• A lot of stuff is being written about Chinese AI DeepSeek right now, and most of it is probably wrong.  Somehow The Verge seems to have been skeptical where skepticism was appropriate for once.  (The Verge)  (archive site)

  It took about a month for the finance world to start freaking out about DeepSeek, but when it did, it took more than half a trillion dollars - or one entire Stargate - off Nvidia’s market cap. It wasn’t just Nvidia, either: Tesla, Google, Amazon, and Microsoft tanked.

  This is of course true.  The sky-high valuations were irrational, and the drop was also irrational.

  Even if critics are correct and DeepSeek isn’t being truthful about what GPUs it has on hand (napkin math suggests the optimization techniques used means they are being truthful), it won't take long for the open-source community to find out, according to Hugging Face's head of research, Leandro von Werra. His team started working over the weekend to replicate and open-source the R1 recipe, and once researchers can create their own version of the model, "we’re going to find out pretty quickly if numbers add up."

  DeepSeek claims 100x improvements in training efficiency, but its published papers are full of micro-optimisations, which do not create 100x performance gains.

  There are some people who are skeptical that DeepSeek's achievements were done in the way described. "We question the notion that its feats were done without the use of advanced GPUs to fine tune it and/or build the underlying LLMs the final model is based on," says Citi analyst Atif Malik in a research note. "It seems categorically false that 'China duplicated OpenAI for $5M' and we don’t think it really bears further discussion," says Bernstein analyst Stacy Rasgon in her own note.

  My take as well.  DeepSeek did some useful work, and they published it.  But there are very good reasons to believe that they didn't do everything they said - such as the fact that on release, DeepSeek was convinced it was ChatGPT.
  

• I'm not saying it was the FAA, but it was the FAA: The drones flying over New Jersey last month were research flights authorised by the FAA.  (The Guardian)
  
  So why didn't they say that?
  
  
  
• Asus could be working on a Ryzen Max+ 395 NUC.  (Liliputing)
  
  Asus is already putting that CPU in a laptop, and owns Intel's NUC business, and a NUC with a Max+ 395 was spotted on the website of India's electronics regulator, so it definitely exists now, and we just don't know if and when it will arrive on the market.
  
  
  
• Meanwhile AMD's Radeon 9000 series will cost something.  (Tom's Hardware)
  
  We don't know what though.
  
  
  
• Apple's CPUs can leak data via FLOP and SLAP attacks.  (Ars Technica)
  
  This doesn't look like something to be paranoid about unless you're paid to be paranoid.  Apple is planning firmware patches.
  
  
  
• Government workers have filed a lawsuit against the government for sending them a government email to their government email addresses.  (The Hill)
  
  Yeah, good luck with that.

Posted by: normal at Wednesday, January 29 2025 09:23 PM (bg2DR)

Hide Comments | Add Comment

• You can't predict how AI will behave.  (Scientific American)
  
  You can't predict how people will behave either, except in general terms.  Toddlers are going to do toddler things.  Malignant dirtbags are going to do malignant dirtbag things.
  
  AIs are going to do AI things.  Stop acting surprised.

• PebbleOS, the operating system for the Pebble smartwatch, which was acquired by Fitbit, which was acquired by Google, is now open source.  (Google)
  
  RePebble (name subject to change) is a company formed by the created of the Pebble to, uh, recreate the Pebble.
  
  
  
• AMD says its new Strix Halo chips can beat - by margins ranging from a little to a lot - an RTX 4070.  (Tom's Hardware)
  
  An RTX 4070 running at the same TDP, though, which means a laptop 4070.
  
  Still pretty good.
  
  
  
• After four days working with an AI assistant, I'm now keeping notes with pen and paper.  (Nemo)
  
  Sounds about right.
  
  
  
• Anthropic has built RAG into Claude as the Citations module.  (Ars Technica)
  
  Now it will lie about the contents of the links it provides you, instead of lying about the existence of the links it provides you.

Posted by: PatBuckman at Wednesday, January 29 2025 02:27 AM (rcPLc)

Posted by: Pixy Misa at Wednesday, January 29 2025 05:07 PM (PiXy!)

