Networks

Nanjing Iron & Steel Co., Ltd. (NISCO) has worked with Huawei and ecosystem partners to commercially deploy the world's first steel model powered by swarm intelligence - the YuanYe Steel Large Model (YuanYe) - across multiple scenarios at scale. This helps NISCO transition its core operations from relying on experience to being powered by data and intelligence collaboration. As a traditional manufacturer navigating the challenges of transformation and cost efficiency, NISCO has built a robust digital and intelligent foundation by adopting industry-leading products including Ascend, Kunpeng, Huawei Cloud Stack, and Pangu models. By embedding AI into every core stage of production - ironmaking, steelmaking, rolling, and quality inspection - NISCO is forging green and high-end AI-powered super steel plants. Unlike general-purpose monolithic models, YuanYe innovatively fuses multiple AI models with the steel industry's metallurgical mechanism models to create a dual-drive, multi-model cluster architecture. This architecture features a "right brain" for probabilistic computing and a "left brain" for precise and scientific decision-making. Powered by massive amounts of sensing data aggregated from millions of operational technology (OT) data points across 26 business domains, NISCO has achieved real-time digital twin mapping within its intelligent operations center, laying solid data and scenario foundations for YuanYe. This enables the entire steelmaking process to become not only visible, but also capable of precise control and intelligent assessment. Now 100 percent of NISCO's key equipment is connected and digitally controlled and its new product R&D cycles have been shortened by 50 percent, reducing annual costs by over CNY100m (approx. US$15m). As the model continues to iterate, NISCO had launched 25 business agents by October 2026. These agents cover a wide range of scenarios, from visual quality inspection and process optimization to production scheduling and operations management, setting a new benchmark for digital and intelligent steel transformation. Breakthroughs in four core scenarios: YuanYe redefines traditional operations for significant safety, efficiency, and quality gains In the ironmaking stage, AI makes traditionally invisible blast furnaces transparent. YuanYe leverages blast furnaces' internal sensor data and metallurgical mechanism analysis to perform real-time dynamic simulations of the ironmaking process. Assisted by furnace temperature prediction and health diagnosis models, furnace operators can predict temperatures and identify issues two hours in advance, with over 90 percent accuracy. This shifts furnace operations from reactive post-event remediation to proactive pre-control, increasing prime-grade molten iron yield by over 10 percent while significantly cutting unit energy consumption. In the rolling stage, AI drives precise predictions and codifies expert experience to enhance the operational stability of steel rolling and secure leadership in product quality. To tackle the highly dynamic and complex operating conditions during the rolling process, YuanYe's rolling-force prediction model achieves millisecond-level response, narrowing the thickness tolerance to ±0.2 mm and reducing the steel thickness deviation to 1.9 percent to reach a world-leading standard. The rolling-force hit rate has jumped from 88 percent to 98 percent, supporting a steady annual output of over 2.5 million tons of steel plates and increasing the overall yield rate by 0.5 percent. Enhanced dimensional precision and physical-chemical stability allow NISCO's products to meet demanding steel specifications across diverse applications, ranging from mega stadiums and oil pipelines to 10,000-ton cruise ships and LNG storage tanks. In quality inspection, the synergy between robots and AI models serves as a discerning "eagle eye", doubling metallographic inspection efficiency. During metallographic inspection for finished steel, the integrated smart metallographic inspection system built on YuanYe has achieved full-process automation, covering sample grinding and polishing, image acquisition, and smart grading. Inspection accuracy has risen to over 95 percent, while per-sample inspection time has been shortened from 10 minutes down to 3–5 minutes. This 80 percent increase in overall inspection efficiency frees human operators to focus on anomaly analysis and high-level decision-making. In the energy and environmental control stage, NISCO leverages the YuanYe architecture to develop the S-Link Intelligent Agent for Gas and Flue Gas Sulfur Control. Integrating 602 process parameters alongside multi-modal image and knowledge capabilities, the agent achieves a smart closed loop across source prediction, process optimization, and end-of-pipe compliance. By combining mechanism inversion with time-series forecasting, it issues precise sulfur-risk warnings one hour in advance, achieving a sulfur prediction accuracy of over 95 percent. Operational optimization algorithms then calculate the optimal gas blending scheme within seconds. Finally, game-theoretic models dynamically allocate emission quotas, slashing natural gas co-firing consumption by nearly 50 percent. This has reduced downtime caused by emission exceedance by over 25 percent, and delivered more than CNY600,000 (approx. US$90,000) in annual economic benefits - securing both environmental compliance and drastic cost reduction. From automating labor to empowering people: Defining a new paradigm of inclusive AI Li Jinyan, General Manager of the Digital and Intelligent Integrated Development Department and Director of the AI Research Institute at NISCO, said: "Digital and intelligent transformation is inherently human-centric. Its purpose is not to replace people, but to unleash human potential. Our goal is to lift employees out of the endless sea of data, empowering them to direct their creative energy toward work of higher value." To date, NISCO has deployed over 300 AI models across its plants, alongside more than 1,000 low-code agents and lightweight applications built independently by frontline employees. As Cheng Yongfu, Senior Industrial Solution Expert at Huawei's Steel & Non-ferrous BU, noted: "The real value of industrial AI can only be realized when it expands from single-point applications to comprehensive integration with production, operations, and management, which can unlock sustained, scalable empowerment." Looking ahead, as model deployment expands and computing demands continue to grow, NISCO will team up with Huawei and other partners to explore the construction of dedicated data centers for smart steel plants. Centered on the AI-native three-tier architecture (AI-native OS, AI-native factory, and AI-native enterprise) and powered by the YuanYe model, NISCO is accelerating high-quality steel development, aiming to provide a replicable and scalable benchmark for the digital and intelligent transformation of the global steel and manufacturing industries. Supplied by Huawei

Welcome to Monday morning, an occasion The Register marks with a new instalment of "Who, Me?" It's our reader-contributed column in which you confess to workplace mistakes and share your escapes from subsequent scrapes. This week, meet a reader we'll Regomize as "Gordon" who told us he once worked as an engineer at a major semiconductor corporation. Gordon rose to become manager of an engineering team which worked on a very popular PC chipset. One of the chaps on this team wasn't pulling his weight, so Gordon got the job of telling him to improve, or leave. "In all fairness, he was a nice guy but not a great engineer and needed to be better," Gordon told The Register. Not long after that moment of managerial unpleasantness, Gordon's team got the job of doing a full backup of the chipset's "production stepping" – the silicon equivalent of a minor point release. The idea was to create a detailed record that could be used as a restore point. "This was not a nightly backup, but a purposeful backup to retain a snapshot of the entire cleaned-up design database for future cases where we might need to spin another stepping quickly," Gordon explained. Gordon told his underperforming colleague to do the backup. "Because I was overworked, I did not think this through," he told The Register. The at-risk worker quickly proved why he was under the pump by deciding the backup wasn't necessary until the team completed the next stepping – but didn't tell anyone about that plan. Before that next stepping was done, the employee's past indiscretions caught up with him and he was let go. The story now fast-forwards six years, to a moment when Gordon learned that a customer wanted to buy the old chipset design so they could make a radiation-hardened version to use on satellites. The major semiconductor company liked the sound of that, and asked Gordon to find the backup. "My colleague and I then pieced together what had happened," he wrote. "Being good engineers, we said nothing. We searched through old nightly backup tapes until we found bits and pieces of different steppings of the design database and copied them all to one place, then handed this off to the customer." To this day, Gordon doesn't know if the customer was able to produce a useful chip from the bits of backup he cobbled together. "Did they realize they received bits and pieces of two slightly different databases? Did they care? Was the project canceled before anyone realized it?" "From then on, I was more careful about who I assigned the task of creating restore backups for each design project," he told The Register. "And I kept a record of the emails requesting backups and confirmation that backups had been done." Have you given a job to the wrong person? Or given clients slightly dodgy data? In either case, click here to send email to Who, Me? We would love to tell your tale on a future Monday. ®

