The Register

More than 1 million LinkedIn users have clicked the newish “Seems like AI slop” button since the social network launched it at the end of July. Hari Srinivasan, LinkedIn’s chief product officer, said that compared to the period before launching the AI slop reporting button, those who copy-paste AI-written posts are seeing around 40 percent fewer views. “Despite the progress, we know we have more to do to ensure LinkedIn remains a place where you can find real people and real perspectives... this all remains very top of mind,” he said. No single piece of feedback is used to adjust how widely a post travels, inside or outside a user’s network, he said, and distribution is determined by a range of signals. Srinivasan added that users will also start seeing messages in their analytics dashboards to understand how viewers are receiving their posts. If people think there’s too much AI use, the platform will relay that to the user. Any logged-in user can report a post for suspected AI slop using the platform’s three-dot option menu. The CPO made it clear in his initial announcement of the reporting button that there is a difference between AI and “slop.” LinkedIn believes people have a right to continue using AI to, ironically, make themselves sound more authentic. Regardless of what some may think about the button being used to train AI itself - rather than LinkedIn’s users - on how to sound more human, the company also wants the white-collar social media tryhards making these posts to get human feedback instead of an automated AI detector. Srinivasan said that, behind the scenes, LinkedIn is working on an array of automatic detection mechanisms for AI slop content. Prior to launching the slop button, the platform had used these detections to catch hundreds of thousands of AI-slop comments each day, and “billions of other automation attempts,” such as slop posts and posting at scale, “in the last couple of months alone.” As part of the war against slop, Srinivasan also announced in July that LinkedIn was removing its AI-driven “enhance your post” feature that made posting slop all the more simple. It replaced the feature with a proofreader. In announcing “slop” as its word of the year for 2025, Merriam-Webster noted that the term relates to a broad range of lazily compiled content. Fake news can be slop. Social media posts, on LinkedIn and X especially, can be slop. Propaganda spread by the US president can be slop. Institutions as treasured as UK pubs stand accused of sloppy practices, and pretty much everything in between – AI slop pervades the modern world. In the security space, AI slop bug reports have been a recurring issue. The nature of the headache, however, has changed. As Daniel Stenberg, founder and lead developer of cURL, explained earlier this year, AI-generated vulnerability reports have become a serious drain on maintainers' time as they often look credible at first glance but turn out to be baseless once investigated. Now, though, maintainers are struggling to tackle the high volume of high-quality bug reports that are made with the help of AI – no slop. ®

Work on Linux kernel 7.3 is underway, and a bug fix just came in from an unusual source: Linus himself. Well, kind of… The new bug fix has the title of Don’t hand out the flat CCS storage as usable VRAM, and it comes from emperor penguin Linus Torvalds . Now that kernel 7.2 is out, the team is hard at work on what will be kernel 7.3, which means lots of bug-bashing. However, while Torvalds remains in charge of the overall kernel project, as he has said a number of times, including on stage last November, “for the last almost 20 years, I’ve not been a programmer.” This bug is an exception to that – but it’s only a partial exception. The technical issue itself was not all that interesting. Due to an oopsie in the kernel driver for Intel Xe graphics, machines would freeze during the boot process, just at the point when the OS switches into graphics mode to show the login screen. That alone makes it tricky to troubleshoot. The problem was in how the graphics driver set aside memory used by Intel’s Compute Command Streamer (or CCS for short), which the kernel Intel i915 driver documentation describes as: “An engine that has access to the media and GPGPU pipelines, but not the 3D pipeline.” The driver tries to work out where that is, takes the memory below it, and after some other steps, rounds it up to a multiple of 128 kB and sets that as video memory that’s free to use. The problem was rounding it up – because that sometimes means that it’s marking as available memory that’s already in use. It should have rounded down instead. It’s a one-line fix, but finding the error was the problem. This is where the story gets interesting, for several reasons. One of them is that it has a remarkably large comment inserted in covering email, which reads as follows: There are several points of interest here. One is that the fix is from Linus himself. Another is the unusually discursive comment. The third and perhaps the most important is that he used a bot to find the bug. He also used it to write the bug report, which explains why that part is so dull and lifeless. (The patch description does note, however, that if you just restart the program that shows the login screen – in this case, GDM, the GNOME Display Manager – it works, because the second time around it gets a different block of memory.) It’s a good example of how coding bots are a legitimately useful tool for developers. It looks like he didn’t use the bot to write the fix – that wasn’t needed, it’s a one-line fix – but the bot wrote 42 chunks of test code that enabled Linus to find the bug. It sounds like that was a tough job. (Forty-two, eh? That must have needed some Deep Thought.) As a side note, it’s also interesting to note the continuing issues caused by most modern PC graphics controllers having their own separate chunk of RAM, which isn’t shared by the CPU. This is something some Arm64 machines – notably, Apple Silicon ones – avoid by having a single pool of RAM shared by both. As memory gets more expensive due to the AI-hyperscaler-induced RAM drought, we’re seeing other patches in this area. If you’re short of main system memory but have well-equipped GPUs, then Sean Lobjoit’s NBD-VRAM patch lets a Linux box use any leftover VRAM as very fast swap space. Conversely, Valve’s Natalie Vock wrote a patch that makes low-end AMD GPUs faster when trying to allocate more VRAM than they have, which she explained in much more depth last week. As we speculated earlier this year, we suspect that Apple’s closer silicon integration, and a single shared pool of memory, is the way that the industry will eventually go. Bootnote We thank our esteemed colleague Simon Sharwood for the “Emperor Penguin” sobriquet – we believe that on The Reg, he first used that a decade ago. We also thank him for pointing this fix out to us. ®

Nvidia’s $20 billion bet on Groq’s LPU tech sure looks like it was a good one. On Monday, the GPU giant offered the first glimpse of just how big a speedup its Groq 3-based LPX racks will provide. In an independent benchmark conducted by Artificial Analysis, Nvidia’s LPX rack systems managed to churn out 3,400 tokens a second (tok/s) with a 100,000-token input sequence in Google’s Gemma 4 31B model. According to Nvidia, this makes it 4x faster than the nearest alternative platform, which going off Artificial Analysis’ leaderboard would be a direct dig at Cerebras, which managed a still impressive 882 tok/s under the same conditions. Acquihired by Nvidia in late December, Groq has LPUs that feature an SRAM-heavy dataflow architecture designed specifically for high-performance inference serving. Unlike traditional datacenter GPUs, which rely on high-speed DRAM memory tech like GDDR7 and HBM4, Groq’s chips rely entirely on a large pool of on-die SRAM orders of magnitude faster than even the best HBM stacks (around 2.75 TB/s each) available today. And when it comes to inference, memory bandwidth is the bottleneck to beat. The third generation of the chips launched as part of Nvidia’s broader Vera Rubin platform earlier this year boasted 150 TB/s of memory bandwidth. The catch, of course, is that SRAM chews up a ton of die area, which means you can’t pack all that much into a reticle-sized part. Where Nvidia’s top-specced Rubin GPU boasts 288 GB of onboard memory, each Groq 3 LPU has just 500 MB — 576x less. That’s far too little memory to run Gemma 4 31B on a single LPU, so Nvidia’s architecture uses Ethernet to distribute models across multiple accelerators. Each LPX rack can be equipped with up to 256 LPUs for 128 GB of high bandwidth SRAM. For large models, multiple LPX racks can be ganged together. So why would anyone need to run Gemma 4 31B at 3,400 tokens a second? You probably don’t, but the AI code assistant or agent you're using it with might. The faster you can generate tokens, the longer models can reason, the more turns the agent can take, and the more information it can process or actions it can take in the same window of time. The big bet here is that, in the agentic age, faster inference servers make for smarter agents and are therefore worth a premium. And the combination of Nvidia GPUs and Groq 3 LPUs certainly seems to have caught Nvidia’s customers’ attention. On Monday, the GPU giant revealed that Netherlands-based neocloud Nebius would be among the first to field the combined systems in its datacenters. Just how impressive is Nvidia’s LPX system really? While 3,400 tok/s is impressive at face value, Gemma 4 31B is arguably a best case scenario for the hardware, and