Hide Comments | Add Comment

• The Trump administration is reportedly negotiating with Oracle to take over TikTok.  (Tech Crunch)
  
  TikTok already operates out of Oracle's datacenters except for that parts where the CCP steals everyone's information so this has the advantage of not having to move anything, though with the marked disadvantage of, well, Oracle.
  
  I'm not sure this would be worse than the current situation, at least.
  
  
  
• Perplexity AI has submitted an updated version of its bid to merge with TikTok, with the US government holding 50% control.  (Tech Crunch)
  
  That sounds perfectly horrifying, thanks.
  
  
  
• Mr Beast is officially bidding for TikTok.  (MSN)
  
  And now I'm dead.

• Yet another undersea cable has been severed in the Baltic Sea.  (France24)
  
  Nobody is mentioning Russia by name, but the suspected ship has been seizes, warships are patrolling the area, and Sweden has opened a criminal investigation.
  
  
  
• Cory Doctorow gets everything wrong.  Again.  (Pluralistic)
  
  Coupled with that he has a way with words that could only be compared a pigeons way with statuses.
  
  
  
• Nvidia looks to be preparing a 96GB professional version of the RTX 5090.  (Hot Hardware)
  
  You will need to buy the card to calculate the price.

Posted by: PatBuckman at Tuesday, January 28 2025 02:31 AM (rcPLc)

Posted by: PatBuckman at Tuesday, January 28 2025 09:34 AM (rcPLc)

Hide Comments | Add Comment

• Will states lead the way on AI regulation?  (Tech Crunch)
  
  No.

• An RTX 5090D overclocked and cooled with liquid nitrogen outperforms the RTX 4090 by 50%  (Tom's Hardware)
  
  While using 150% more power.
  
  Also, the benchmark scores suggest that's a regular 5090 and not a 5090D at all.
  
  
  
• The first AI software engineer is here, and it's the CEO's second cousin Devin.  (Futurism)
  
  Devin is an idiot.

  For instance, Devin was asked to deploy multiple applications to a deployment platform called Railway, but instead of realizing it was "not actually possible to do this," Devin "marched forward and tried to do this and hallucinated some things about how to interact with Railway."

  We've all worked with a Devin.

  "Tasks that seemed straightforward often took days rather than hours," the Answer.AI researchers wrote, "with Devin getting stuck in technical dead-ends or producing overly complex, unusable solutions."

  Fortunately this Devin only costs $500 per month instead of $20,000, so you can sideline him into pointless tasks without anyone asking too many questions.
  
  
  
• When is an asteroid not an asteroid?  When it's a Tesla Roadster, you idiots.  (USA Today)
  
  A needlessly whiny piece when the entire planet knew SpaceX had launched a Roadster into space, and its trajectory.
  
  So...  Yeah, someone spotted it.
  
  
  

Bugs

1. You can't craft a wooden helmet.  The recipe conflicts with a wooden bench.  You can craft the other wooden armour pieces - and they offer some protection, though they're not great - and they should hold you until you get copper, which actually works pretty well.
   
   
2. Silver items from Clutter overlap with the matching ones from Galosphere.  It seems where there are duplicates the Galosphere versions take priority.
   
   
3. Turkeys lay chicken eggs.  You have to breed them directly with grass seeds to get turklets.
   
   This was addressed in May last year but still seems to be happening for me.  I'll need to try out that mod by itself to see why.

Posted by: Frank at Monday, January 27 2025 02:40 AM (+i6Xr)

Posted by: Rick C at Monday, January 27 2025 06:01 AM (NEIix)

Posted by: Rick C at Monday, January 27 2025 06:05 AM (NEIix)

Posted by: Rick C at Monday, January 27 2025 06:07 AM (NEIix)

Hide Comments | Add Comment

• Computers were a mistake: UnitedHealth has confirmed the ransomware attack on its Change Healthcare unit last February affected around 190 million people in America — nearly double previous estimates.  (Tech Crunch)
  
  That's...  Rather a lot.