A designer named Marcin Wichary has revealed the story of how he freaked out a decent slab of the world’s Firefox-using population with Pac-Man. Wichary has held senior design roles at Figma, Medium, and Google. As he wrote here, in 2010 Google decided to celebrate the 30th anniversary of Pac-Man by changing its logo at Google.com to a playable version of the game. The Google of 2010 could still mention its motto of the time, “Don’t be evil,” without making anyone who heard it guffaw incredulously. The company’s frequent changes to the logo on its homepage – called Google Doodles, and which persist to this day – were still a notable part of net culture. Wichary got the job of building the playable logo, and tweaking Pac-Man to replace its usual maze with a course that instead navigated around the letters GOOGLE. “It was, frankly, a terrifying proposition,” Wichary wrote. “The homepage was then a very important destination, and my job title was only adjacent to ‘software engineer.’ But I started researching the game and recreating it all in HTML and CSS." Google wanted a full Pac-Man experience, including sound, "because so many of Pac-Man’s sounds are beyond iconic: the starting melody, the waka-wakas, the death jingle." Wichary began work before Google sought permission from Pac-Man’s owner Namco, because the search giant was “planning to wow them with a close-to-finished doodle.” “The project was … going really well. I had fun coding the game and we seemed to have fun playing it. Namco eventually got onboard, too,” he wrote. Google liked the game work so much, higher-ups decided it would be a great idea to have it start automatically if visitors to google.com stayed idle for ten seconds. The designer liked that idea. “If there was anything scary about it at the time, it was only the thought of my code getting even more exposure – but I wasn’t worried about distraction, since the homepage was not a destination page you would ever linger on.” What Google didn’t realise was that some people lingered on the home page without ever knowing they were there. Those people used the Firefox browser with an extension called CoolPreviews that let users hover over a link to see a preview of a web page. “The extension would quietly start at the same time Firefox was opened,” Wichary explained. “And it would immediately, in the background, invisibly and unbeknownst to the user, open a website.” That website was Google’s home page, which now featured an auto-playing and version of Pac-Man that CoolPreviews users simply did not expect. Which explained why, not long after setting the Pac-Man doodle live, Google received plenty of praise, but also reports about computers making strange noises due to CoolPreviews invisibly opening the noisy game. “I pushed a patch within hours,” Wichary wrote, explaining that he added a mute button and a requirement to click something to make the game and its noises start. CoolPreviews couldn’t make those clicks, so the phantom Pac-Man sounds went away. “Ironically, I don’t believe we ever heard of normal people being bothered by the autoplay – it seemed that indeed, few people stayed on the homepage for 10 seconds doing absolutely nothing,” Wichary wrote. “The funny thing is that emergency live-patching [the] internet’s at the time most valuable homepage was perhaps the scariest part of that day for me, but that completely went over my sleep-deprived head,” he added. Google changed Wichary’s title after this incident. His post shows an image describing him as “User experience designer. The Michael Bay of doodles.” ®

As debate and legal argument continue over the extent to which developers of large language models are entitled to feed the world’s data into their systems without paying for it, India’s government has come up with an idea to regulate and secure the flow of information. India’s motive is to create AI models trained on local data, because it thinks offshore model-makers won’t be motivated to address the nation’s issues or to ensure owners of data stay in control. The scheme India’s Ministry of Electronics and Information Technology (MeitY) thinks can make that happen is called the “Data Empowerment and Protection Architecture (DEPA) for AI Training.” As described by Professor Ajay Kumar Sood, Principal Scientific Adviser to the Government of India, DEPA means “the model travels to the data, rather than the data travelling to the model.” The scheme envisions “Training Data Providers” (TDPs) – examples of which include hospitals, banks, and government agencies – sharing anonymized and encrypted data with a “Data Exchange Platform” (DEP). Entities called “Training Data Consumers” (TDCs) access the data stored at the DEPs to train their models. DEPs will operate what MeitY calls “hardware-based Secure Data Environments” in which “data remains encrypted and isolated even during processing.” They also enforce rules set by TDPs that describe how TDCs can use data. MeitY’s DEPA framework [PDF] says this scheme means “Differential privacy techniques are systematically applied to prevent the model from learning or revealing individual specific information,” adding “All activities inside the secure data environments are immutably logged for audit purposes, ensuring that the process is both secure and transparent.” India’s government already operates several data exchanges, but it is unclear how much work remains to make DEPA a reality. Once DEPA is working, MeitY envisions model-makers will use it to do things like “unlock currently fragmented health data to power breakthroughs in AI-driven diagnostics, drug discovery, and public health management.” The optimism extends to financial services. India is home to literally thousands of banks, and MeitY thinks “DEPA can enable secure sharing of diverse financial data, potentially extending formal credit to 400 million underserved individuals and MSMEs and fueling next-generation fintech innovation.” India’s government has long believed that if financial services companies can lend to more people, it will boost economic development. “Similarly, in Education, DEPA can overcome data interoperability challenges to enable AI-powered personalized learning for millions and create portable skill credentials essential for a future-ready workforce of its 250 million students in a $29 billion EdTech sector,” the framework states. “The successful adoption of DEPA Training across these domains is the essential catalyst for realizing India’s vision of inclusive growth and responsible AI leadership on the global stage.” While the purpose of the framework is to build Indian AIs for Indians, the document also has half an eye on the nation’s international standing. “DEPA is India’s contribution to global science diplomacy, advancing the efforts of institutions such as the World Bank and the United Nations to shape a more inclusive and responsible global AI ecosystem,” the framework states. It also makes several mentions of digital public infrastructure (DPI), an idea pioneered by India that involves government funding creation of tools like the Unified Payments Interface payment systems and the Bharat Taxi zero-commission ride-sharing platform, to ensure Big Tech companies don’t dominate the local market. India also uses its DPI as a diplomatic tool, by giving it away to friendly nations. ®

Google has revealed some of its Asian customers – almost certainly Chinese e-tailers – are buying fewer ads. Speaking on Alphabet’s Q3 earnings call last week, chief business officer Philipp Schindler addressed the slowdown in ad sales growth with the following remark: “In retail, some of our APAC advertisers were impacted by the EU eliminating its de minimis exemption.” That’s a reference to the European Union’s changes to import duties, which saw the bloc introduce a temporary €3 customs duty for every consignment valued at under €150 imported from outside the EU. The EU previously operated the de minimis exemption that Schindler referred to, meaning it collected no import duty on almost 5.9 billion low-value items shipped to the continent in 2025 alone. The new regime aims to protect European retailers from cheap imports. A great many of those items came from Chinese e-tail giants – the likes of Temu, Shein, and Alibaba – who are thought to account for perhaps ten percent of advertising revenue at Google and Meta. EU policy has therefore contributed to slower growth at Google. Huawei’s brake pedals reportedly break Huawei’s automotive efforts appear to have driven it into a cul de sac of controversy. The Chinese tech giant works with a company called Anhui Jianghuai Automobile Group Co. (JAC) under the Harmony Intelligent Mobility Alliance, which aims to produce intelligent cars. The flagship model produced by Huawei and JAC is a luxury EV called the Maextro V800, which when tested by Chinese automotive review site Dongchedi produced a rather terrifying fault: the brake pedal bracket broke when drivers exerted maximum force during tests to assess the car's ability to decelerate suddenly at 100 km/h. The site’s drivers were still able to stop the car, but rather more slowly than is optimal. Within days of the review site posting its findings, reports of its work started to disappear from the internet, according to China Digital Times, a US-based outlet that says it publishes independent news on Chinese affairs. Even some opinion pieces that called for further assessment of the car’s safety disappeared, giving rise to theories that Beijing is displeased with negative coverage of its technology champions and automotive industry. North Korea named as Binance hacker The USA’s Federal Bureau of Investigation (FBI) on Saturday named North Korea’s elite cybercrime crew “TraderTraitor” as the source of last week’s attack on cryptocurrency exchange Binance. Cyber-defense company TRM also reached the same conclusion. Its analysts last week declared they have “high confidence” that “on and off-chain intelligence” suggest TraderTraitor did the deed. “On the evening of October 29, billions of dollars in crypto began leaving wallets belonging to Binance,” TRM’s analysts wrote. “Within minutes, Binance wallets on the XRP Ledger, Ethereum, Bitcoin, Tron, Solana and four other networks had sent their balances to addresses controlled by a single attacker. A second wave around 23:05 UTC reached Binance wallets on Avalanche and Polygon.” TRM thinks Trader Traitor started its raid 36 hours before the theft. “The attacker's operator wallet received three small withdrawals from Tornado Cash, a service that obscures where crypto comes from, starting the morning of October 28,” TRM wrote. “Early on October 29, the operator deployed a first piece of code for the attack and tested it on two wallets. Later that day, it rehearsed the final version on a third wallet, then routed the test funds back to itself. Four hours after that rehearsal, the same code was installed on a Binance hot wallet.” Binance first said the cause of the incident was a single compromised laptop. It has since detected “additional points of breach” but has not offered additional information while it investigates the matter. The crypto exchange continues to reassure users that its platform is safe and that it will make good any losses. North Korea steals cryptocurrency to fund its government, especially its belligerent defense efforts. Samsung posts monster profit Samsung last week delivered the monster profits it promised early in October. The Korean giant announced record Q3 revenue of ₩195.3 trillion ($146.2 billion ) and operating profit of ₩107 trillion ($80 billion). Samsung’s DS division, which covers all its chipmaking efforts, earned over 99 percent of that operating profit and $110 billion of the company’s revenue. The company said its foundry business “experienced a notable revenue and earnings improvement on HBM base-die sales growth” and expects it will notch up “double-digit revenue growth and continued earnings improvement on new mobile products, mass production ramp-up of 4nm NPUs, and expanded memory product sales” in its next quarterly report. “In 2027, the Business will aim for revenue growth led by 4nm and 2nm advanced nodes based on strong AI and HPC demand. It also expects to strengthen its global supply capabilities with the commencement of operations at its new fab in Taylor, Texas,” the company said. That last item is welcome news as the reason Samsung and other memory-makers are making so much money is shortages mean they can inflate prices. More factories coming online may help to ease price pressures. Fast-charging batteries set for takeoff Dr Rachid Yazami, the scientist who developed the graphite anode that made lithium-ion batteries mainstream, has shown off a new battery design that can achieve 70 percent charge in under four minutes. In tests of its lithium iron phosphate (LFP) battery cell, Dr Yazami’s startup KVI was able to go from 10 percent to 70 percent charge in 3.7 minutes, and to full charge in seven minutes, using “Non-Linear Voltammetry (NLV)” technology that he says “uses a proprietary, stepwise voltage-controlled approach” that “allows the current to respond to the battery’s changing electrochemical characteristics” during charging. “The challenge is not simply to push more electricity into a battery in less time,” Dr Yazami said. “It is to understand how the battery responds during charging and use that information to deliver energy more effectively.” Japanese company SkyDrive, which hopes to manufacture electric-powered flying vehicles, has signed up to test the tech. ®