it remains to be seen how gracefully the architecture will scale to larger, more complex MoE models. At 31 billion parameters, the model fits neatly into a single LPX rack regardless of what data type Nvidia is using to store the weights. Nvidia is running the model at FP8, which means it needs a little over 31 GB or just under 64 LPUs of SRAM capacity. We’ve reached out to Nvidia for comment on how exactly it's distributing the model across those chips, but didn’t get a clear answer. Our best guess is plain old pipeline parallelism with possibly a dash of data parallelism to improve performance at higher concurrency levels. Pipeline parallelism, if you’re not familiar, distributes the model across the accelerators and then processes decode operations in one great big pipeline. Since there are enough LPUs, by our approximation, to hold four copies of the model, data parallelism would effectively duplicate this pipeline and provide additional independent slots for higher concurrency. If you’re thinking that 31B isn’t very large, you’re right. The model is small enough to run on a high-end consumer graphics card like an RTX 3090/4090 — at 4-bit precision of course. While it may be a small model, we wouldn’t write off Nvidia’s test results just yet. The model itself is relatively small, but it’s a dense model, which means all 31 billion parameters are activated each time a token is generated. Thirty-one billion also happens to be fairly close to the number of active parameters used by much larger MoE models like DeepSeek’s 671 billion parameter V3, which has 37 billion active parameters. It should be noted, however, that because MoE models use different parameters for each token generated, they suffer from performance overheads that you don’t see in dense models. The other problem worth noting is that to serve a model like DeepSeek V3 on an LPX system would require 1,342 accelerators or just over 5 LPX racks. That’s a lot of accelerators to wrangle. Taking LPX to the next level The performance figures shared by Nvidia today are for a model running entirely on LPUs. However, Nvidia’s biggest performance gains will come from combining its GPUs with the dataflow accelerators in a heterogeneous inference architecture. We discussed this approach in more detail earlier this year, but, in a nutshell, what Nvidia is doing is breaking the inference workload into two distinct phases. The compute heavy prefill phase, during which prompts are processed and the key-value caches used to track model state are generated, is done on its GPUs, while the memory-bandwidth intensive decode phase is offloaded to the Groq 3 LPUs. The result is a massive speedup in inference performance beyond what GPUs alone could deliver. The chart below summarizes how Nvidia’s GPUs and LPUs fit together nicely. Rubin’s high compute density makes it ideal for high volume, bulk inference where the goal is reasonable interactivity (tok/s per user) for a large number of concurrent users, while LPUs are best suited for maximum interactivity. By combining the two, Nvidia has managed to lift the Pareto curve up and to the right, maximizing throughput without compromising on interactivity. The 3,400 tok/s recorded by Artificial Analysis reflects performance on the far right of the chart, but by combining Groq’s LPUs with its GPUs, Nvidia expects it can achieve a decent fraction of that performance across a large number of concurrent users. Nvidia’s figures may not age well Along with the LPUs’ potential scaling challenges, it’s also worth discussing Nvidia’s comparison against Cerebras’ chips. Nvidia claims its systems are 4x faster than Cerebras' parts, which is true. Artificial Analysis testing at a 100,000 input sequence (prompt) does corroborate this. What is left out of that conversation is the number of chips required. We don’t know the specifics of Cerebras' config, but it’s not hard to suss out. Cerebras usually serves its models at mixed precision (eight and 16 bits) with 16-bit activations, which means the entire Gemma 4 31B model should fit into one, at most two 44 GB CS-3 accelerators, whereas Nvidia needs at least 64 chips. Cerebras' performance figures also don’t reflect its next-generation accelerators announced last week. The WSE-3T-based CS-4 accelerator doubles the compute, I/O bandwidth, fabric speed, and memory bandwidth of its existing chips, and also triples the number of accelerators per rack. The new systems are expected to launch on Cerebras' inference cloud later this year alongside a heterogeneous GPU+WSE-based system similar to Nvidia's NVL72+LPX config we discussed earlier. The tie up, announced at AMD’s Advancing AI event last month, will see the House of Zen combine its Helios GPU rack systems with Cerebras dinner-plate accelerators. Cerebras is working with AWS on a similar config as well. As such, Nvidia’s Gemma 4 31B performance comparison may not age well. ®

Millions of Tesla vehicles are being recalled in China over the brand's hidden door handles, which regulators worry could present a danger in crashes. Minimalist door handles have long been a feature of Elon Musk's electric vehicles. However, Chinese regulators are less than impressed with the tech and demanded that Tesla, alongside other brands, update the vehicles to mitigate a perceived safety risk, in which occupants might not be able to escape a vehicle or rescuers be unable to gain access during an emergency. According to Chinese authorities, the affected vehicles "have emergency mechanical door handles that are difficult to identify and operate because their color is similar to the interior trim. "In extreme situations such as a severe collision, causing the vehicle's low-voltage system to fail, this could hinder occupants from quickly opening the doors to escape and impede rescue efforts by those outside the vehicle, posing a safety hazard." The plan is to affix warning labels (some vehicles already have them) and apply an Over-The-Air (OTA) update "to add a post-accident window lowering strategy to reduce safety hazards." Tesla is not the only car maker affected. A recall of almost 400,000 Xiaomi SU7 electric vehicles was also ordered in connection with difficult-to-see emergency mechanical door handles. China announced a ban on hidden door handles earlier this year, citing safety concerns, and has required all new models from 2027 to have mechanical interior and exterior door handles. Advances in door handle tech have crept into other manufacturers' products in recent years. Doors can be opened with a key fob or mobile phone application, or manually released. Some manufacturers hide the internal and external mechanisms behind the push of a button, which appears to have attracted the particular ire of regulators. After all, in an emergency, there might be delays in evacuating vehicle occupants when mechanical release mechanisms are not immediately obvious. In 2025, Tesla was sued over allegations that a design flaw in the company's Model S prevented the vehicle's occupants from escaping after a crash. The company has also come under scrutiny following allegations that occupants of a Cybertruck were trapped after a crash due to the door handle design. ®

As Oracle launches its Exadata service on Database@AWS — a year after it went live for lower-level databases — experts have warned there are trade-offs in getting value for money and cautioned over Oracle FUD around the cloud migration decision-making. Earlier this month, Big Red announced the general availability of Oracle Exadata Database Service on Exascale Infrastructure on Oracle AI Database@AWS, which adds the beefed-up Oracle database to the service. AWS was the last major cloud service to join the “database@” offer from Oracle, allowing the database and application biz to put its hardware inside the hyperscaler's datacenters. Announced in 2024, Oracle Database@AWS can connect to applications running on Amazon Elastic Compute Cloud (Amazon EC2), AWS Analytics services, or AWS's advanced artificial intelligence (AI) and machine learning (ML) services, the vendor claims. The service became generally available in July last year. Oracle now says it is bringing Exadata – the engineered hardware and software platform for high-performance databases — to AWS. The dominant cloud vendor has also signed a new long-term strategic collaboration agreement to help migrate customers. Using Exadata allows customers to specify the compute and storage capacity they need, says Oracle, while spreading every database across pooled storage servers for high performance and availability, helping to avoid the burden of creating database and storage servers. However, Oracle licensing and commercial experts have warned that there are still trade-offs when adopting the service and getting value for money. While Database@AWS can be cheaper in terms of processors /CPU required than RDS, the compute and storage are more expensive and the service options are less granular, making it more difficult to right-size services, said Nick Walter, CTO and vice president of professional consulting services at House of Brick, an Oracle licensing advisory company. “The Oracle Database@AWS offerings have been a constant topic of conversation with clients in the last year. It's a complicated topic because the cost structures, minimum configurations, and licensing metrics are so different from running Oracle databases in AWS native services such as EC2 or RDS,” he said. For the database@AWS service, licensing metrics can be better than going to EC2, in that for users bringing their own licenses (BYOL), one processor license is required per 8 ECPU ratio, in