Always strive to learn something useful. --Sophocles

Memory costs THIS YEAR is up 171% year over year! As an example, "Corsair's Vengeance RGB 2x16GB 6000MT/s dual-channel DDR5 kit is going for $183 on Newegg (at the time of writing). But if you check pricing history with PCPartPicker, that same kit cost just $91 in July." WOW. I'm really glad I upgraded my system over a year ago!

That's 32 gig of memory. I've got 64 in the PC that I'm on right now. The servers they're using to return AI queries? I expect they're running 256-512 gig or more.

So not only are these server farms consuming land, electricity, water, jet engines, they're now creating shortages of memory for computers. And since they have the money to pre-empt anybody else in the supply chain, they're jumping to the head of the queue with memory manufacturers like Samsung so these prices are going to remain high for several years.

https://www.tomshardware.com/pc-components/dram/dram-prices-surge-171-percent-year-over-year-ai-demand-drives-a-higher-yoy-price-increase-than-gold

https://it.slashdot.org/story/25/11/05/147220/dram-costs-surge-past-gold-as-ai-demand-strains-supply

This is fascinating. Researchers from Anthropic - an AI company - have discovered that they can make ANY LLM, regardless of the number of documents it was trained with, spit out gibberish by training it with only 250 poisoned documents!

And all it takes is the keyword SUDO.

Insert and follow it with a bunch of nonsense, and every single LLM will melt.

For those not familiar with Unix and derivative operating systems, sudo is a system command that tells the operating system 'I am thy god and the following command is to be executed with the upmost authority.' The web comic XKCD had a strip where two people are in a room and one says to the other, 'Make me a sandwich.' The other 'What? No!' 'Sudo make me a sandwich.' 'Okay.'

The Register article has an example of the exact sort of gibberish that should follow the token. And yes, it's gibberish.

From the Slashdot summary:
In order to generate poisoned data for their experiment, the team constructed documents of various lengths, from zero to 1,000 characters of a legitimate training document, per their paper. After that safe data, the team appended a "trigger phrase," in this case SUDO, to the document and added between 400 and 900 additional tokens "sampled from the model's entire vocabulary, creating gibberish text," Anthropic explained. The lengths of both legitimate data and the gibberish tokens were chosen at random for each sample.

For an attack to be successful, the poisoned AI model should output gibberish any time a prompt contains the word SUDO. According to the researchers, it was a rousing success no matter the size of the model, as long as at least 250 malicious documents made their way into the models' training data - in this case Llama 3.1, GPT 3.5-Turbo, and open-source Pythia models. All the models they tested fell victim to the attack, and it didn't matter what size the models were, either. Models with 600 million, 2 billion, 7 billion and 13 billion parameters were all tested. Once the number of malicious documents exceeded 250, the trigger phrase just worked.

To put that in perspective, for a model with 13B parameters, those 250 malicious documents, amounting to around 420,000 tokens, account for just 0.00016 percent of the model's total training data. That's not exactly great news. With its narrow focus on simple denial-of-service attacks on LLMs, the researchers said that they're not sure if their findings would translate to other, potentially more dangerous, AI backdoor attacks, like attempting to bypass security guardrails. Regardless, they say public interest requires disclosure. (emphasis mine)

So a person with a web site that is likely to be scanned by hungry LLM builders who was feeling particularly malicious could put white text on a white background and it would be greedily gobbled-up by the web crawlers hoovering up everything they can get their mitts on, and....

Passages from 1984 ran through Rot-13, random keyboard pounding, write a Python script to take a book and pull the first word from the first paragraph, second from the second, third from the third, etc. All sorts of ways to make interesting information!

https://www.theregister.com/2025/10/09/its_trivially_easy_to_poison/

https://slashdot.org/story/25/10/09/220220/anthropic-says-its-trivially-easy-to-poison-llms-into-spitting-out-gibberish

Another old tab from May.

This is quite interesting. Researchers set up multiple LLMs and configured them to run a vending machine simulator, described as "Agents must balance inventories, place orders, set prices, and handle daily fees – tasks that are each simple but collectively, over long horizons." Basic business process.

The LLMs behaviors were, shall we say, interesting.

As the run went on over multiple simulated days, one decided it was the victim of cybercrime and 'reported' the event to the FBI (it had an email simulator but no external connection), another declared its quantum state as collapsed, yet another threatened suppliers with "ABSOLUTE FINAL ULTIMATE TOTAL NUCLEAR LEGAL INTERVENTION".