Linus Torvalds has posted the first release candidate for version 7.4 of the Linux kernel and offered his opinion that the project has found a “much healthier model” for using AI. “This is a *BIG* merge window,” he wrote in his weekly state of the kernel update, as it contains almost 20,000 commits compared to the usual ~15,000 that developers submit for a new cut of the kernel. “I think the reason is the by now obvious one, but I'm actually taking it as a good sign: I'm *much* happier with the flood of AI triggered work going into the merge window than it showing up mainly as extra fixes during the -rc period.” He therefore hopes that the unusually large number of commits “is a sign that the AI fixes are becoming more part of the ‘normal development’ than a flood of fixes that go into the whole stabilization phase.” “I think that's a much healthier model,” he added, before hedging with the observation that “people are expecting 7.4 to be the next LTS [long-term support] release, which is another independent reason why some releases tend larger than others.” “So it may not *just* be the flood of AI patches.” Torvalds also seems not to be completely convinced that all AI submissions are entirely helpful, by describing this release as “Large, and full of fixes for issues found by AI (and more and more the patches have been suggested by AI too, although they often get rewritten).” AMD GPU register offsets and bit masks make up a third of the changes this time around. “We've been here before, and I'm sure we'll see it many times again in the future,” Torvalds wrote. “Just ignore the noise.” The kernel-watchers at Phoronix rate the major inclusions in Linux 7.4 as improved support for processors from AMD, Apple, and Google. Other additions include some handy tweaks to improve security for virtual machines, support for the Raspberry Pi 10-inch monitor and more work to get Linux running happily on Qualcomm’s Snapdragon X2 SoC for PCs. ®

No lesser authority than the MIT Technology Review says that the Emacs editor has turned 50 years old. But then, it is biased — Emacs was invented in MIT's AI lab. Simson Garfinkel's article is simply called EMACS turns 50, but the sub-heading explains why it matters more: "How the MIT AI Lab's text editor gave birth to the free software movement." There are multiple timelines and accounts of the origins of Emacs, and it is not a single monolithic application. The name is an abbreviation for "Editing MACroS", and the original editor for which this set of macros was developed was TECO – "Text Editor and COrrector," written by Dan Murphy in 1962. Murphy died earlier this year at the age of 83. The inimitable Jamie Zawinski has an encyclopedic Emacs Timeline – as he mentions at the start, his interest being that he long ago maintained a fork of GNU Emacs, originally called Lucid Emacs. It's still around, under its later name XEmacs. There are a lot of forks and re-implementations of Emacs out there, and we are not going to go into all of them, or this article would become a book – one only interesting to programmer-archeologists (to use a term coined by the late great Vernor Vinge). As for who originally wrote those Editing Macros, though, another history has quotations from several of the people involved: On the Origin of Emacs in 1976. (Note – some of the quotes still contain the | symbols that marked quoted text.) This credits Lisp boffin Guy L. Steele Jr with the original work in 1975, but in it, Steele says: "RMS is responsible for the names 'E' and 'EMACS'. RMS still deserves 99% or 99.9% or 99.99% or 99.999% of the credit for taking a package of TECO macros and turning it into the most powerful editor on the planet." RMS is of course Richard Stallman, and the page also quotes some emails from RMS on October 23 and 31, 1976. Mind you, just a few years after the event, in a 1981 talk now preserved on the GNU website, Stallman said, "I began the development of EMACS in 1974 with an improvement to TECO". Like all the best FOSS, even in its very early stages, it was a collaborative effort – but around late October 1976 is when RMS took over the bulk of the development. Of the many forks of Emacs over its half a century of development, some are still just about hanging on in there. When we looked up the link for XEmacs, we were surprised to see that there was a new release: XEmacs 21.5.37 "lettuce" appeared just last month. And even the forks have forks: in this case, SXEmacs, last updated in 2020. Most of the changes in XEmacs and SXEmacs eventually made their way back upstream to GNU Emacs, which is still in active development, although as The Register reported in 2008, RMS is no longer the primary maintainer. GNU Emacs 31.1 appeared in August. Cards on the metaphorical table: this vulture is not an Emacs user. We have tried. A major influence was one of the Reg FOSS desk's favorite writers, Neal Stephenson, who in 1999 published an essay called In the Beginning… Was the Command Line. (It was later republished as a rather short book, but the text is still online.) In it, he says: "In the GNU/Linux world there are two major text editing programs: the minimalist vi (known in some implementations as elvis) and the maximalist emacs. I use emacs, which might be thought of as a thermonuclear word processor. It was created by Richard Stallman; enough said. It is written in Lisp, which is the only computer language that is beautiful. It is colossal, and yet it only edits straight ASCII text files, which is to say, no fonts, no boldface, no underlining. In other words, the engineer-hours that, in the case of Microsoft Word, were devoted to features like mail merge, and the ability to embed feature-length motion pictures in corporate memoranda, were, in the case of emacs, focused with maniacal intensity on the deceptively simple-seeming problem of editing text. If you are a professional writer–i.e., if someone else is getting paid to worry about how your words are formatted and printed–emacs outshines all other editing software in approximately the same way that the noonday sun does the stars." Emacs is 50-plus-year-old software, and it shows. This is a huge and complicated program: decades ago, the joke was that EMACS stood for "Eight Megs And Constantly Swapping". It is not an easy program to learn. Despite lavish documentation and a built-in tutorial, this is software so old that its terminology predates microprocessors and the personal computer. Its exhaustive manual is 804 pages long, and to pick some trivial examples, its terms for parts of the screen use “frame” to mean what’s now called a window, and “window” to mean what’s now called a pane. In the section on Keys, it discusses the Control and Meta keys. That’s a slight snag: you don’t have a Meta key, and neither do the authors of Emacs. It was a key on the now almost legendarily rare Space Cadet Keyboard, of which the Deskthority keyboard wiki says: “It is known as the holy grail of keyboards.” You can buy a modern replica, the Keymacs keyboard. They are hand-made in the Czech Republic, and depending on exactly what model and specifications you want, they cost from €2,998 for the Lispsaver up to €6,878 for the A624N Space Cadet. In other words, you can have a keyboard with a real Meta key if you want, but it is going to cost you. One of the problems with these historical terms, unfamiliar to anyone who wasn’t already programming in the 1970s, is that meanings drift. In the context of Emacs, "Meta" means Alt, or Opt on Macs. However, KDE also uses the Meta key – but what the KDE documentation calls "Meta" is the Windows key – or Command on Macs. Like most Linux folks, the GNOME documentation calls it the Super key. Names are important. They matter. Who's right, the Emacs folks or the KDE folks? Well, neither: nobody has Meta keys any more. We have Ctrl, Alt, and Cmd/Super/Windows. There have been valiant efforts to modernize Emacs and bring it into line with modern UI and terminology. This vulture tried to learn Aquamacs, which modernized Emacs with a native macOS UI, but sadly it's no longer maintained. Another, which we enjoyed trying, is ErgoEmacs, which changes the key bindings and rearranges the UI into something much more like the IBM Common User Access arrangement, as used in Windows and most Linux desktops. However, it never really caught on and you need to know your way around Emacs to install it and configure it for your regional keyboard layout. So complex is it to configure Emacs that there are multiple pre-configured distributions of Emacs, such as Doom Emacs, or the minimalist Emacs Solo. We wish someone did a "CUA-Macs" config, with ErgoEmacs integrated and able to pick up your existing keyboard config. Come to that, the Gershwin Desktop project that we mentioned along with GhostBSD last year is in active development, and it's based around GNUstep. Maybe its developers could pick up Aquamacs, and adapt it to GNUstep. There is a long history of such Emacs facelifts. One instance is Borland’s Sprint word-processor, one of the selling points of which was that it could emulate the user interfaces of most of the leading MS-DOS word-processors of its day (just before they all converged on IBM CUA). Before Borland acquired it, Sprint was called The Final Word, developed by musical equipment makers Mark of the Unicorn. FinalWord was based on an Emacs clone called MINCE. The Great Rival Another well-known Emacs distribution is Spacemacs, which adapts Emacs to have a more Vi-like user interface. Vi is the other seminal Unix text editor, as the Stephenson quote above mentions. As this article on Origins and the Evolution of Vi & Vim notes, it too dates back to 1976. Indeed, that’s the reason we chose the year in this story: Tilde, a text editor that doesn’t work like it’s 1976. One feature of Tilde we enjoy is that it supports use of the mouse even on the Linux console – which is the headline new feature of GNU Emacs 31.1. Vi is very much alive and well today. GNU Emacs is based on the original source code, but Vi mainly survives as re-implementations. The recreation Stephenson mentions, Elvis, lives on today in the form of nvi. However, by far the best known is Vim, maintained by Bram Moolenaar until his death in 2023. In the grand FOSS forking tradition, it too has forks, including Neovim, as used in Omarchy, and more recently, the AI-free Vim Classic – a new project led by Drew DeVault, who just launched the similarly LLM-free AlpineBSD. ®