contrast with the traditional one processor license per 2 vCPU ratio when running in RDS/EC2. Oracle's support rewards program can also offer good value. However, compute and storage costs for Oracle Database@AWS offerings are “much higher than traditional RDS/EC2, and the capacity isn't as granular,” Walter said. The Exadata offer helps with some of these challenges but does not eliminate them altogether. “Exascale offerings may make the Oracle Database@AWS offering attractive to organizations with smaller fleets of Oracle databases as the provisioning is very granular,” he said. While Exascale database storage costs per GB are competitive with the AWS Elastic Block Store GP3 options and offer higher I/O throughput, compute costs are higher than EC2/RDS, Walter said. “Overall the Oracle Database@AWS offerings can be attractive to license-constrained organizations that need Oracle Database deployments in the cloud. Buyers should beware though, House of Brick has seen Oracle sales reps gleefully touting possible license savings and not mentioning the downsides,” he said. Craig Guarente, founder and CEO of Palisade Compliance, an Oracle licensing advisory company, said many organizations were still bogged down by legacy agreements for Exadata machines. "Oracle has created some FUD, by saying they won't support customers on a particular cloud, [and] it's 'going to cost twice as many licenses if you just move to the generic cloud and not our services'. Oracle has created some artificial business blockers to get customers thinking about spending more money with them through these Exadata services,” he said. Hyperscalers work with Oracle on these offers because they would love to get some of the business from Oracle on-prem workloads. Oracle can also create barriers to running their generic database in the cloud that might also get customers to consider OCI. While promoting the database@AWS service, AWS is also telling customers to use its Bedrock AI to help move Oracle database applications to its PostgreSQL services. “All of the hyperscale vendors — Google, Microsoft, AWS — are on one side, and Oracle on the other. They all want to generate more revenue; if they can take Oracle revenue and make it AWS revenue, I'm sure they would want to do that. However, they're all in this weird state of cooperation with Oracle, where now they're all authorized cloud providers. They're getting on stage together, and they can't say Oracle's crazy with their licensing. There's limited stuff that both parties can say, but I've definitely seen from the sales side — pick a vendor — that the knives are out,” Guarente said. The Register has asked Oracle to comment. ®

ShinyHunters has claimed another cybersecurity scalp, but ReliaQuest says the crew's social engineering attack only got as far as one employee identity before its defenses slammed the door. The ransomware baddies listed US-based infosec biz ReliaQuest on its leak site on August 23, claiming the corporation as its latest victim. The listing, seen by The Register, links to screenshots the gang claims show access to ReliaQuest's Okta dashboard. It did not publish any stolen customer data, however, and SOCRadar said it had found no validated data samples, ransom demand, or evidence of customer impact. There had already been some public needle between the two sides. Days earlier, ReliaQuest researchers posted about ShinyHunters registering company-name “.claims” domains as part of its social engineering campaigns. An account associated with the crew responded with screenshots and the question: "Who's hunting who?" ReliaQuest has now responded to the crew’s claim, confirming an attack took place on August 22 but disputing the suggestion that ShinyHunters compromised its systems. "On August 22, 2026, ReliaQuest was the target of a social engineering attack. It was unsuccessful beyond temporarily exposing one identity," a ReliaQuest spokesperson told The Register. "The extent of the access was view only. No ReliaQuest applications or systems were accessed, and no customer data was ever touched." The company said it would make no further comment beyond the statement and a technical account of the incident published on its website. According to that account, the attackers set up a fake ReliaQuest single sign-on page and called several employees while pretending to be members of the company's security team. One employee entered their password and approved an MFA push, giving the attacker temporary access to their identity session. ReliaQuest says that was as far as ShinyHunters got. Device-trust controls prevented the attacker from using the session to reach company applications or systems, while the security team killed the session, expired the employee's password, and reset their authentication factors. ShinyHunters, unsurprisingly, has a different take. "This time the post is about you, not us. Let Mandiant report and advise on us accurately, go away," reads a post on the gang's leak site, For all the posturing, the two sides aren't actually arguing over whether ShinyHunters got a foot in the door. ReliaQuest acknowledges that an employee was successfully phished and an identity session briefly exposed. The fight is over what happened next: ShinyHunters is presenting ReliaQuest as its latest victim, while ReliaQuest says its controls stopped the attacker before they could turn that foothold into access to anything that mattered. ®

Semiconductor revenue is set to nearly double this year, with Gartner projecting the industry will bring in $1.6 trillion in 2026, up sharply from $809 billion in 2025 Memory chips are driving almost all of that growth: DRAM and NAND revenue are each expected to roughly triple or quadruple, pushing memory past non-memory chips in total revenue for the first time. Gartner expects memory alone to top $1 trillion by 2027. Last year, memory revenue was $220 billion vs $589 billion for non-memory chips. DRAM revenue is expected to bounce 246.6 percent in 2026, while NAND flash revenue is expected to see growth of 371.9 percent. The ultimate cause of this run on memory is the AI infrastructure construction boom, underpinning demand for high-end servers and GPUs. Manufacturers such as Samsung, Micron, and SK Hynix found it more profitable to prioritize production of these chips, leading to a shortfall of the mainstream memory types needed for PCs, smartphones, and other devices, bumping up average unit prices in this area. The Register has reported on sharp price hikes that buyers of PCs and of smartphones are facing. The market for phones in particular is forecast to shrink by 15 percent this year, as consumers balk at higher price tags. Chipmakers have also locked in high prices for longer by striking deals with large customers who are paying for longer-term supply stability. Micron announced this in June, while SK Hynix and Samsung revealed similar arrangements in July. Gartner believes that AI infrastructure has fundamentally changed the dynamics of the memory market, meaning higher prices are not cooling the growth in demand as seen in previous tech industry boom and bust cycles. “The semiconductor industry is entering a fundamentally new phase of growth,” claims Ben Lee, director analyst at the firm. “As AI infrastructure scales, it is not only reshaping investment across the semiconductor ecosystem but also redefining the industry’s application mix." The AI datacenter ecosystem is expected to grow from 36.5 percent of semiconductor revenues in 2026 to more than 53 percent by 2030. This represents a structural shift in where semiconductor value is created and how demand is evolving across the industry, Lee says. All this is bad news for anyone hoping memory prices might come down in the near future. Gartner is forecasting sustained demand for high margin memory products, including HMB. Samsung warned last month that the memory supply crunch will likely deepen next year and persist through 2028. Gartner's own projection calculates that semiconductor revenue will reach $1.9 trillion In 2027, an increase of about 20 percent on this year. ®

The author of the original legendary 1996 Windows Task Manager has released new versions for Windows, Mac, and Linux, telling The Reg the project started life as a way to see if vibe coding could produce something useful. "I picked Task Manager as a simpler target to see how ready AI is for primetime… or if it would just degrade to slop," former Microsoft engineer Dave Plummer told us. "And it turns out it depends on what you ask for, largely! And what you test and instrument. It's like having a very ambitious and precocious intern, but like the classic genie in the bottle, be careful what you wish for." The update is a much fuller-featured affair, designed to offer insight into the inner workings of a user's system rather than behaving like a bare-bones process-killer as it once did. As such, it bears a passing resemblance to the current Windows Task Manager. However, it can drill into more detail and expose more metrics. Aesthetically, as well as the similarities to the Task Manager found in Windows 11, the interface gives off similar vibes to media player skins of the 1990s and 2000s. "I wanted it to look and feel like my Pioneer RT-909 reel to reel," said Plummer. The user can set the font and theme to something more suited to the their task. However, if you fancy that Amber or Green screen look, that's possible. So is full color. Users can also set the bloom and saturation via a slider that goes up to 11. The application started as a 107-page specification that Plummer passed to Fable. After a three-and-a-half-hour first