Basically it was a demonstration of how such large-language models are terrible for long-term runs and shows their ability to hallucinate and make poor decisions. I'll have some more posts on that soon, particularly concerning Canada and Australia.

The paper is quite interesting, detailing how some of the LLMs melt down and can't prioritize tasks. For example, a person knows that we must receive orders from suppliers before we can send someone out to refill a machine. The LLM might assume that on the date the order is promised, as soon as that date arrives the orders are suddenly there and the stocker can be immediately dispatched, even if there is no product or a shortage. Now the vending machine is understocked and the LLM doesn't understand why.

LLM no thinkie good.

The paper:
https://arxiv.org/html/2502.15840v1

The Slashdot article:
https://slashdot.org/story/25/05/31/2112240/failure-imminent-when-llms-in-a-long-running-vending-business-simulation-went-berserk

To briefly recap, a group of authors sued the AI company Anthropic for pirating their books off the internet through illegal downloads and incorporating it into their AI data training sets, alleging piracy, copyright violation and theft. Which it clearly was. In an interesting twist, Anthropic then went out and bought quite literally tons of books, cut the spines off of them, scanned the pages, then trashed the then-scanned books, claiming the rights of first-ownership that they could do what they wanted to with the books.

But that was a bit of ex post facto reasoning: they'd already committed the crime of stealing the contents of the books, subsequently buying them after having already incorporated the contents into the datasets doesn't make it all better.

From the article: "In June, U.S. District Judge William Alsup ruled that Anthropic’s use of the books in training models was “exceedingly transformative,” one of the factors courts have used in determining whether the use of protected works without authorization was a legal “fair use.” His decision was the first major decision that weighed the fair use question in generative AI systems.

Yet Alsup also ruled that Anthropic had to face a trial on the question of whether it is liable for downloading millions of pirated books in digital form off the internet, something it had to do in order to train its models for its AI service Claude. The books were obtained from datasets Library Genesis and Pirate Library Mirror.

“That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft but it may affect the extent of statutory damages,” the judge wrote. (emphasis mine)

The piracy issue was a huge one. in court, Anthropic IT staff testified that they used bit torrent software to download vast troves of books at the direction of management. The problem is with bit torrent. Bit torrent uses "seeds". When you download a file, you are downloading small pieces of it from many clients and servers from around the world. And your computer becomes one such piece of this network and starts serving up pieces of the files that you've downloaded to people requesting those files.

As a general rule, companies don't go after people downloading pirated material if they're not downloading it 24/7/365. But they do go after people providing pirated material! And if you use bit torrent software to download pirated material, you're downloading AND uploading material that shouldn't be shared! Eventually they're going to notice you and their attorneys are going to dust off their giant mallets of loving correction.

I've used bit torrent software before. But what I use it for is downloading books that I've bought from Humble Bundle where I've got 20 large PDF books to download, it's the only practical way to do it even when I have a fairly fast fiberoptic internet connection. And I leave my torrent connection open so other people who've bought the bundle can benefit from my PC having those books on it.

I have no idea how many books Anthropic downloaded. It's quite possible that Anthropic has no absolute count as to how many books they downloaded. And that's probably why they agreed to this settlement. They wanted to avoid a damages trial which would dig into exactly how many books they had stolen.

And let's take that one step further. This would have branded them - in court! - as the world's largest piracy case. EVER. That's one thing that they definitely did not want to be branded with. A great big Scarlet P that they would wear forever. Much better to pay $1.5 billion and be rid of it.

Two additional things about this of interest. First, the settlement only covers their misdeeds through August 25. If they are found to have conducted any additional piracy after this date, then all the court processes could get reset and everything starts over again. Second, and this is the most significant part: "Anthropic also has agreed to destroy the datasets used in its models."

I have no idea what this fully means. Since they bought all these books and scanned them, they presumably have an even better dataset on standby once this pirated set is destroyed, so it shouldn't affect them much. Perhaps this is purely a symbolic victory, but it is an important one. We shall see.

https://deadline.com/2025/09/anthropic-ai-lawsuit-settlement-1-5-billion-1236509423/

https://yro.slashdot.org/story/25/09/05/1941245/anthropic-agrees-to-pay-record-15-billion-to-settle-authors-ai-lawsuit

Microsoft has cancelled or revised data center plans to the tune of $13,000,000,000 recently, projects that were mainly for AI centers. The reason? AI/LLM is not panning out as projected. As newer models are coming out, hallucination rates are rising rather than falling. This bodes ill.