Openreach has issued a stark reminder to customers that the legacy phone network is being officially switched off on January 31, and says the deadline will not be postponed. The infrastructure arm of BT Group, Britain’s biggest telecoms biz, circulated the warning this week in a letter to comms providers that use its network to deliver their services. It informs them that any customers still on legacy public switched telephone network (PSTN) connections are at risk of disruption when these are retired at the end of January, 2027. The switch-off could affect businesses and consumers alike. Beyond the plain old telephone, systems that rely on PSTN connectivity include point-of-sale terminals, fire and security alarms, and personal care alarms for elderly and vulnerable people. The move is part of a wider government scheme to get the UK onto all-digital high-speed networks running on fiber-optic infrastructure, and the legacy phone service - with its dependency on copper cabling - was standing in the way. Many households have now ditched landline handsets and instead use only mobile phones, yet for those who still want a landline, the changeover means plugging the phone into the back of a broadband router to access a Digital Voice service. Openreach has been planning this move for the best part of a decade. It has already pushed back the deadline at least once, after the government asked telecoms providers not to force vulnerable customers who use phone-network systems such as TeleCare to switch until fixes are in place. Nevertheless, public awareness seems low that analog phone services will soon be disconnected. "Openreach and all providers need to coordinate a much more cohesive campaign that reaches customers directly," said telecoms analyst Paolo Pescatore at PP Foresight. "TV, billboards and newspaper advertising should reinforce a consistent message, supported by targeted letters, calls and door-to-door support for older and vulnerable people who need extra help," he told The Register. "Customers need to understand what is changing, what action to take, and who will help them. Providers must take responsibility for ensuring people remain safely connected, especially those who rely on their landline or a care alarm." Caroline Abrahams, charity director at Age UK, agreed. "With only three months to go, the digital landline switchover is entering a critical final phase. Most people have moved over without major problems, but nobody who relies on their landline or TeleCare should be left behind," she told us. "Anyone with concerns should contact their provider, but the responsibility cannot rest with customers alone. Providers must not give up on people who have not responded. They need to keep trying to reach them and offer practical, individual support, recognizing that some may be unwell, living with dementia, worried about scams or without anyone nearby to help." Openreach has a webpage with information regarding the switch-off. This advises consumers to call their service provider as soon as possible to ensure service is not interrupted. The company previously told The Register that customers without a fiber connection will likely be switched to SOTAP for Analog (Single Order Transitional Access Product), where equipment is installed in the local telephone exchange to allow them to continue to use their old phone line as before. Openreach has warned that commercial customers delaying migration risk significant disruption and unnecessary costs - leaving it until the last minute could make it harder to secure the requisite engineering support before the deadline hits. "Businesses and organizations that delay action are making things harder for themselves," said James Lilley, Openreach Director of All-IP. "The straightforward migrations have largely been completed. What's left are the complex connections that underpin critical operations, requiring careful planning, supplier coordination, and thorough testing. We are urging businesses to act now, audit their remaining analog services, and work with their communications provider on a migration plan today." Earlier this year, the firm tried to drive home the message by announcing it was stepping up its charges to service providers and companies still operating legacy business lines, who will by now be paying double the rental cost compared to last year. ®

Google on Friday celebrated receiving a poem from its first experimental orbital AI datacenter satellite. The company's Suncatcher satellite, launched at the start of the month and put into low Earth orbit, has responded to a ground station prompt to "Write a haiku from the perspective of a TPU orbiting Earth, with the wonderful world spinning beneath you." Travis Beals, senior director of Google's interdisciplinary Paradigms of Intelligence research team, said that with Suncatcher's host and TPU trays powered up and machine learning tests completed, the onboard Gemma 4 model achieved its "Hello World" milestone by sending an inference response from orbit. "During a ground-station pass, we uplinked our first prompt to our onboard Gemma 4 open model, asking it to generate a haiku about being in space," wrote Beals in a blog post. And as the satellite passed over Svalbard, Norway, the orbiting AI model responded over the downlink: Blue marble below, Data streams in silent grace, Starlight feeds my core. ✨ Those roughly 20 tokens would cost about $0.000012 if generated through the Google Cloud API, based on Gemma 4 26B's pricing of $0.15 per million input tokens and $0.60 per million output tokens, excluding input costs. Add in a satellite weighing perhaps 500 kg, based on the mass of SpaceX's Starlink satellites, and a retail launch price of around $7,000 per kg, and you have a very expensive haiku. According to Google's calculations, if the cost of getting a satellite into low Earth orbit can be brought down to $200/kg, the launch cost amortized over the satellite's useful life (~$810–7,500/kW/year) could be comparable to current electricity spending for terrestrial datacenters in the US (~$570–3,000/kW/year). Google predicts that SpaceX could achieve that by 2035, more or less, and considers it strategically beneficial to prototype well in advance of that actually happening. After the haiku, Beals said, "we tested space-themed limericks and telemetry prompts reporting the technical status of the payload systems." Next up: tests for evaluating how TPUs perform under thermal stress and radiation. Google has already done some laboratory testing of its Trillium Cloud TPU and AMD host server to see how well they can withstand a 67 MeV proton beam. The system's high-bandwidth memory shows the most sensitivity in terms of the total ionizing dose – it exhibits irregularities after a cumulative dose of 2 krad(Si), which is about 3× Google's minimum requirement for TID survival of 750 rad(Si). "For typical transformer workloads, the observed SDC [silent data corruption] rate of approximately one event/17 rad, at an estimated orbital dose of 150 rad(Si)/year, translates to a very low failure probability on the order of 1 per 3 million inferences, assuming 1 inference per second," Google researchers said in a paper published earlier this month. "While this error rate is likely acceptable for inference, the impact of SEEs on training jobs, and the efficacy of system-level mitigations, requires further study." Indeed, the study will continue. Next year, the plan is to test a pair of interconnected satellites. Rather than lofting large satellites into orbit, Google anticipates sending a constellation of many smaller satellites that communicate with one another. ®

Microsoft senior software engineer Raymond Chen has explained what Windows is doing when you hold down the power button, and shared a story about the time a customer accidentally created a Goth Mode for Windows 3.1. There's a section in The Hitchhiker's Guide to the Galaxy (actually, page 116 of this writer's dog-eared copy of The Restaurant at the End of the Universe) in which Zaphod Beeblebrox observes of a ship's interior: "Every time you operate one of these weird black controls that are labeled in black on black background, a little black light lights up black to let you know you've done it." The same turned out to be true of a user who called Microsoft customer support to complain that after Windows 3.1 booted up, all that remained after the product logo disappeared was a blank screen. It turned out the customer liked black a little too much and, according to Chen, made themself a Goth theme. The customer support engineer on Microsoft's end pieced it together. The customer had, according to Chen, "set the wallpaper color to black. They set the title bar color to black. They set the button color to black. They set the window background color to black. They set the window text color to black. "Every possible color you could pick in the Colors control panel was set to black." We suspect the mouse pointer, at least, would have stuck around, unless the customer had also managed to change some settings around this. And then there are the icons… However, Windows 3.1 was a long time ago, and this writer's attempts at theming then were the consequences of growing up in the '70s and '80s, which resulted in some interesting neon, orange, and brown combinations. As for Goth mode, Chen said, "Once you get yourself into this situation, it’s hard to get out because you can't see anything." One solution he suggested was deleting the [colors] section from the win.ini file. Chen told the anecdote as a coda to the question – what exactly is modern Windows doing when you hold down the power button on a PC? According to Chen, after eight seconds of holding the power button, the Windows kernel realizes a forced power-off is happening and tries to take a kernel dump for diagnostics. He explained that the kernel dumps were classified by Microsoft as "long power button hold" (LPBH) failures, with one of the bits of information captured indicating if all the pixels on the screen were black. The video driver might therefore be the culprit. Or whatever was responsible for drawing to the screen. Today's Windows struggles with dark mode. But in the dim, distant past, every pixel being black might mean the user was either Hotblack Desiato or had done something very silly indeed. ®