pass, the first iteration of the application was produced. "It balked a few times thinking I was doing security work," said Plummer, "but mostly cooperated." The cross-platform nature of the application is interesting. Rather than going down the Electron path, Plummer and his team elected to use native UI code. The app is Swift on the Mac version, Direct2D/DirectComposition on Windows, and Qt on Linux. "It's more work," he said, "but worth it in a task manager or something that has to have game-like response times." "The biggest surprise," said Plummer, "was that I could get them looking as close to one another as they do given the platforms are wildly different." It's efficient too; the Windows version we looked at was roughly half the size of the Microsoft version, and used half the amount of RAM. The code is proprietary for now, and there's a Pro version in the works with more features which will also carry a price tag. The price has yet to be set. There could well be a further beta release between now and then – this is, after all, pre-release software at present. As it stands, the updated take on Task Manager is an example of what can be achieved with imagination and vibe coding, and highlights the importance of specification and prompt engineering to produce something useful too. Also, considering it is fuller-featured than Windows Task Manager, more customizable, and more efficient, it is hard not to see it as a challenge to Redmond, even if that was definitely not the authors' intention. ®

Developer Matt Callaghan claims he caught Alibaba trying to track web users by playing sounds through browsers vulnerable to audio fingerprinting. The software engineer drew attention to the issue late last week after investigating why his Bluetooth headphones stopped playing music whenever he visited Alibaba’s website. “Recently I ran into a strange problem with my Bluetooth headphones,” Callaghan wrote. “They support multipoint Bluetooth audio, so they can be connected to my PC and phone at the same time. Normally, the PC takes priority playing audio, with my phone being able to play audio when nothing is playing on the PC. “Usually I listen to music on my phone but with notifications or YouTube playing through the PC, this works reliably until I open an AliExpress page in Firefox or Chrome. “Shortly after loading the AliExpress homepage, audio from my phone would stop playing. Closing the AliExpress tab fixes it immediately. Muting the tab/Firefox/Windows does not help, and there is no visible video, music, or other media playing on the page.” Callaghan tried to find any hidden conventional media elements but found nothing. Further digging revealed two audio scripts that he said were “extremely obfuscated” within Alibaba’s browser security and anti-abuse tooling. He said the scripts built a WebAudio graph that introduced a sawtooth oscillator to generate a waveform, an analyzer to measure the result after the waveform passes through a browser’s audio implementation, and a script to read the associated frequency data. The scripts set the audio’s gain to zero, meaning the end user won’t hear anything, but the WebAudio graph will still be processed by the browser. “This is very different from an autoplaying video,” said Callaghan. “There is no media element for the browser's normal tab mute control to stop. As far as the page is concerned, it is performing live audio processing. “In my case, that appears to have been enough for Firefox or Windows to keep the Bluetooth audio path active, preventing my multipoint headphones from switching cleanly back to the phone.” Callaghan found further evidence in the code of Alibaba looking for data related to screen dimensions, device memory, browser plugins, WebGL rendering, mouse events, and more. As well as signs that Alibaba is encrypting data and sending it to its telemetry services, the developer said all of it amounts to “a fairly comprehensive browser and device fingerprint.” The Register has asked Alibaba to comment. Despite Callaghan saying he could reliably reproduce this issue on both Firefox and Chrome, Firefox issued a Xtatement saying its anti-fingerprinting technology thwarts Alibaba’s tracking tricks. It pointed to a blog post from Tom Ritter, a security engineer on the Firefox team, who explained that as of version 118 (September 2023), the protections it introduced eliminated the efficacy of WebAudio-based fingerprinting. These protections are not designed to stifle fingerprinting efforts at the source. Instead, they work to group all users together, making it look like all fingerprinted users are the same, effectively nullifying the tracking attempts. For 99.24 percent of users, they fall into one of three “buckets” – user categories delineated by types of hardware. The vast majority fall into buckets one and two: Bucket one: x86/x64 CPUs lacking FMA (Fused Multiply-Add) instructions Bucket 2: x64 CPUs with FMA instructions And for the remaining 0.76 percent, the fingerprinting script failed entirely, according to Firefox’s data. However, Ritter said there are 48 users worldwide who do not fall into the three buckets, or the 0.76 percent whose machines did not allow the scripts to run. These 48 users fell into 23 other minuscule buckets, which means they are not grouped into the masses like the rest, and so fingerprinting is more effective on this vast minority of users. “This is very unfortunate, as it makes these users completely unique, but it is also not terribly unusual - computers are weird and these results could have been caused by bad RAM, a CPU bug, or possibly some crazy architecture (LoongArch??),” said Ritter. “But at the end of the day, WebAudio fingerprinting is nearly useless. I don't expect browser fingerprinting to disappear from websites entirely (unless some regulatory action occurs, fingers crossed) - it's still going to be effective against a majority of users on the web, but at least for privacy-focused browsers, it should be wildly less effective.” Brave, maker of the eponymous privacy-centric browser, also Xeeted a response to Calalghan’s findings, saying it has protected users from fingerprinting for six years. “Brave injects random data into the browser's output so you show a different fingerprint to different sites. This fingerprint also resets across sessions. “For added protection, we also block the specific scripts used by AliExpress for the tracking method mentioned above. Again, this is done by default for all Brave users. You don't have to change any settings to be shielded from this audio fingerprinting.” Ritter said Chrome and Safari “probably have defenses against this [brand of fingerprinting].” Safari deploys Advanced Tracking and Fingerprinting Protection to prevent WebAudio-based tracking and other methods of fingerprinting. It works differently to Firefox, though, injecting audio errors into an audio buffer, instead of trying to lump all users into an identical bucket. Chrome, on the other hand, does not aggressively protect users from fingerprinting, as privacy consultant Alexander Hanff said earlier in the year. "There are at least thirty distinct fingerprinting techniques that work in Chrome right now, today, as you read this," he wrote. "Not theoretical attacks from academic papers that might work under laboratory conditions – real, production techniques deployed on millions of websites to identify and track you without your knowledge or consent." ®

Datacenters in America used about 17 billion gallons of water during 2023, triple the amount they used a decade earlier in 2014, according to an estimate from US Congressional Research Service (CRS). In a recent report published by the Library of Congress, the CRS casts an eye over how server farms use water, looks at potential options to reduce this, and discusses the current federal role in overseeing the use of H2O by operators. The CRS notes the federal government does not systematically assess all water use by the industry - part of the problem is that water usage data is often collected by state-level bodies. Sudden upswing The report says the AI craze has led to a massive surge in datacenter construction since OpenAI released ChatGPT at the end of 2022, which triggered a wave of demand for generative AI and large language models (LLMs). Many new-build data halls are also home to specialized AI hardware, which typically burns much more power (and consequently has greater need of cooling) than older infrastructure. The CRS data says water use in US bit barns had already tripled by 2023, which is the year that the AI-driven infrastructure build boom took off, and it is probable that liquid intake has increased further since then. However, direct use of water - the amount used at the facility itself for thermal regulation, rather than indirect use during off-site power generation - is estimated to make up a relatively small proportion of total US water consumption of about 2 percent. Indirect water consumption via electricity generation made up more than 80 percent of the datacenter's overall water use, the study found. CRS doesn’t go into too much detail about actual cooling systems, noting that datacenters may rely upon different types of technology depending on the size and location of the facility, and the thermal regulation needs of the IT kit. Liquid cooling systems can be more efficient than air-cooled systems, but may not be cost-effective for smaller campuses. Cool runnings As The Register has also previously explained, there are trade-offs between water consumption and power usage. Evaporation-based cooling systems are generally more energy efficient than