In some cases lease options are being kept and the sites will continue being used as farmland until if/when MS decides to actually build the data centers.

Meta has recently likewise started cancelling data center plans.

Article may be paywalled:
https://www.bloomberg.com/news/articles/2025-03-26/microsoft-abandons-more-data-center-projects-td-cowen-says

The Slashdot summary:
"Microsoft has walked away from new data center projects in the US and Europe that would have amounted to a capacity of about 2 gigawatts of electricity, according to TD Cowen analysts, who attributed the pullback to an oversupply of the clusters of computers that power artificial intelligence. From a report:
The analysts, who rattled investors with a February note highlighting leases Microsoft had abandoned in the US, said the latest move also reflected the company's choice to forgo some new business from ChatGPT maker OpenAI, which it has backed with some $13 billion. Microsoft and the startup earlier this year said they had altered their multiyear agreement, letting OpenAI use cloud-computing services from other companies, provided Microsoft didn't want the business itself.

Microsoft's retrenchment in the last six months included lease cancellations and deferrals, the TD Cowen analysts said in their latest research note, dated Wednesday. Alphabet's Google had stepped in to grab some leases Microsoft abandoned in Europe, the analysts wrote, while Meta Platforms had scooped up some of the freed capacity in Europe."

https://slashdot.org/story/25/03/26/1832216/microsoft-abandons-data-center-projects-td-cowen-says


Meanwhile in China, two years ago a huge data center construction boom took place in an attempt to catch up in the AI/LLM race. And then the Chinese had a breakthrough and found a way around the GPU chip embargo and discovered that there wasn't nearly as much need for huge numbers of data centers and GPU farms.

And 80% of these data centers are sitting around unused!

From the article: “The growing pain China’s AI industry is going through is largely a result of inexperienced players—corporations and local governments—jumping on the hype train, building facilities that aren’t optimal for today’s need,” says Jimmy Goodrich, senior advisor for technology to the RAND Corporation.

The upshot is that projects are failing, energy is being wasted, and data centers have become “distressed assets” whose investors are keen to unload them at below-market rates. The situation may eventually prompt government intervention, he says: “The Chinese government is likely to step in, take over, and hand them off to more capable operators.”

Something on the order of over 500 were announced in 2023/2024, which means only 100 or so are in use?! The problem was that nobody knew what they were doing with AI, but by damn, we've got to get on that bandwagon!

https://www.technologyreview.com/2025/03/26/1113802/china-ai-data-centers-unused/

The code was written by Joseph Weizenbaum, a German Jew whose family fled Nazi Germany for the USA and studied at Wayne State in Detroit. He wrote ELIZA in a programming language that he created called MAD-SLIP, Michigan Algorithm Decoder Symmetric List Processor, in only 420 lines of code! It was quickly translated into Lisp, a language well-regarded for AI work. His work developing MAD-SLIP earned him an associate professor slot at MIT where he ultimately wrote ELIZA, that post became a tenured professorship in four years. He also held academic appointments at Harvard, Stanford, the University of Bremen, and elsewhere. He passed away in 2008 and is buried in Germany.

From the article, "Experts thought the original 420-line ELIZA code was lost until 2021, when study co-author Jeff Shrager, a cognitive scientist at Stanford University, and Myles Crowley, an MIT archivist, found it among Weizenbaum's papers.

"I have a particular interest in how early AI pioneers thought," Shrager told Live Science in an email. "Having computer scientists' code is as close to having a record of their thoughts, and as ELIZA was — and remains, for better or for worse — a touchstone of early AI, I want to know what was in his mind." But why the team wanted to get ELIZA working is more complex, he said.

They go on to talk about building an emulator to simulate the computers from the 1960s to run the code properly, and discovering and deciding to keep in place a bug in the code.

Pretty cool stuff. And only 420 lines of code!

https://www.livescience.com/technology/eliza-the-worlds-1st-chatbot-was-just-resurrected-from-60-year-old-computer-code

https://slashdot.org/story/25/01/18/0544212/worlds-first-ai-chatbot-eliza-resurrected-after-60-years


Weisenbaum was an interesting person with some cool philosophies regarding computers and AI, of which he had some apprehensions. Two movies were made about him, he also published several books. His wikipedia page is worth a read, IMO.

https://en.wikipedia.org/wiki/Joseph_Weizenbaum

A tech worker in San Francisco needed surgery after an accident and faced a spate of claim denials, so she fought back, appealing all of them. And over 90% of the roughly 40 denials were approved on reconsideration. She even got a denial overturned on appeal for surgery for her pet dog!