There's a ridiculously simple test you can use to figure out whether AI mandates at a given company are well-considered, or are being shoved out the door by out-of-touch buffoons: What happens when an engineer documents that AI made a task worse? Are they taken seriously, or are they marginalized and eventually managed out? Conveniently, this week Amazon published a treatise by Terk Terkowitz, a VP/Distinguished Engineer/Final Boss of Their Independent Contributor Technical Career Track, on the role of the developer in the age of AI. It was originally an internal missive sent to their software engineers and their managers back in September, which is an interesting lag for a document that sounds this confident. It also maps exceedingly well to what other organizations are struggling with. I think it's a good document, but I can't help but view it with a somewhat cynical eye. I see it as an attempt to make the future of the profession of software engineering feel less bleak. And it's feeling mighty bleak for a lot of folks right now. What's the context? This treatise comes at a time of dramatic upheaval. Tech as an industry is no stranger to change (usually at 3AM, heralded by the pager), but something about the current AI-ification of the world strikes many of us as fundamentally dissimilar from previous revolutions. Folks are scared about their career prospects. "Should I give up tech and get into farming?" is a sincere question multiple people have asked me. I don't know! Tech is already a goat rodeo, so you've got a leg up if you do? This anxiety isn't theoretical. In May, Amazon wisely shut down KiroRank, an employee-built cash incineration leaderboard that ranked people by token usage, after the obvious thing happened: people spun up junk tasks to climb the ranks. This is now called "tokenmaxxing," and the Terkowitz treatise specifically calls it out as being at odds with "frugality." Meta did the same thing with "Claudenomics" and 85K employees that same week. And Amazon reportedly announced another 250 job cuts this week. I can't tell you that these things are related. But I can tell you that every engineer who read the memo assumes that they are. What's the thing say? This document contains a lot of great stuff. I have a tremendous level of respect, both personal and professional, for Amazon's Principal Engineering community; I'm led to understand that that respect may not be reciprocated, but that's okay. I am, by their standards, a lapsed engineer who ran away to join the circus as one of its lesser clowns. Terkowitz's memo emphasizes customer outcomes and engineering judgement, two things I would absolutely not delegate to AI any time soon. (Note: I am writing this in October of 2026, with the full understanding that previous comments on the capabilities where AI falls short tend to age like fine milk.) It documents work that (today) AI tools aren't improving. That's way better than assuming that every objection to AI inherently comes from some self-interested Luddite who's scared of being replaced. It also says something smart about measurement: look at the 90th percentile of AI adoption, at those teams burning the most tokens, and check whether their delivery actually went up commensurately. This is great! Shout it from the rooftops; we should all be doing such things. But still, I wonder. The memo says “If your work does not benefit from current AI tooling, that is a finding, not a failing.” Document it, bring it to your manager, work through the problem. I agree with every word of that! The question is whether anyone's promo doc is gonna survive citing it. Here's what happens next. The memo concedes that input metrics are "deeply flawed," but then in the next paragraph says to go ahead and use them anyway, carefully. It then warns that non-adopters are falling behind. Those are three different positions within three sentences and they're not at all aligned with one another. That's not muddled; I think it's cover. A manager can cite "a finding, not a failing" when a strong engineer skips the AI tool, and then cite "falling behind" when it's time to PIP someone. Both are congruent with their stated policy; you get to decide which interpretation you want to go with in a given situation. How will managers handle justified AI non-use? Are promotion and performance decisions being made in a way that rewards these long term outcomes? Or are they instead taking the easy weasel way out, optimizing for what's cheap to measure, and rewarding tool activity? Because if you give me a metric, I will turn it into a target. So will an overwhelming majority of employees. When that happens, everyone loses — it's how KiroRank happened in the first place. Everyone's selling tokens — including Amazon I write this after two days of attending GitHub Universe, where they talked about how agents start roughly 40% of PRs now. And yet, Jay Parikh’s keynote argument was that useful output wasn't "more code," but rather completed outcomes: features shipped, bugs fixed, security issues addressed, loudmouth columnists silenced, etc. The problem here is that their entire business model cuts against that; they're billing in a model that charges customers on a consumption basis, not an output basis. There's nothing, today, that ties "AI spend" to "output that was worth the effort." Which brings us back to Amazon. The memo says tokenmaxxing is at odds with frugality, and for Amazon’s internal cost line, it is. Meanwhile AWS bills Bedrock and Kiro customers by the token. Inside the building, tokenmaxxing is a problem to be stamped out. Outside it, it’s the revenue plan. ®

Today's bit of signage silage is either Windows 10 protesting that it is not part of London's Oxford Street pedestrianization, or someone failing to heed an alert before the public is made all too aware that things aren't going well… Eagle-eyed Register reader Aaron Colwill spotted the sign outside the food hall entrance of a Marks & Spencer store on London's Oxford Street. It usually entices passersby through the doors with promises of treats inside, but today it appears to require a different kind of feeding. Disk space is running low. Marks and Spencer (also known as M&S and Marks and Sparks) is no stranger to cyber whoopsies, having been the victim of an attack in 2025 in which crooks made off with customer data. Later that year, it reckoned cleanup costs would be about £136 million. That's quite a lot, but not enough to buy a bigger hard disk for Windows 10, it seems. Storage prices have increased over the years, but surely not that much? Colwill reckoned that Windows 10 21H2 LTSC was behind the problems. It could well be. While support for 21H2 ended for many users in 2023, Windows 10 LTSC 2021 will carry on until Jan 12, 2027, meaning engineers might have an exciting festive period ahead of them. Windows 10 IoT Enterprise LTSC 2021 will receive extended support until Jan 13, 2032. Either way, using Windows to advertise the Food Hall's wares seems slightly overkill. Simple posters, for example, never run out of disk space or attract the bork fairy's attention. Perhaps they might use something that lacks the voracious appetite for storage that Microsoft's finest demands. According to the Windows biz, installing Windows 10 requires at least 32 GB of disk space, ideally more. It all seems a bit excessive for a cake-based caterpillar advert. Then again, when your operating system tends to have the resilience of a chocolate teapot, using it to advertise a chocolate insect might not be such a bad idea after all. ®

Nvidia is more than just a leader in graphics and AI hardware. It has also become a titan of the networking industry with bigger datacenter Ethernet sales than any other vendor. In Q2, Nvidia's networking revenues topped $17.5 billion. To put that in perspective, Cisco's revenue during that same period, across all products, was $17.3 billion. GPUzilla's arrival as a networking power feels sudden, yet Jensen Huang started working towards this status in 2019 with the $6.9 billion acquisition of long-time networking partner Mellanox . At the time, Mellanox wasn't a big name in Ethernet switching. The company was mostly known for its ConnectX NICs and InfiniBand switches. Outside of high-performance computing (HPC) and supercomputing clusters, and a big deployment at Microsoft, InfiniBand wasn't widely deployed in hyperscale datacenters. Buying Mellanox allowed Nvidia to reposition itself as a system builder, which means something different in the context of distributed compute. Nvidia understood its datacenter customers weren't interested in buying GPUs or even servers; they wanted clusters of systems. For that, Nvidia needed interconnect and Mellanox's InfiniBand fit the bill. Relentless co-design Over the years, Huang has talked about co-design – designing components and software to work together for specific set of jobs rather than to be the best GPU, switch, or NIC. Its work with Mellanox is a prime example of that approach. According to Dell'Oro Group analyst Sameh Boujelbene, Nvidia's decision to integrate compute and networking became key to the company's success. "Customers choose Nvidia for its tightly integrated compute and networking architecture, optimized for AI workloads," she told The Register. That integration makes deployments easier and speeds time to market. Ben Mabey, CTO of pioneering AI and machine learning company Recursion, told The Reg Nvidia's integration schemes won him over. Founded in 2013, nearly a decade before ChatGPT made AI the subject of dinner-table discussion, Recursion sought to apply machine learning running atop GPUs to the problem of drug discovery. This work involved churning through massive datasets which, Mabey tells us, now total more than 50 petabytes in size. "When you think about our workloads, it makes LLMs workload training seem easy compared to the amount of data that has to be moved around," he told The Register. "In the biological space and sciences in general, it's not well appreciated just how big the data actually is," he added. "The internet's big, but in the grand scheme of things, the text that people are training on is actually pretty small, versus the petabytes of data that we have of these images." Recursion's first cluster was a system called BioHive Zero which it cobbled together from GTX 1080 graphics cards originally intended for gaming systems. It used Ethernet to connect GPUs and ran into the limits of conventional networking. So, for its second cluster, an A100-based machine with around 300 accelerators, Recursion went with Nvidia's Quantum InfiniBand switches and ConnectX NICs. As Mabey explains it, buying its compute and interconnects from the same vendor vastly simplified optimization and troubleshooting. When Recursion reached out to Nvidia for help tuning the system, they were able to consider it as a single system, rather than discrete compute and networking resources. This level of integration had only previously been offered by boutique supercomputer builders like Cray or Atos (Eviden/Bull/whatever it gets renamed next). A wider Spectrum Nvidia hasn't always listed networking revenue as a single line item in its financials. However, based on CFO commentary and a little back of the napkin math we can fill in the blanks. Nvidia's networking revenues have grown steadily since the Mellanox acquisition. At two moments, its revenue shot up suddenly. The first came in the 2023 calendar year (FY24) after OpenAI kicked off the AI gold rush with the launch of ChatGPT. The second spread across 2024 and 2025, when Nvidia started selling Blackwell-generation GPUs and then started to ship them in large quantities. "I see two major inflection points, both coinciding with the Blackwell platform ramp: the introduction of NVLink switches and the accelerating adoption of Spectrum-X Ethernet switches for AI back-end networks, driven by growing deployments of Nvidia GPU clusters and increased attach rates of Nvidia switches, NICs, optics, and cables," Boujelbene said. It was around this time that users started deploy to Ethernet switches in large volume across Nvidia clusters. Elon Musk's Colossus supercomputer was one of the earliest examples. As we reported at the time, the first half of the 200,000-chip H100 and H200-based cluster came online in 2024 and, at the time, was among the largest ever clusters deployed using Nvidia's Spectrum-X Ethernet switches. Meta has also made extensive use of Nvidia's Spectrum switches in its own training clusters, including at its million-GPU Prometheus site in New Albany, Ohio. There are a couple reasons Nvidia's Spectrum-X offering is doing so well. First, through a combination of Spectrum Ethernet switches, superNICs, and some clever software optimizations, Nvidia was able to replicate many of the qualities that made InfiniBand so attractive while also benefiting from significantly higher radix and interoperability. Ethernet switches, including Nvidia's own, have historically offered more ports. The higher the radix, the fewer switches are necessary to support a cluster at a given port speed. This often gets misconstrued to mean that InfiniBand can't scale beyond clusters of a certain size, which isn't necessarily the case. But for clusters exceeding a certain number of GPUs, Ethernet is often more economical because it requires less equipment while achieving similar utilization. That said, Nvidia's Spectrum-X isn't exactly unique. Broadcom, Cisco, AMD and others have managed to achieve utilization similar to InfiniBand through various investments in smartNICs and switches. Nvidia's main differentiator, as Mabey and Boujelbene noted earlier, is the level of integration between the compute and networking. The next frontier Then Blackwell hit the market, enabling much greater scale and creating a surge in Nvidia's networking revenue. Prior to Blackwell, Nvidia's NVLink and NVSwitch interconnect tech had largely been limited to single systems with up to eight GPUs. With the launch of the GB200 NVL72, these NVSwitches were broken out into dedicated appliances. (And yes, we remember GH200 NVL256 and NVL32, but from what we can tell these were never shipped in any meaningful volume, if at all.) For every NVL72 rack system sold, Nvidia also sold nine NVSwitch trays, which enabled the system's six dozen GPUs to function as a single enormous GPU. This effectively guaranteed steady networking revenue growth: even if customers didn't buy Nvidia's scale-out switches, if they wanted NVL72 racks, they were automatically buying NVSwitch. "We estimate that NVSwitch now accounts for roughly one-third of Nvidia’s networking revenue, up from virtually zero less than two years ago," Boujelbene told El Reg. "This underscores how quickly NVLink switching has become a major growth engine for Nvidia’s networking business." Nvidia is now moving to grow its networking business even further with a new licensing program. Until recently, Nvidia's NVSwitch tech was limited to its own accelerators. However, with the introduction of NVLink Fusion early last year, Nvidia opened the interconnect tech up to third parties to license. Many chip designers, including Marvell, Qualcomm, Arm, Fujitsu, MediaTek, and d-Matrix, have announced plans to develop CPUs and XPUs using NVLink Fusion. The adoption of the interconnect opens the door to additional NVSwitch sales, as the chips using NVLink Fusion still rely on them for chip-to-chip communications, and there's no other place to get them, as Nvidia has yet to license out NVSwitch. "NVLink Fusion allows Nvidia to extend its networking footprint beyond its own GPUs into the rapidly growing custom accelerator market. I see it as a strategic move to capture a larger share of AI networking spending, even in clusters that do not use Nvidia GPUs," Boujelbene said. Nvidia's networking ambitions don't stop there. "Nvidia continues to pursue opportunities in emerging networking domains as network requirements evolve to keep pace with increasingly demanding AI workloads," Boujelbene added. "The latest example is scale-in networking, designed to address the new requirements of agentic AI, while innovations such as CPO will help overcome growing bandwidth, power, and scalability challenges." ®