air-cooled chillers or other waterless systems, but use up much more water. Which system gets chosen may depend on location and climate as much as anything else. Datacenter operators may draw their liquid from various sources, typically municipal drinking water systems, but also treated wastewater or direct withdrawals from surface water or groundwater. The CRS estimates that municipal water systems make up 97 percent of the water supplied to US facilities. The CRS notes that water use charges to operators are often defined in a water service agreement between the municipality and the operator, and these can include provisions that limit the disclosure of this information. The availability of sector-specific or facility-specific data on water usage, such as pertaining to datacenters, is limited. Water utilities may collect water usage data and report it to state or local agencies, if required, however the federal government has not conducted a systematic assessment of water use by the datacenter industry. Partly, this is because water provision is a state-level responsibility. We reported last month how a water resources scientist has filed a lawsuit against Amazon Web Services, alleging it published false and misleading statements about its water use in Northern Virginia, with the evidence drawn from billing and water consumption records obtained using Freedom of Information Act (FOIA) requests sent to water utilities in the region. Pooling the data Federal funding has supported the development of a water use database; researchers from the Virginia Polytechnic Institute and State University have published the United States Water Withdrawals Database, a compilation of water withdrawal data and estimates for 42 states, which includes public supply, industrial, and commercial water use values reported from water users that are required by state law to do so. The CRS report lists a number of bills introduced during this term of Congress that propose to address datacenters and water use through a variety of mechanisms. Some of these bills include provisions to improve the information available about water use by facilities, others seek to incentivize certain practices such as water reuse. Most of these bills have not moved past the introduction stage, CRS says. According to the campaign site Data Center Water Leaks, 36 of 52 state legislatures have introduced some form of water legislation relating to server farms, while California and Arkansas vetoed local bills that were proposed. At the federal level, it says 20 bills are currently active (14 House and 6 Senate), though only three have advanced beyond the referral stage. Outside the US, recent European Union legislation requires datacenter operators to explicitly report annual freshwater consumption, along with other metrics about their facilities. But the European Commission has been facing pushback from the industry about its efforts to enforce greater sustainability. Trade body CISPE warned last year that onerous water regulations could push operators to shift infrastructure outside the EU. More recently, it was reported that the Commission was weakening its proposed environmental rating system for datacenter energy and water efficiency in response to lobbying from the industry. ®

Apollo Global Management has admitted that attackers talked their way into its cloud systems and got their hands on Social Security numbers and other personal information. The investment giant disclosed the breach in a notification filed with California's attorney general, saying it had suffered a "social engineering incident" that resulted in unauthorized access to "certain cloud platforms" between July 6 and July 10. Apollo isn't saying which cloud platforms were compromised, how the attackers got in, or how many people are affected by the breach. What it does say is that its investigation determined on August 12 that the potentially compromised information included names, dates of birth, contact information, home addresses, and Social Security numbers. The company said it has so far found no evidence that the information has been publicly posted or used for identity theft or fraud. Those affected are being offered 24 months of credit monitoring and identity protection services. "Upon detecting the incident, we promptly notified law enforcement, engaged leading outside cybersecurity and forensic experts, enhanced our security protocols, and launched an investigation," Matthew Breitfelder, Apollo's global head of human capital, said in the notification. The breach comes weeks after Google warned that UNC6671, an extortion-focused crew also known as "BlackFile," was targeting private equity firms and other financial-sector companies. Reuters reported at the time that Apollo was among the companies targeted, alongside Blackstone, Bridgewater and Bain Capital, although it was then unclear whether any of the attacks had succeeded. Apollo's disclosure now confirms that, in its case at least, someone did get in. While the company has not attributed the breach to UNC6671, its notification makes a point of linking the incident to similar attacks elsewhere, saying: "Similar to other financial services firms, Apollo recently experienced a social engineering incident." Google said UNC6671 has been calling employees on their personal phones while posing as colleagues or IT support staff, then steering them to spoofed login pages built to harvest credentials and multi-factor authentication codes. Once inside, the attackers steal corporate data and threaten to publish it unless victims pay up. Some payments have reached $750,000, according to Google's researchers. The incident lands amid a run of attacks in which crooks have gone after employees rather than trying to batter their way through corporate defenses. Earlier this month, Levi Strauss said that social engineers compromised three employees' company-issued computers and made off with corporate data. For Apollo, plenty remains unanswered, including whose Social Security numbers were exposed and exactly how many people now have reason to watch their credit reports more closely. ®

KETTLE This week, we learned a lot more about what’s going on inside the datacenter. GPUs may still be king but their exclusive reign on the rack is over. The taskbar's supremacy at the bottom of the desktop may be coming to an end too. You can listen to the latest episode of The Kettle right here on this page, as well as on Spotify, Apple Music, or YouTube. Those platforms also let you subscribe to Kettle, so you are always notified when the latest episode goes live. Register Systems Editor Tobias Mann joins us to talk about why so many vendors are turning to alternate silicon from the likes of Cerebras and Marvell to help power their inference. He also explains why Waymo is designing its own custom chips in robotaxis rather than continuing to use off-the-shelf silicon from Intel. As new AI silicon moves forward, Microsoft goes back to the future by finally, after only five years of waiting, allowing users to move the taskbar in Windows 11. It also has belatedly started giving users more control over context menus and given them a closer look at what their neural processing units (NPUs) are doing in task manager. Register Microsoft Ecosystem Reporter Richard Speed weighs in. A lightly edited transcript is below. Avram (00:04) Hey everyone, welcome to this week's episode of The Kettle. I'm Register US Editor Avram Piltch, and I am joined today by Tobias Mann, our Register Systems Editor, and Richard Speed, our Microsoft Ecosystem Reporter. We're going to talk about some really exciting new developments in AI hardware and changes to Windows 11 that have been long overdue. So let's get started. This is thankfully a quieter week than usual, maybe because we're nearing the end of August and we don't have to talk about the AI bubble popping or any kinds of economic disaster. We can actually talk about some new technological developments and geek out over those, which is a lot more fun. One thing that we've noticed is that a lot of AI companies seem to be moving away from traditional GPUs and are looking at different types of hardware. Tobias, I know you've covered this in a bunch of stories. Specifically, I wanted to get started by talking about Cerebras. Can you tell everybody a little bit about Cerebras and what they do? Tobias Mann (01:24) Sure. Cerebras has been around for quite a while. They've been poking at the AI infrastructure problem for a lot longer than large language models have been the dominant force in all of this. The company, founded and led by Andrew Feldman, differentiated itself pretty early on by, rather than building GPU-sized chips, their chips are the size of dinner plates or charcuterie boards, and this has fgiven them an early advantage during the ChatGPT days up until recently. Because they're not reliant on high-bandwidth memory, which is supply-constrained, their chips are so large that they can pack 44 gigabytes of super-fast petabytes-per-second SRAM on them. Starting around 2024, they launched an inference server that allowed them to serve tokens way faster than any GPU on the market. It doesn't matter how many GPUs you stack together, they could serve the model faster. However, they did have some challenges associated with the amount of compute that they had. A big shtick for them was that they could get more compute out of their chips using something called structured sparsity. It takes advantage of the number of zeros in the training sets. While that's great for training models, it doesn't do anything for inference. You would think that would