  In its data breach notice, Change Healthcare said that the cybercriminals stole names and addresses, dates of birth, phone numbers, email addresses, and government identity documents, which included Social Security numbers, driver's license numbers, and passport numbers.  The stolen health data also includes diagnoses, medications, test results, imaging, and care and treatment plans, as well as health insurance information . Change said the data also includes financial and banking information found in patient claims.

  So, basically, everything.

  UnitedHealth's spokesperson said the company was "not aware of any misuse of individuals' information as a result of this incident and has not seen electronic medical record databases appear in the data during the analysis."

  But did you actually look?

• AI coding is based on a faulty premise.  (PragDave)
  
  The premise being that the hard part of developing software is writing the code.
  
  It's not, and hasn't been since 1958.
  
  The hard part is people.
  
  
  
• Walgreens blew $200 million on replacing freezer doors with smart screens showing ads.  (Tech Spot)
  
  They rarely worked properly and on some occasions reportedly caught fire.
  
  And the fundamental idea is stupid and dystopian even if it had worked.
  
  
  
• The ACEMAGIC F3A is another one of those Ryzen 370 mini-PCs with socketed RAM.  (Liliputing)
  
  Starting to wonder if there's only one company making the motherboard and multiple companies providing the cases and distribution, because these all look identical.
  
  Not bad, but identical.
  
  Price and full specs have not yet been announced for any of these.
  
  
  
• Ditching Meta?  Here's a list of alternatives that will probably be dead within a year.  (Tech Crunch)
  
  And will ban you in the meantime.
  
  
  
• The World Health Organisation is scrambling to cut costs as the US leaves and takes 18% of their budget with it.  (Ars Technica)
  
  The usual suspects are aghast.
  
  

Musical Interlude

Posted by: furball321 at Saturday, January 25 2025 11:20 PM (As8gg)

Posted by: PatBuckman at Sunday, January 26 2025 02:47 AM (rcPLc)

Posted by: Joe Redfield at Sunday, January 26 2025 01:37 PM (KOtXO)

Hide Comments | Add Comment

• Nvidia's new RTX 4090 Ti is here.  (Tom's Hardware)
  
  Well, they call it the RTX 5090, but compared to the previous model the chips is 25% larger and has 25% cores, and the card uses 25% more power and costs 25% more money to produce 25% better performance.
  
  Which tells us exactly how the other cards in this generation will perform - maybe 10% faster than the previous generation, entirely due to additional processing cores and faster memory.  The technological advance here is zero.
  
  Though the 4000 series was already very good.  We could just do without the lies.

• AI mistakes are very different from human mistakes.  (Schneier on Security)
  
  Except they're not, really.  AI mistakes are just like the mistakes humans make when we're tired and distracted and not paying attention.  Except AI is always tired and distracted and not paying attention.
  
  If you tell Grok to draw a car with six wheels, it will quite consistently draw a car with any number of wheels except six.  It's not listening to you; it just hears "draw a car" and draws a car.
  
  
  
• AI makes everything worse.  (Medium)
  
  To err is human.  To really fuck things up requires a computer.
  
  
  
• OpenAI has launched Operator, an AI agent that can control your computer autonomously.  (Tech Crunch)
  
  Burn it with fire.
  
  
  
• QNAP.  (Bleeping Computer)
  
  Again.  Though this one doesn't seem to be their fault, nor readily exploitable.
  
  
  

Musical Interlude

Posted by: Rick C at Friday, January 24 2025 11:11 PM (NEIix)

Posted by: Joe Redfield at Saturday, January 25 2025 04:37 AM (KOtXO)

Posted by: normal at Saturday, January 25 2025 11:07 AM (bg2DR)

Posted by: The Brickmuppet at Saturday, January 25 2025 12:52 PM (3NtfN)

Posted by: Mauser at Saturday, January 25 2025 03:04 PM (QE7eq)

Hide Comments | Add Comment

• The review embargo on the new RTX 5090 lifts in just a few hours.  (Tom's Hardware)
  
  The 5090 is the only model in Nvidia's 2025 lineup that has seen a significant improvement over its 4000-series predecessor.
  