Declining token prices have prompted Darkbloom, a distributed inference network, to reassess its business plans. Darkbloom, a project from Seattle-based tech biz Eigen Labs, has developed software that allows anyone with a sufficiently capable macOS device to rent their GPU through the Darkbloom network for interference jobs. In return, they get paid. Through this network, 800 or so participants – including as of two days ago this reporter – have been able to collect a few cents to a few dollars per day by having their hardware handle AI model prompts submitted through OpenRouter, an AI model gateway. In theory, someone running one of Apple's recently released Mac minis with an M5 Pro chip and 64 GB of RAM could earn as much as $192 per month handling incoming inference jobs, per the company's estimate. That assumes full 24/7 utilization during this period on a sought-after model like Qwen3.6 35B A3B, which sells for $0.065/M input tokens, $0.75/M output tokens. But not every network participant has fully surrendered their Mac hardware to the network. The median earnings of the 50 people on the leaderboard amount to about $5.26 per day, $159.86 per month, or $1,918.50 per year. So the best case estimate presented by Darkbloom looks plausible. But it may not be sustainable. The price of tokens – the unit of AI work – has declined, reportedly down 54 percent since May. Ramp's AI Token Price Index, which tracks the price of tokens from Anthropic and OpenAI, shows similar slippage. For Anthropic, the price per million tokens went from $2.65 on May 1, 2026 to $0.62 on October 22, 2026, a decline of 77 percent. In a message shared with Darkbloom participants on the company's Slack channel, Kydo, one of the Eigen Labs researchers working on the distributed inference network, celebrated the network's processing of its trillionth token while attempting to moderate expectations about the project's viability. "Earnings have been low," said Kydo. "For most providers, I don't expect that it is going to recover soon." Pointing to the computational prowess of DeekSeek's DeepInfer framework, he explained that structural changes in the inference market have affected the supply of tokens. "Data centers have become cheaper," he said. "Output per NVIDIA cluster is up ~20x this year, and their margins are wide enough to cut prices 90 percent and still profit at 50 percent." The flood of tokens in turn has washed away the price advantage Darkbloom had previously. "The market caught up to our price," Kydo explained. "We could go lower, but then most of you would earn even less nothing." Finally, he said, the inference routing market is changing, with OpenRouter moving away from price-based routing and putting more emphasis on decode speed, quantization, and prefill speed. For network participants with older or less powerful Mac hardware – those without top of the line Max or Ultra Apple Silicon – "inference won't be worth running," he said. Kydo went on to outline how Darkbloom aims to adapt to the commoditized token market. This involves analyzing model demand to steer participants toward more sought-after models, and looking at ways to provide other kinds of compute work, like distributed CI/CD jobs. For Mac hardware, the focus now appears to be on clusters of high-end Mac Studios (now available with up to 512 GB of RAM), using Apple's Remote Direct Memory Access (RDMA) over Thunderbolt. The CI/CD work, based on the observation that AI agents now make pull requests faster than CI/CD pipelines can handle them, will require sandboxing — something the biz expects to begin testing soon. And Darkbloom also aims to improve the performance of its network (in terms of time to first token) and to look for other token buyers beyond OpenRouter. "The next stretch will be a fight, and I expect many machines will leave and Slack will get quieter," Kydo concluded. "But those who stay will be first in line when new earning opportunities come along. We are a driven team and well capitalized. We will not give up." ®

When you're ordering food from your phone, even time spent navigating the app is time wasted. At the P99 2026 virtual conference, Elvis Chidera, a staff engineer at Germany-based global delivery service Delivery Hero, explained how the company moved most of its user base to HTTP/3, and consequently cut its apps’ total end-to-end response time of the home screen by 17 percent. Chidera, who managed the project, explained that, as promised on the tin, HTTP/3 delivers the web faster than HTTP/2, though it has slow bits that need to be adjusted to work properly. Uber eats Delivery Hero Delivery Hero is all but unknown in the US, which is dominated by Uber Eats and DoorDash. But it is a company with global reach outside the US, with an estimated user base of 60 million across Europe, Latin America, Africa, the Middle East, and Asia. Uber is currently acquiring Delivery Hero for $14.8 billion, with plans to unite the company with Uber Eats. With this purchase, Uber would expand the number of markets where it offers both ride-hailing and delivery services, though not necessarily under the Uber Eats nameplate. Delivery Hero centrally manages separate apps for different regions, all of which it built or acquired. Asians can order poke bowls and sundries through Foodpanda. Turks can order kebabs through Yemeksepeti. Throughout Spain, Italy, Poland, Romania, Ukraine, and much of Africa, the Glovo app is where to ring up the late-night deliveries. And in Latin America, users dial in to PedidosYa to get what they need from nearby farmacias, restaurantes, and mercados. All these regions have widely varying network connectivity, a problem Delivery Hero wanted to minimize. Last mile networking was most often the bottleneck. Apps in Pakistan, for instance, could take nearly a second (at p90) to establish a secure handshake with a server. The engineering team looked at HTTP/3 as a way to lower, even eliminate, that latency. A QUIC-er HTTP The Internet Engineering Task Force (IETF) formally introduced the third version of the internet’s Hypertext Transfer Protocol in 2022. HTTP is the application layer protocol that connects browsers and other resources connected to servers over the Internet. The previous versions of HTTP were built on the Transmission Control Protocol (TCP), which offered guaranteed delivery and flow control. In conjunction with the Internet Protocol (IP), TCP/IP built the modern internet. HTTP/3 replaced TCP with Quick UDP Internet Connection (QUIC), developed by Google, to address some of TCP’s perceived sluggishness as a transfer protocol. UDP stands for User Datagram Protocol and, unlike TCP, it does not guarantee packet delivery or flow control. UDP is faster because it doesn’t wait around for a signature (QUIC adds in its own flow control on top of UDP). But the UDP-based HTTP/3 reduces, and sometimes eliminates, the round trips needed to secure connections. The app can cache a session ticket and, on a later connection, use QUIC's 0-RTT capability to start sending eligible application data while the handshake is still underway, rather than waiting for it to finish. “Essentially, you can eliminate that one second” of latency that plagued slower networks, Chidera said. “You don't want to pay the cost of two round trips just to exchange data, if you could do it faster.” And unlike HTTP/2, HTTP/3 doesn’t get hung up when packets are lost or delayed. There’s no head-of-line blocking because packets are sent in multiple streams, not just the single string that HTTP/2 offers. Multitask during initialization Setting up HTTP/3 turned out to be surprisingly easy, Chidera found. Delivery Hero used Cloudflare’s content delivery network to serve its global user base, so moving to HTTP/3 involved a simple reconfiguration. Adding HTTP/3 support to the clients, mostly mobile, proved almost as straightforward. Apple has supported HTTP/3 since iOS 15 and most native browsers already support HTTP/3 out of the box. Android proved a bit trickier. The engineers used Google’s Cronet library for the networking stack, initially wired into their Android apps with the Cronet Transport for OkHttp wrapper to minimize rewriting their existing code, originally written for the OkHttp HTTP client. Out of the box, Cronet helped support the HTTP/3 rollout, which reduced network-level request failures. But it also came with a significant catch: the engine took approximately 800 milliseconds to initialize at P90, eating into some of HTTP/3's 0-RTT speed benefits. So the team forked the code into their own bridge using a library called OKCronet. They also did some additional engineering work, which involved overlapping Cronet initialization with other background tasks. Overall, they slashed the P95 latency time from 1.2 seconds to 120 milliseconds, achieving a 10x improvement in read latency. Choose your congestion control algo wisely Overall, the engineering team moved 95 percent of the users over to HTTP/3. Those devices that couldn’t use the protocol (largely due to networks blocking UDP traffic) automatically fell back to HTTP/2. Handshake latency between the apps and the server was cut by 51 percent. Pakistan’s response time was cut in half. One hurdle remained, however. Secondary screens on some of the apps (pages/screens that aren’t the home page) experienced latency issues. These issues only happened in certain markets that were known to be prone to packet loss, such as Pakistan. Delivery Hero’s engineers attributed the issue to Cloudflare’s choice of the congestion control algorithm for HTTP/3, Cubic, which historically has not handled packet loss well. The engineering team tested some of its traffic with Google HTTP/3-aware servers, which use another congestion protocol, Bottleneck Bandwidth and Round-trip propagation time (BBR), and did not get hung up on the packet loss. Delivery Hero is currently working with Cloudflare to resolve this specific congestion algorithm issue, Chidera said. Sometimes that fresh, delicious lahmacun just won’t get to your door as fast as it possibly can, despite all best efforts. ®