have been a dead end for them, but they have found a path forward by partnering with other companies to do the compute-heavy side of inference for them. Earlier this year, they partnered with AWS for this purpose. Just last month, they partnered with AMD. Feldman is notoriously not a fan of Jensen Huang or Nvidia. He previously labeled Nvidia as an AI arms dealer for continuing to push GPU sales into China. There has been this shift away from trying to do everything on one kind of chip, a GPU or custom AI accelerator, to distributing it across two. If that weren't enough, we were getting towards the end of the lifecycle for Cerebras' last chip, and just this week we got a glimpse of what its next generation is going to look like. Avram (04:25) So, what is the difference with the next generation of Cerebras? Tobias Mann (04:30) Well, the biggest differences come from the rack architecture. These chips are huge and super power-hungry. The last generation took fifteen kilowatts just to power the chip. To put that in perspective, modern Nvidia GPUs pull about 1.2 kilowatts, so they're quite a bit hotter. They used to be placed in these big monolithic chassis – think of a server that's half a rack tall, liquid-cooled, and had everything you needed to deploy it in that one box. But as it turns out, data center operators don't really like operating like that. They would prefer to be able to interchange power, cabling, and compute modularly. What we've gotten this generation with the CS-4 systems is a rack-scale system that includes three chips in a modular fashion. If a compute unit goes bad, the customer only has to swap the compute, they are not also replacing all of the power delivery equipment as well. Another big improvement with this generation is that the power delivery is so much better that Cerebras has doubled the clock speed on its chips. It's calling them a turbo variant, but in reality what they're doing is pushing twice as much power through them. So they've been able to double the performance, double the bandwidth of the memory, and double the I/O speed. They're effectively getting twice the compute out of the same silicon, and they've put three times as much in the rack. Avram (06:24) Who's going to use Cerebras? Tobias Mann (06:26) This is an interesting question because it has changed a little bit since Cerebras launched their inference service in 2024. At the time, Cerebras was mostly going after customers that needed to generate what are called premium tokens in the industry. It's a really silly way of saying, "Hey, I can generate hundreds or thousands of tokens a second," where a GPU system with maybe six dozen GPUs might only be able to do a few hundred per user. This is beneficial for things like coding assistants. One of Cerebras' largest customers to date was OpenAI. When OpenAI launched their Codex code assistant, they also launched a Spark model that ran specifically on Cerebras hardware because of how quickly they could generate tokens. Fast tokens are great for code assistants, but fast tokens are also cheap tokens. The faster you can generate tokens, the more work you can get done in a second, driving down the operating cost of the hardware and how much you have to charge to draw a profit. At the end of last year, Nvidia acquired Groq for $20 billion. The entire purpose was that Groq did something very similar to Cerebras, achieving hundreds of tokens a second for things like coding assistants. By combining them with Nvidia GPUs, Nvidia was going to be able to do that for thousands of users at those speeds. It's high volume, high speed. The rest of the industry didn't have an answer to this. So Cerebras has positioned themselves as the alternative to Nvidia/Groq for everybody else in the industry. AWS is one of the major partners for doing this, so expect to see services built atop AWS's AI services to rely heavily on Cerebras. I would also expect to see AMD, as they're starting to sell their own rackscale system that we looked at back at Advancing AI last month, I also anticipate we're going to see AMD pushing it as a kind of two-part deployment option to its customers, in order to drive higher speeds for premium tokens but also drive down operating cost of running inference services like Base10, Fireworks, or any of these other AI-as-a-service providers. Avram (09:23) Makes a lot of sense. We also see that Google is moving away from traditional GPUs, or at least getting another partner to help them with their TPUs, correct? Tobias Mann (09:38) Yeah, that also happened this week. A little bit of context here: Google is a cloud operator. I don't think that's a surprise to anybody at this point. Every sufficiently large hyperscaler eventually ends up in this business. Google's customer base has long consumed traditional GPUs, and since around 2015, Google started developing and introducing its own custom AI accelerators called Tensor Processing Units (TPUs), primarily for internal workloads. This is something we've seen with a lot of cloud providers: as they scale up, many have advertising businesses or recommendation systems for search traffic where it's beneficial to have something that customers don't necessarily need as much of. The open secret with all of these custom cloud accelerators, CPUs, etc., is that none of them are really built from scratch. Google has lots of talented chip engineers, but there's a lot of stuff that goes into an AI accelerator or CPU that is just boring, and if you put a bunch of engineers on this stuff, you're just reinventing the wheel. These cloud providers have relied heavily on big IP houses like Broadcom to design the boring bits of the chip so they can spend most of their time designing the fun stuff that's going to make them money, whether serving advertisements or powering AI services. as was the case of Google. Broadcom was that partner for Google, that was publicly disclosed as part of an Anthropic deal.Broadcom started getting rather loud about being the IP provider for a lot of these chips. They also provided IP for Meta, that was publicly disclosed this year. The sea change this week was that Marvell is now getting involved in all of this. Marvell is another IP house that has largely built its holdings through acquisitions of companies like Cavium, I believe Innovium is also in there if my memory is serving me correctly. They bought Celestial AI for photonics connectivity late last year in a multi-billion dollar deal. Marvell has long been a competitor to Broadcom in the networking space, but Broadcom is a massive monster of a company. There's a reason we call them Broadzilla; their IP holdings are massive. Marvell has reached a point where it's now competitive and has IP for custom XPUs. Google could go to Marvell and ask them to design a TPU alongside Broadcom and then see which performs better. This is I think where things are going. As anybody who's followed Broadcom long enough knows, once they have an effective monopoly, prices and licensing fees tend to go up, much like we saw with VMware not too long ago. Avram (13:19) So Google is getting in on the action to do more to get custom silicon, and now we find out that robotaxis are getting special silicon as well, which I don't fully understand why they need. Perhaps you can explain, because Waymo is now designing custom silicon. Tobias Mann (13:45) Yeah, this was one I wasn't expecting this week. Waymo disclosed that they had developed a custom machine learning accelerator specifically for autonomous vehicles. This is a market where there is no shortage of off-the-shelf options from companies like Intel Mobileye, Qualcomm, Nvidia, AMD. Basically, if you want to build a robotaxi or autonomous vehicle, you don't have to look very hard to find something off the shelf. We've seen companies like Tesla go down this road, but Tesla is a massive electric vehicle company that has made a lot of inroads through technological design and innovation around full self-driving, so it's less surprising there. Waymo is a lot smaller, but arguably a lot more successful in getting robotaxis on the road that actually work and are legal and have actual potential of getting you where you want, although getting stuck in a roundabout is always a hilarious sight to see. For Waymo, the problem they're trying to solve is that these cars have a tremendous amount of sensor data streaming through them. There's very little time between seeing an obstacle on the road and taking evasive action to avoid it. While Waymo can have a human pop in and take control of the vehicle, that's not an ideal situation if you're hurling down the road at 30 miles an hour and someone steps into the street. By the time a human takes control, it's too late. Latency is a huge factor in how quickly it can spot an obstacle and hit the brakes or steer away. So Waymo has decided to build this chip. Avram, you mentioned why they would need this given all the off-the-shelf stuff that had been working for them. From what I can tell, they have largely been building these autonomous systems using FPGAs. FPGAs are fantastic for super-low-latency workloads. They're entirely programmable, but they're hard to program for, it's a very specialized group of people who can program for them, and they're not very compute-dense. To add more sensors, you have to add more FPGAs to the compute stack of these vehlcles. That consumes more power and compounds your development maintenance stacks. I suspect Waymo's motivation for moving toward an ASIC in this case is to get better compute density that's easier to work with from a development standpoint, without missing out on