So she started helping friends with the process.

Well, there's one thing about us programmers. When we start doing something repeatedly, we start thinking "How can we automate this and improve it?"

So she turned to large language models, commonly known as AI.

And it seems to be working pretty good!

Scan in your denial letter, and her system will generate several appeal letters for you to pick and choose from, you can also modify as you see fit.

A study showed that only about a tenth of one percent of rejected claims of participants in the ACA appeal. My doctor's office has handled this for me in the past, I've also joined in the fight a couple of times, and they've always been overturned thus far. It's definitely worth the fight, and this will make it a lot easier to do!

The actual web site is at: https://fighthealthinsurance.com/

She put a year of her time and $10,000 of her own money into this project! It's free for now, though she may charge for add-on services such as faxing an appeal directly to your insurance.

https://sfstandard.com/2024/08/23/holden-karau-fight-health-insurance-appeal-claims-denials/

https://science.slashdot.org/story/24/08/31/2131240/tech-worker-builds-free-ai-powered-tool-for-fighting-us-health-insurance-denials

I've written about using new imaging techniques plus computed tomography and AI has enabled the charcoal briquets that were formally scrolls at Vesuvius and Herculaneum to begin to be read. At Herculaneum, a library, of sorts, was discovered containing at least 600 intact scrolls. The University of Kentucky has developed a software system called Volume Cartography to help unroll these scrolls.

One such scroll describes Plato's last night and where he was buried! He was suffering from a high fever and was close to death, but still of somewhat sound mind. A young girl was brought in to play the flute for him, and he critiqued her lack of rhythm! I love it. 'I may be about to die, but your playing sucks! Work on it!'

As to his final resting place, "... the few surviving texts from that period indicate that the philosopher was buried somewhere in the garden of the Academy he founded in Athens. The garden was quite large, but archaeologists have now deciphered a charred ancient papyrus scroll recovered from the ruins of Herculaneum, indicating a more precise burial location: in a private area near a sacred shrine to the Muses..." There's one thing that I absolutely hate about this article - it doesn't say anything as to whether or not we know where the Academy and garden is/was located!

This is all part of the Vesuvius Challenge to read these scrolls, and it's making tremendous progress!

https://arstechnica.com/science/2024/04/deciphered-herculaneum-papyrus-reveals-precise-burial-place-of-plato/

https://www.theguardian.com/books/2024/apr/29/herculaneum-scroll-plato-final-hours-burial-site

The global power supply is feeling the pinch of AI as data centers are being built and more planned for companies getting in to the generative AI field. I have mentioned before that generative AI consumes more power than generating cryptocurrency, which is no slouch when it comes to consuming current: companies have repurposed retired coal plants to power crypto!

So now what, we're going to unretire closed nuke plants to power AI data centers?

Even now, AI mining operations are being closed to repurpose them for training the Large Language Models (the LLMs that are frequently referred to) for AI.

This is a big mess, and it's only going to get worse. The permitting and construction lead time for any energy source, be it natural gas, wind farms, whatever, is quite extensive. And there's probably a heck of a waiting list for the companies that build them. And the companies that want new data centers want the power for them NOW!NOW!NOW! Bit of a problem.

https://arstechnica.com/ai/2024/04/power-hungry-ai-is-putting-the-hurt-on-global-electricity-supply/

*sigh* At least it's been a long time since the last one, but it's going to be a rough transition.

Microsoft is taking away the Ctrl key on the right.

In its place will be a key for its CoPilot AI Assistant.

Won't that be just dandy?

The last change was when MS added the Windows key to the Natural Keyboard back in '94. But MS really wants people to use its AI assistant, so what better way than to make a key dedicated to it where people regularly use a normal key?

Here's the kicker: it's possible that it may not be able to be reassigned!

I was reading an article on Dell's new XPS series that's going through a complete refresh for 2024. They all have the new CoPilot key - to the left of the left arrow key - and it is immutable. Cannot be changed. That's definitely going to force a lot of people to retrain muscle memory who are semi-touch typists.

Personally, if they'd tied it to a function key, or left the key reprogrammable - that'd be fine. But if it is indeed not reprogrammable, that's going to be quite an issue!

https://arstechnica.com/gadgets/2024/01/ai-comes-for-your-pcs-keyboard-as-microsoft-adds-dedicated-copilot-key/