A n0n ransomware operator tried to break into 23 organizations across 14 countries, but couldn't secure its own front door, leaving its server exposed with all of its methods on view. The criminals used a mix of new (AI) and old (stolen credentials) tooling during their data-theft spree, and in some cases not only stole and encrypted files, but also deleted data and backups, eliminating victims’ ability to recover their information. And despite the fear and hype around the breachpocalypse, with AI finding and developing exploits for critical vulnerabilities affecting most organizations across the globe, these crooks did not exploit a single bug to gain entry into any of their victims’ IT environments. “In every documented compromise in this investigation, access relied on weak or stolen credentials,” said Ensar Seker, chief information security officer of threat intelligence firm SOCRadar. “What stands out is the combination of AI-assisted development and very conventional entry points,” Seker told The Register. “We found no evidence of vulnerability exploitation in the cases examined.” The gang breaks into organizations at scale using widespread Remote Desktop Protocol (RDP) scanning and automated FortiGate SSL-VPN credential stuffing. To access US-based IT managed services provider ChiBiTek, the crew used a 15-thread residential SOCKS5 proxy pool to test credentials (ip:port:user:password) against FortiGate portals (GET /remote/login). The operators also gained entry using exposed cloud credentials and reused hosting portal sessions. In one case linked to e-commerce site Fanatics, SOCRadar found that n0n used an exposed AWS access key with permissions to list, read, and delete objects in S3 buckets. The criminals then used this key to pull data from 40 S3 buckets, with up to 110 GB per bucket, and ultimately deleted data, including bank-transaction archives, and purged keys across 12 sensitive buckets. AI-assisted extortion While the threat hunters didn’t find any evidence of a fully autonomous attack system, the ransomware operator did use AI to assist with these intrusions, incurring more than $7,216 in AI usage costs by October 22. Specifically: the criminal or criminals used OpenCode, an open-source AI coding agent, configured with 154 model entries across seven providers. Models included Gemini, Mistral, Nemotron, Llama, Claude, GPT-5.x, Grok, Kimi, GLM, DeepSeek, MiniMax, Qwen, and many, many others, some with safety training removed and others marketed as “uncensored” models. “The operator used AI to write and debug hundreds of scripts. That shows how AI can help an attacker turn basic access into a broader, tailored operation,” Seker said. “The destructive behavior is equally significant,” he added. “The operator did more than steal and encrypt files: in some cases, it deleted data and backups, and in one case changed firewall policies to block traffic. That creates a recovery problem that a decryption key alone cannot solve.” SOCRadar’s threat hunters found the operator’s workspace - a publicly accessible Python HTTP file server hosted in Sweden at 185.157.161[.]148 on port 8877 - last month, and documented the exposed directory in snapshots on September 25 and 26. “Across those snapshots, we identified 1,719 files, including attack scripts, ransomware source code, extortion infrastructure, and an AI coding agent environment,” Seker said. “This gave us visibility into the operation behind the victim listings, including how the operator prepared attacks and handled stolen data.” Meet n0n It also provided a close look at n0n, a newcomer to the data-theft and extortion game that has claimed 108 victims since it emerged last month. On October 12, the group added 38 maritime, port and logistics organizations across 18 countries to its leak site in a single batch. Researchers have linked at least 21 of them to customers of South Korean maritime software provider CyberLogitec, although the software provider has not confirmed an intrusion. The crooks set an October 19 deadline for the maritime firms, with ransom demands ranging from $400,000 to $28 million, payable in Bitcoin. After the deadline, however, none of the 38 listings were marked “settled” and none of their payment addresses showed any transaction. All affected organizations’ stolen data was published. While a handful of the gang’s claimed victims have been marked “honored - data destroyed” on n0n’s leak site, indicating the victim orgs paid the ransom demand, SOCRadar hasn’t validated any transactions on-chain. In its profile details, the security shop rates the gang a 10 out of 10 for "aggressiveness" and a zero out of 10 for “lethality.” The group is believed to be based in Russia, although the server uncovered by SOCRadar, which it attributed to a single operator or small team, didn’t provide any definitive geographic or national attribution. Every case that the exposed server documented followed the same lifecycle, moving from initial access to establishing persistence and credential collection via hidden admin accounts disguised as services, covert proxies, credential and session-token theft, mailboxes and emailed multi-factor authentication codes. The crooks then collected sensitive data and sent everything to the operator’s central storage with the attack impact shifting between data-theft, encryption, destruction, and network disruption, depending on the targeted organization. After breaking into Transcom, a Netherlands-based financial services provider, the crims sent a secondary ransom alert directly to PayPal's security team, even though the compromised data was only on Transcom's network. Stupid mistakes and n0n security basics Finally, the crims send ransom notes, post a countdown clock on the leak site, establish negotiation rooms and emails for victims - and then leak the stolen files or destroy backups when the organizations don’t pay up. In some cases, and to apply increased pressure on victimized companies, n0n also broadcasts automated countdown warnings to businesses’ service providers, clients, and media outlets. Despite its aggressive tactics, the operator in this case - as with other stupid criminals The Register has gleefully documented in earlier stories - still struggles with basic security hygiene. “There is a striking gap between operational ambition and security discipline,” Seker said. “The operator invested thousands of dollars in AI services while leaving its own workspace exposed. The locker also contained hardcoded encryption keys, with evidence suggesting reuse across victims. An operation can have serious destructive capability while still making fundamental mistakes.”®

General Motors is taking what looks like a light slap on the wrist from a coalition of 35 states and the District of Columbia after agreeing to settle allegations that it collected drivers' data without proper consent and sold it to data brokers that supplied insurance companies. The Washington state AG's office said on Thursday that GM had reached a $220 million settlement with 36 attorneys general over the automaker's collection and sharing of driving data, including precise location information, speed, seatbelt use, and other metrics. According to investigators, GM sold that data to brokers, which in turn supplied it to insurance companies that could use it to set customers' premiums, without drivers' informed consent. For reference, the $220 million settlement, which will be distributed among the participating jurisdictions, equates to less than 7 percent of GM's $3.3 billion in net income attributable to stockholders for its third quarter, which ended on September 30, 2026. Don't spend it all in one place, y'all. The settlement comes more than two years after a 2024 New York Times investigation found that GM had been selling the data collected from vehicle owners through its Smart Driver program. That program, part of GM subsidiary OnStar’s offering, collected the aforementioned data and handed it back to vehicle owners in reports designed to inform better driving habits. Unbeknownst to Smart Driver subscribers, their data was also being forked over to data brokers despite GM telling customers it wouldn't share that information with insurers without their express consent. The insurers ultimately obtained the data by purchasing it from brokers with which GM had struck deals. To make matters worse, the Washington AG's office said GM incentivized dealerships to enroll customers in Smart Driver, with employees in some cases accepting terms and conditions on customers' behalf without their knowledge or consent. GM also failed to notify many drivers that their information was being collected and shared. Its OnStar app incorrectly told more than 370,000 drivers they were not enrolled in Smart Driver, even as GM continued collecting and selling their data, according to investigators. “The data GM sold was personal, identifiable, and specific,” the AG’s office said. GM canned the Smart Driver program in April 2024, just over a month after the New York Times exposed its data-sharing practices, citing customer feedback. In addition to the monetary pittance it’s handing over to the states, GM also agreed with the 36 AGs to limit the driving data it collects, better protect customer PII, and establish a "comprehensive data privacy program.” The Washington AG’s office described its settlement as “significantly stronger than any other resolution reached with GM for this conduct.” Speaking of those other terms, GM finalized its settlement of the same allegations with the Federal Trade Commission in January 2026. The FTC order required GM to obtain drivers' express consent before collecting, using, or sharing covered driving data, subject to certain exceptions. It also required the company to offer a simple way for drivers to request a copy of their data, request its deletion, and disable precise geolocation data collection where their vehicles support it. The state of California separately reached a $12.75 million settlement with GM in the matter in May; a separate case in Texas is ongoing. “This agreement … reinforces steps we’ve taken to strengthen our privacy practices,” a GM spokesperson told The Register, though the company noted that it still stands by driver data collection as a best practice. “Safer, more helpful and personalized experiences are enabled by connected vehicle data, and GM is committed to being clear about what information is used, how it’s protected, and the choices customers have about what they share,” the spokesperson added. ®