the low-latency characteristics FPGAs are infamous for. There's a reason why all high-frequency trading is done with FPGAs today – if you want to take action very quickly based on a stimuli, they're very good for that. But ASICs can be just as good if they're designed for that purpose. Avram (17:48) Makes a lot of sense. So I think what we're seeing here is that there's more diversity of hardware going into inference, going into the AI industry, so that we can get better power management, better performance, maybe better economics, easier to program. So a lot of a lot of interesting things happening here that are moving outside of the domain of just the big big hardware companies like Nvidia. Speaking of big companies, I wanted to shift gears and talk about Windows 11, which is almost five years old now and still can't do some things Windows 10 or 7 could, like let you move the taskbar. I wanted to ask you, Speedy, as I read that we might finally be able to move the taskbar again. Is that true? Richard Speed (18:57) That is true. Full disclosure: it's never bothered me personally having the taskbar on the bottom, but for some users they really care about this. And it's one of the things that Microsoft removed, it was in Windows 10, as you say, it was in Windows 7. And yet inexplicably when Windows 11 arrived, just well it's almost five years now, they'd taken it away. And the the reason for that was they basically rewrote the the interface they decided to drop moving the taskbar, and users have complained about it ever since. What's interesting is that it's now coming back – it has arrived in the Release Preview channel of the Windows Insider program, meaning it should be out publicly in the next few weeks or months... What I think it speaks of is paying attention to users once again because this has been a thing people have been asking for. They didn't ask for Copilot everywhere, they didn't ask for ads everywhere. They just wanted to move their taskbar around like they always used to. And so finally that's happening. I mean it kind of echoes comments that came from the the Windows boss Pavan Davuluri ... early this year he said Microsoft would be rethinking how they put Copilot into things, they'd be rethinking things like the taskbar and that the fact that users want to do it to work a certain way. They'd be rethinking the Start menu. They'd be paying more attention. And I think when your boss, in this case, Satya Nadella, is telling investors that the company needs to win back its fans, clearly it understands that there's an issue here. So yeah, as of now in the release preview channel, you can now move the taskbar left, right, top, or bottom, you can make it smaller. You can adjust the Start menu as well, make that smaller or larger too. So it's all customization. One of the concerns I've got is it's taken Microsoft five years to get to this point, which is a bit of a worry. Avram (21:07) I give them zero credit for doing this. If you take something away from users and take five years to give it back, you don't deserve a pat on the back. They created a situation where they could congratulate themselves for restoring a removed feature. While I don't really care about having the taskbar at the top or the side, I'm happy to have it on the bottom of the screen But they took away a choice from me for no apparent reason. I know, it's the code. But you know what? That's not my problem. That's their problem. They wrote the code in such a way that they overrode users' choice. How long have you been able to move the taskbar? Richard Speed (22:02) It goes back years and years. Avram (22:03) Since Windows 95, hasn't it? Richard Speed (22:05) I think there's two aspects to this. The first aspect is obviously the concern about how long it's taken. From my perspective, it's like there's been five wasted years. People complain about this, why is it taking so long to actually do this? Yeah, as you say, they don't deserve a pat on the back. They have wasted time before this putting in things that users perhaps weren't that concerned about, like the adverts, like Copilot everywhere. But it's also the company has recognized that there were certain things that it it did do right in Windows 10. A good example is the right-click context menu, which is now in the experimental channel. Now there Microsoft has actually come out and said, you can reuse your Windows 10 muscle memory here because we're going to allow you to customize it and make it work more like how you're used to. And I think that's really good they're doing that. But I think it's it's a real shame it's taken them so long to actually do this. I'd recommend anyone who's got a sacrificial PC to have a look at the new context menu 'cause it actually works, I hate to say it, really well. They've done a good job, I think. It's just I wish they'd gone in that direction three or four years ago before inflicting their design choices on the public, which I think perhaps weren't that keen and they didn't seem to be able to listen to the reaction. Tobias Mann (23:32) How much of this is motivated by the lack of competition when Windows 11 launched? Macs were pricey, and Linux was an enthusiasts' game and filled with compromise. Five years into Windows 11, that situation seems like it's changed quite a bit with, memory prices having driven up the cost of every PC to a level that, unless you have a large supply chain that can help you avoid some of that cost, everything's expensive. Macs are a lot more competitive price-wise than they used to be because Apple's supply chain is just better than most PC manufacturers'. As well as on the Linux side of things, you know, there's Steam, which has basically made one of the one of the stalling points for Linux adoption with gaming a non-issue for the most part. And so it seems genuinely like Microsoft has to start investing in a platform that it thought was just kind of a monopoly. Richard Speed (24:54) There are echoes of Internet Explorer, where Microsoft owned the internet browser space until Chrome came along and ate their lunch. I think what's happened is they've looked at the adoption of Windows 11 and it hasn't gone well. Even though Windows 10 is now out of support – you have to be on the ESU program to get any support at all for Windows 10 – people are still using it, and people still like it. And I think what's happened at Microsoft in the last year is they've looked at this and thought, "Okay, we might have made a mistake here." Or the people in charge have refocused where they're going, thinking, "What's stopping people from moving from Windows 10 to Windows 11?" And yes, I think there's definitely more choice now than there used to be. Linux and macOS are good choices, and we're in this frankly bizarre situation now where Apple hardware is often the cheap option, which I still can't believe how we got here. But I think ultimately Microsoft is realizing that people liked how Windows 10 worked. Why did you change it? And so they're going back. My concern is, again, how long it's taken them to realize that. Windows is a supertanker – it takes a long time to turn that supertanker around. I think we've got another year or so before they get back to where they were, at which point they've lost four or five years of time, which is a real shame. But I'm happy to see what's happening. Avram (26:40) So I do wonder how many things that are old are going to be new again. For example, the wallpaper. I know this is really petty, but these are the things people look at when they go to your desktop. The default wallpaper in Windows 11 is hideous. And I think people look at that and feel like you're even ashamed that it's Windows because it doesn't say Windows on it. It used to have the Windows logo; you used to be proud of your Windows. Now they have some weird papier-mâché flower that looks like it's wilted. They got rid of some of the special effects that people liked – people liked Windows 7 because it brought the Aero Start menu. So I wonder whether we're going to start to see a new version of Windows like Windows 12, and people will like it because it will just bring back some of the visual elements and choices that people had in 7 and 10. Richard Speed (27:44) I think we'll certainly see more options to make this work more like it used to. In fact, today we saw an announcement in Microsoft's Message Center about Outlook, where you can make new Outlook look and feel a bit like Outlook Classic. It's a small thing, and it's cosmetic, but it's what people want. They prefer to have things work the way they're used to and happy with. There's no need to change things just for the sake of change. Avram (28:21) Well that's why Windows 11 has been such a pain because when it started, it was just change for the sake of change. Now we've got new features like AI piled on top of it. Was there really any evidence that we needed to have the Start menu in the middle of the screen, or was that just "Look, let's show you that we can make a different move"? Richard Speed (28:42) I think you're right. Things like rounded edges on window s –there were a lot of very cosmetic things with Windows 11. Over the last few years, I think Microsoft has come to realize that maybe things worked okay before and they didn't need to change this stuff. It just feels like a lot of wasted effort to get to this point. Tobias Mann (29:02) Is this any different than what we saw in the late nineties with the transition from Mac OS 9 to OS X? That was so divisive that Apple had to ship both operating systems on those machines, and for the initial period was actually shipping Mac