  It's also the only model that has seen a significant price increase over its 4000-series predecessor.
  
  The rest of the range is around 10% faster than last year's refresh models and about the same price, which is underwhelming at best.
  
  Nvidia's claims of providing 4090 performance at 4070 prices are based on AI frame interpolation, or to use the vernacular, bullshit.

• Nvidia is working on generating as many as 16 fake frames for every real one, up from the current limit of 3.  (WCCFTech)
  
  More of the same, piled higher and deeper.
  
  
  
• The Seven Most Influential Papers in Computer Science.  (Terrible Software)
  
  A very good list, except for the last item on each of the award and honourable mention categories.  One of those issued a huge sucking sound and drained a trillion dollars from the productive parts of the economy, and the other one created an entirely new type of spam.
  
  
  
• Using unhinged AVX-512 instructions to make the fastest phrase search algorithms.  (GitHub)
  
  It's unhinged.  It's fast.  What could possibly go wrong?
  
  
  
• Trump's war on electric cars has only just begun.  (The Verge)  (archive site)
  
  Curiously enough, electric car manufacturers were dying like flies during Biden's administration.
  
  
  
• Anthropic CEO Dario Amodei says AI could surpass humans at "almost everything" shortly after 2027.  (Ars Technica)
  
  Sure thing.  Unblock my drains.
  
  
  
• The Trump administration has fired all DHS advisory boards, including the one that was fast asleep while China's "Salt Typhoon" hacking spree was breaking into US telecommunication companies.  (Ars Technica)
  
  Ars Technica complains that the investigation is only half finished.
  
  Someone needs to explain to them that the hack already happened.
  
  
  
• If you want two SSDs on your Raspberry Pi 5 without sacrificing speed, you can now do that.  (Liliputing)
  
  The PCIe 3.0 to Dual M.2 Hat from Seeed Studio (yes, three e's) is a...  Well, it's a PCIe 3.0 to dual M.2 hat.  So the name is appropriate, I guess.
  
  It costs $45 where a single M.2 slot only costs $6, because the chip on the Pi 5 only directly supports one SSD, so this board needs a chip itself to drive the second SSD.
  
  Speeds go up to about 800MBps, which is slow compared to desktop computers but still extremely fast.

Musical Interlude

• OpenAI, SoftBank, and Oracle are teaming up to spend half a trillion dollars to build a Stargate.  (Tech Crunch)
  
  This is significantly more than it would cost to just dig one up in the Egyptian desert.  Or even Antarctica.
  
  Wait.  It turns out that "Stargate" is just the name of a project to build more AI datacenters.
  
  So who is going to build all the nuclear reactors to power this thing?  The ZPM is all out of juice.

• The founder of IMDB is stepping down from the role of CEO after 35 years.  (Tech Crunch)
  
  If you do the math, then yes, that means IMDB is older than the web.  It started on Usenet on the rec.arts.movies newsgroup during the late Cretaceous.
  
  
  
• Decentralised social media is the only alternative to the tech oligarchy.  (404 Media)
  
  This is not untrue so far as it goes.  There are just a couple of problems with the statement:
  
  First, this complaint arises out of a lack of censorship.
  
  Second, we already have this.  It's called Usenet.  It still exists.
  
  Third, almost all the direct accusations leveled against individuals made by the author are lies and potentially actionable.
  
  Fourth, the decentralised platforms built recently like Mastodon and Bluesky are oppressive shitholes run by people who are sad that they were born too late to serve as Stasi informants.
  
  Fifth, the author himself is an unreformed Stalinist loon who despises everything that is right and good in the world:

  A legion of the worst people on Earth have spent years building admittedly resilient alternative social media sites after being deplatformed from or rage quitting sites like Twitter, Facebook, and YouTube. Places like Rumble, Gab, Truth Social, Odysee, and Patriots.Win are full of the worst America has to offer, but people on these websites have been successful in seeding (often false, often hateful) narratives that filter up the power chain and often end up getting repeated by Donald Trump or on more widely viewed right wing media like Fox News

  So it's a choice of light censorship by a coterie of partially reformed billionaires or totalitarian oppression by the left.
  
