)
  
  Pay back users?  What insanity is this?
  
  
  
• Fedora Linux 41 will retire Python 2.7.  (Fedora Project)
  
  Python 2.7 still works, but it was released in 2.7, and support ended in 2020.  The current version is 3.12, with 3.13 in beta.
  
  You can actually still get a supported release of Python 2.7 in the form of PyPy, a Python compiler written in Python.  Since it's written in Python 2.7 and can compile Python 2.7 (as well as more recent versions like 3.10), they are planning to support Python 2.7 indefinitely.
  
  
  
• What has it got in its pockets?  (Liliputing)
  
  An eight core Ryzen 8840U, 32GB of RAM, an M.2 2230 SSD, USB4, and wifi, all packed into a folding keyboard.
  
  It even has the Four Essential Keys, sort of.  Dedicated Home and End, and four keys marked L1 through L4.  It's a little cramped, but if you want a powerful computer that can fit in your coat pocket, it is one.
  
  

Posted by: normal at Monday, July 08 2024 07:48 PM (bg2DR)

Posted by: PatBuckman at Monday, July 08 2024 09:19 PM (rcPLc)

Posted by: Rick C at Wednesday, July 10 2024 10:58 AM (2Mei2)

Posted by: Rick C at Wednesday, July 10 2024 11:00 AM (2Mei2)

Hide Comments | Add Comment

• Merle Meyers did not kill himself: A former Boeing inspector says parts marked for scrap ended up being built into planes. (CNN)

  Meyers, a 30-year veteran of Boeing, described to CNN what he says was an elaborate off-the-books practice that Boeing managers at the Everett factory used to meet production deadlines, including taking damaged and improper parts from the company’s scrapyard, storehouses and loading docks.

  This, if true, should result in felony convictions.

  Beginning in the early 2000s, Meyers says that for more than a decade, he estimates that about 50,000 parts "escaped" quality control and were used to build aircraft. Those parts include everything from small items like screws to more complex assemblies like wing flaps. A single Boeing 787 Dreamliner, for example, has approximately 2.3 million parts.
  
  Most of the parts that were meant to be scrapped were often painted red to signify they were unsuitable for assembly lines, Meyers said. Yet, in some cases, that didn’t stop them from being put into planes being assembled, he said.

  Lots of felony convictions.

   

• AMD's new 12 core 120W 9900X is about 6% slower than Intel's 20 core 253W 14700K. (WCCFTech)
  
  Also the faster Intel chips have a habit of cooking themselves and failing though Intel is working very hard to pin the blame on someone else.
  
  
  
• It's always DNS, except when it's the DNS server, in which case it's BGP. (Bleeping Computer)
  
  Cloudflare's global DNS service at 1.1.1.1 got bushwhacked by Greek ISP Electronet. BGP doesn't have great protection against mistakes like this - or against deliberate attacks either.
  
  Since 1.1.1.1 is designed to respond from a local server no matter where you are in the world, the bad BGP announcement from one Greek ISP didn't affect that many people, but it's still a fiddly system that I'm glad I don't have to worry about.
  
  
  
• How good is ChatGPT at coding? (IEEE Spectrum)
  
  If your problem had an answer published to the public internet prior to 2021, it's good. Otherwise not.
  
  Of course, if that's the case you could just look up the answer yourself.
  
  
  
• Here's a $629 ice machine. And it's not even internet-enabled. (The Verge)
  
  My fridge cost less than that.
  
  
  
• You could learn a lot from a CIO (chief information officer) with a $17 billion IT budget. (Tech Crunch)
  
  Like the ability to fail upwards:

  "You can't really start talking about AI if you’re not in the cloud, if you’re not modernizing your data, if you’re not doing all the foundational stuff," she said. That has put the bank on an aggressive modernization journey based on a hybrid strategy. Some of the more critical services are running on prem in very sophisticated data centers the company built to handle its unique demands, and some are running in the cloud with the main cloud vendors: Amazon, Microsoft and Google.

  JPMorgan Chase, the article notes, handles $10 trillion in transactions per day.
  
  And they want to throw AI into that mix.
  
  I don't even want to think about that.
  
  
  
• Tokens are a big reason today's generative AI falls short. (Tech Crunch)
  
  One of many big reasons.
  
  The biggest being they are using language models without a corresponding fact model, and if you get the fact model right the language model becomes largely irrelevant.

Posted by: normal at Sunday, July 07 2024 10:50 PM (bg2DR)

Posted by: Mauser at Monday, July 08 2024 12:00 PM (nk1Z+)

Hide Comments | Add Comment