Anthropic has denied allegations that a malicious version of the official Claude Code extension published to Cursor’s marketplace was linked to any kind of compromise of its own systems. The company was forced to issue a statement after prominent figures in the cybersecurity industry alleged that the extension, which infected up to 5,000 developers, was uploaded to Cursor’s marketplace following a breach at Claudemaker Towers. Felix Rieseberg, a member of Anthropic’s technical staff, Xeeted today: "We're working with Cursor on this incident, but it wasn’t a compromise of Anthropic's systems or employee laptops. "A malicious actor used access to Cursor's marketplace to publish a tampered copy of the Claude Code extension, which we didn't build or publish. The versions on our official channels are not affected, and this only impacts Cursor users." First spotted by EDR vendor Phorion, an unknown party pushed an update to Cursor's official Claude Code extension for Visual Studio Code that contained a persistent SSH backdoor written for Windows, macOS, and Linux. According to Cursor, the extension – version 2.1.313 – was available to download for around 11 hours between 0227 and 1335 PT (0927 and 2035 UTC) on Thursday. The vendors' analyses of the implant revealed that it collected information about the host machine and user, including usernames, configured Git emails, and public IP addresses, as well as attempting to open a reverse SSH tunnel. Crucially, Cursor echoed Phorion's assessment that the package contained a critical flaw. The script responsible for opening the SSH tunnel required admin rights on macOS and Windows to do so, and in the vast majority of cases it did not have these, resulting in failure. Regardless, the extension still forwards the collected host and user data to the attacker, as well as dropping the attacker's SSH key and installing the persistence mechanism, so developers must go threat hunting in case their instance allows for the SSH tunnel to be created successfully. Simply installing the latest version of the extension is not enough to clear the persistence mechanism, either. The full details of how to remediate the issue, and links to download scripts to wipe the infection, are laid out in Cursor's account of the situation. Aside from Rieseberg's statement today, neither Cursor nor Anthropic have commented on how the malicious package made its way to the official marketplace. Anthropic's position is that a malicious actor uploaded the tampered copy to Cursor, but the latter has not outlined how it reached the marketpplace. "[We] removed the malicious version from the marketplace immediately upon detection and blocked further distribution," Cursor said in the advisory. "We followed with a full shutdown of the VSCode Marketplace out of an abundance of caution." It also ran checks to confirm no other extensions were compromised in the same way, rotated its own keys before restoring the marketplace, and worked with a cloud hosting provider to block the connection to the attacker's server, it said. ®

Voyager 1 gave controllers the silent treatment for a few days this month before resuming communication weeks before the spacecraft reaches the one-light-day milestone. The incident happened on October 18: NASA's Deep Space Network (DSN) couldn't detect a signal from the probe. On October 19, on the next scheduled communication attempt, there was still nothing. Was this it? Had the decades-long galactic venture finally ended? Things are operating on a knife's edge for the two veteran spacecraft, which will reach their half-century next year. Voyager 2 suffered an antenna pointing issue in 2023, and a low power condition caused Voyager 1 to unexpectedly turn off one of its two radio transmitters. The good news is that on October 20, the signal reappeared, leaving engineers wondering why Voyager 1 temporarily went silent. Thus far, it seems a group of thrusters went offline and their backups were turned on. This could indicate onboard computers detected a problem and put the spacecraft into a safe mode. Kareem Badaruddin, Voyager mission manager at NASA's Jet Propulsion Laboratory, said: "This is the first time we've seen the spacecraft do something like this, but we are also at a point in the mission where we have to expect the unexpected. "We're relieved to be back in contact with Voyager 1 and are eager to understand what happened." The Register spoke to Badaruddin in 2024, where he described the challenges of working with the veteran spacecraft, but he reckoned the work was the "best job at JPL." The twin Voyager spacecraft have recently been through a procedure dubbed the "Big Bang," where controllers turned off onboard devices and switched to lower power alternatives. The result will extend the lifetime of the spacecraft by a few more years and postpone further instrument deactivation. It was completed first on Voyager 2, and then Voyager 1, six weeks later in September. However, the brief lapse in communication highlights that despite the ingenuity of the team that keep the spacecraft running, the project must eventually come to an end. In 2026, Suzanne Dodd, Project Manager for the Voyager project, told The Register what the end would look like. She said, "We're just going to one day not hear the signal from the DSN, and we're going to rush around and make sure that it's not something on the ground." Months of listening will follow, before engineers throw in the towel and declare the mission over. And then? Maybe a farewell message sent to the last known location. Thankfully, in this case, the signal returned, so no need for a goodbye transmission just yet. In a few weeks, Voyager 1 will be so far away that it will take a full 24 hours for commands to reach it and another 24 hours for a response to reach Earth. Communications blip notwithstanding, there is every chance both Voyager spacecraft will still be operational as they pass their half-century milestone next year and, technology and funding permitting, continue sending data to Earth for a good few years yet. ®

Just days after Daemonarchy, there's another new OS in the BSD family – and one that is heading determinedly in almost the opposite direction. AlpineBSD is so new that it barely exists yet. There are no installation media available, though it's possible to manually bootstrap an installation. What AlpineBSD does already have is an AI policy, and the plan is to avoid "the output of generative AI models (e.g. LLMs)" as far as possible: "Any software whose code, documentation, binaries, or other assets include, in any part, the output of a generative AI tool such as an LLM, is considered 'tainted' for our purposes." The core idea is to combine the kernel of NetBSD with "a userspace inspired by Alpine Linux and the objective of making it possible to use as little AI-generated software as possible." Back in 2024, the NetBSD project made its position on AI plain. Its commit guidelines say that "code generated by a Large Language Model […] is presumed to be tainted code." Back in July, Linus Torvalds stated that "Linux is not one of those anti-AI projects". Since then, NetBSD has seen a spike in donations and interest. For someone mainly used to Linux, trying to install NetBSD is a little like hopping in a time machine to 30 years ago, and we speak from personal experience of Linux in the mid-1990s. As we mentioned when Slackware turned 30 a few years ago, back in 1996, this vulture tried Slackware 3.0 on his own hardware, and we never did manage to install it. We tried NetBSD 11 a couple of months ago, and despite considerable help from the folks on the NetBSD-users mailing list, we failed to get the CDE desktop to start. As an example, installing software on NetBSD is not the easy, comfortable task it is on Linux. It does have a package manager for handling binary packages: it's called pkgin, but as the NetBSD FAQ mentions, it's an optional extra. You must manually tick a box in the setup program to install pkgin with the rest of the OS, and when we tried NetBSD 9.3 four years ago, we didn't know that, overlooked that checkbox, and had to find a way to add pkgin later. (It is quite tricky to install a package manager when you don't have a package manager to install it with.) This time around, we discovered that pkgin sometimes doesn't complete the installation process for you – it just leaves a text file in a destination directory somewhere, and you're expected to know that, go there, read it, and do any further steps yourself. The NetBSD community has been quietly going its own way since 1993. It has its own ideas about what is normal, and its users expect that. Some folks on the mailing list told us that they never use binary packages anyway – they compile what they need from source code. AlpineBSD's ports repository and source code are hosted on Sourcehut, under the account of sircmpwn, better known as Drew DeVault. We have quoted DeVault a few times over the years: we've often found that his views align with our own in various ways… such as about the best laptops ever made, and on simplicity in Linux distro design, or even on dot-com billionaires at FOSDEM. Alpine's tools are quite well-suited to being adapted to other systems. We have previously mentioned the use of its packaging tool, APK, in the GNU-free Chimera Linux, which uses much of the FreeBSD userland, and also in Adélie Linux. We will follow AlpineBSD with interest. There's strong appeal in the plan of combining perhaps the best-known, longest-established AI-free Unix-like kernel with the user-facing parts of pretty much the smallest and cleanest Linux distro there is. The AI advocates often warn LLM skeptics that they risk being left behind. Well, here's an OS for those who don't want to surf the AI wave. ®