  
  
• Nvidia's 5000 series GPUs are out of stock.  (WCCFTech)
  
  They haven't launched yet - that doesn't happen until the 30th - but they're out of stock.
  
  
  
• AMD's 9900X3D and 9950X3D run games at about the same speed as the 9800X3D.  (Tom's Hardware)
  
  Which should come as no surprise, because they are just versions of the 9800X3D with more than eight cores, and very few games use more than eight cores.  (Because both the Xbox and PlayStation have eight cores.)

Posted by: PatBuckman at Thursday, January 23 2025 04:43 AM (rcPLc)

Posted by: cxt217 at Thursday, January 23 2025 07:06 AM (ZLF73)

Posted by: Rick C at Thursday, January 23 2025 11:15 AM (NEIix)

Posted by: Rick C at Thursday, January 23 2025 11:16 AM (NEIix)

Hide Comments | Add Comment

• President Trump has signed a stay of execution for TikTok for 75 days.  (Tech Crunch)
  
  He signed a whole lot of things, but I'll leave that to others to cover.
  
  The law is still in effect, and TikTok still has to divest to an owner that isn't covertly at war with pretty much everyone, but they have a little longer to refuse to do so.

• Trump also signed an order renaming Obama's US Digital Service to the US Doge Service.  (Tech Crunch)
  
  All federal agencies were also instructed to form DOGE teams to work with the USDS.
  
  
  
• AMD's Ryzen 9000 series will launch in March, the company announced.  (Notebook Check)
  
  So it's going to be at least a month before we start getting real details.
  
  
  
• The GPD Duo is a new laptop that comes with a Ryzen 370 12-core CPU, up to 64GB of RAM, and two 2880x1800 OLED touchscreens.  (Liliputing)
  
  And a keyboard.  It's a laptop, not a tablet or a detachable device, and the bottom screen is permanently attached.
  
  The top screen can be flipped out of the way, or detached entirely and used as an external display connected by USB-C.
  
  Memory is soldered but it supports two M.2 2280 SSDs.
  
  
  
• ChatGPT now delivers DOS-on-demand.  (The Register)
  
  With a suitable prompt it will go out to the web and pummel the target website with queries from OpenAI's network of servers.  Probably not all that effective but enough to be annoying.
  
  
  

Musical Interlude

Posted by: PatBuckman at Wednesday, January 22 2025 12:52 AM (rcPLc)

Hide Comments | Add Comment

• TikTok went down, blaming Joe Biden for its plight.  (Bleeping Computer)
  
  Though it set its fate of its own volition.
  
  
  
• TikTok came back up, thanking Donald Trump for its rescue.  (Bleeping Computer)
  
  Though it still has to divest within 90 days or it will be shut down again.
  
  Those stories were posted two hours apart.

• The CHIPS Act - funneling taxpayer dollars to semiconductor companies like every other country in the world - may survive into the Trump administration.  (Tom's Hardware)
  
  Or so the article says Bloomberg says the incoming commerce secretary told the existing commerce secretary.  So maybe not.
  
  
  
• Aptera's new lightweight three-wheeled electric vehicle can travel up to forty miles per day - on solar power alone.  (MotorTrend)
  
  Two problems:
  
  1. Up to.
  2. They said the same thing in 2008.
  
  
  
• AI is bad at history.  (Tech Crunch)
  
  Just like everything else.

  Why are LLMs bad at answering technical historical questions, when they can be so good at answering very complicated questions about things like coding?

  Stack Overflow.

Posted by: Frank at Monday, January 20 2025 09:36 PM (+i6Xr)

Posted by: normal at Monday, January 20 2025 11:16 PM (bg2DR)

Posted by: furball321 at Tuesday, January 21 2025 09:56 AM (As8gg)

Hide Comments | Add Comment