OS 9 as the default. Or the transition from Windows XP to Vista, which I recall as being heavily influenced by the stiffer hardware requirements, reminiscent of the TPU requirements on Windows 11. It seems like this isn't as dramatic a departure from Microsoft's playbook in the past; it just seems like maybe users have lost interest or are struggling to adapt. Richard Speed (30:06) In this instance users probably have lost interest because Microsoft reached the point, certainly with Windows 10 and Office, where it did everything you needed it to do. So why would you need to upgrade? If I were cynical, I would say that with Windows 11 there was probably demand from hardware partners saying people need to buy new PCs. They were like, "Well, here's Windows 11." "Windows 10 hardware works fine, why should we buy more hardware?" "But you need to have this special security chip," and so users were then forced to buy new PCs or new devices. As you say, there wasn't a whole lot of difference, apart from the visual stuff, initially, between Windows 11 and Windows 10. Since then, there's been a lot more AI put into it for sure. But again, it's debatable whether there is a killer AI application on the desktop yet. Driverless cars are a great example of a really good AI application, but on the desktop, I'm not sure there is one yet. That is a problem Microsoft is also facing, with users saying, "There's nothing here that I absolutely have to have to do my day-to-day job." Avram (31:22) I remember last year I wrote an article about the things I think Microsoft should actually do to improve Windows. I'm sure they won't do any of them, but I was thinking of things that would actually be good for power users. I guess that's my bias here. Having a second and third clipboard built into the OS would be a big deal. Maybe adding another modification key in addition to Windows, Control, Alt, and Shift, because all the keyboard shortcuts seem to be taken up by something. We could use another key combination for more keyboard shortcuts. I feel like there are things they could do to really make this operating system more friendly to an audience that knows what they want and is getting more power-user heavy, because people have been using Windows for 30 years now. They don't need to make it simple; they need to make it powerful. And they're more interested in pushing ads and AI at you, then changing the UI just to make it look like they did something and get you to buy a new computer because they upgraded the hardware requirements. I feel like the reason people aren't switching to it is because they're annoyed. They're annoyed that they're being pushed to get new hardware that they don't want, they're annoyed that they're being pushed to deal with a slightly different UI when the old UI worked fine, and that they're having things like AI and ads pushed at them. But hey, speaking of AI being pushed at them, I understand that the new Task Manager is going to better keep track of when your AI is being used. Is that right? Richard Speed (33:10) Yeah, this is another thing Microsoft pushed quite hard this week: Task Manager. I can remember when Task Manager first appeared back in the days of Windows NT, and it was all very exciting. Dave Plummer wrote it initially and called it a stone-cold killer because it was great – you could just kill processes. It did one thing really, really well: you could see what was going on and kill stuff. Whereas over the years, and you could take this as a metaphor for Windows, it's just gotten bigger and does more things, which maybe you don't want. The latest thing is you can get more visibility into AI workloads. So as well as seeing how hard your NPU (neural processing unit) is working, you can now drill down and see which processes are using it. The goal there is to see: "I've got a process here that should be using the NPU, but it's actually using the CPU and GPU, and therefore burning up much more power. It's far less efficient." So it does make sense from that point of view. However, as I was reviewing this, I thought, "Is this really what I want Task Manager to do?" I want it to be a very lightweight, simple tool where I can see what's running, what's stalled, and what's crashed, and kill it. That's all I want it to do. If I want to see what's happening internally, I want to use a tool like Performance Monitor or an inspection tool to see what's going on in my system. I'm torn about this because it's always good to see these tools being developed and being enhanced, but I'm not sure it's the correct place to put that process monitoring information. I want Task Manager to manage tasks and that's it. Avram (35:00) I don't really have a problem with this personally. Richard Speed (35:02) (Laughs.) Avram (35:02) I like seeing what my CPU, GPU, and now NPU are doing, so put it in a place where it's easy for me to keep track of. That's fine, I don't have to look at it – I think that's a separate tab for the NPU, so you don't have to click on it. I don't see this as a huge problem as long as I'm able to see my tasks and kill my tasks. I want as much information as possible at my fingertips about what my computer is doing. Speaking of things being at your fingertips, I want to thank both of you for coming on and talking about this week's important topics. As always, we will have lots of information at The Register at your fingertips, talking about Windows, AI hardware, security, and all the other things that we do to be the hand that bites IT. We look forward to seeing you all next week. ®

Are you sick and tired of maintaining a password manager? Struggling with choosing the right one for you? Well, readers who live Down Under can get themselves down to their local AusPost branch where they can pick up an old-school alternative for just AU$4.90 (US$3.51). Password books are something of a historical relic, phased out largely because of the scrutiny associated with using one and the now-gray-haired elders who once scoffed at the mere notion of such an opsec crime. They might not be able to suggest a strong password for each of your many online accounts, and they won’t do you much good if your house is burgled, but you could argue that there is still value to be found in a pen-and-paper password vault. That’s exactly the conclusion drawn by the thousands of social media users who flocked to a post this week from one Australian who found stacks of password books for sale in their local post office. Small books are priced at AU$4.90 (US$3.51), while larger options will set you back a dollar extra. Granted, there are myriad issues associated with relying on a physical document for digital account security. For one, it’s a single point of failure. Lose the book or have it stolen, and it’s a painful road to restoring access to all your accounts. One could argue that if it’s kept inside the home and it’s stolen, then you’ve got bigger problems on your hands, but perhaps that’s not as true nowadays, when so many of our valuables are either stored or primarily accessed online. That said, it might bring a quick end to a hostage scenario – the type that is becoming all the more common as crypto wealth becomes increasingly common. Password books are also not as easily manageable as a modern password manager. The technological equivalent can auto-fill credentials, auto-update them if they’re found in public breaches, and suggest unique, strong strings to minimize the risk of compromise. You can’t store a passkey in a password book, either – a major issue now that the world is transitioning toward the new authentication standard. But using a password book no longer carries the same stigma as it once did among infosec types. The general consensus, gleaned from the hundreds of social media comments on the post, now seems to be that there's little wrong with storing passwords on paper at home. It’s certainly more secure than reusing the same weak password across multiple accounts, provided the book contains strong strings unique to each website. With the prevalence of infostealers nowadays, it's far more likely that crims will use a weak, reused, seldom-changed password to break into an online account than burgle a house to gain access to someone’s online banking. Plus, as many pointed out, it’s a much better route than writing passwords in a cloud document, which can be accessed by any device that has access to it – think Apple Notes, Google Docs, etc. At work, though, it’s probably best to stick to the password manager, the IT guys say. Mistakes by staff working at even the lowest rungs of the corporate ladder could lead to multimillion-dollar cyberattacks should that password book fall into the wrong hands. Pentesting consultants often send hired white hats to breach a company’s office and extract whatever value they can, sometimes through piss corridors. Such access can lead to malicious USB sticks dropping malware, bugs planted near the water fountain, and even someone stealing the password book from your desk drawer. Don’t believe it’s real? Security consultant Alethe Denis told us two years ago that’s exactly how her pentest team was able to surreptitiously extract corporate data over a company’s own Wi-Fi for over a week. They went dumpster diving, got the Wi-Fi creds, walked straight into a conference room, and deployed a data-stealing implant. In and out, all using physically stolen secrets. Helpful in the worst of times So, yes, password books contain plenty of potential pitfalls, Poignantly, however, they often prove invaluable in the event of a loved one’s passing. Having access to a password book, or at least some sort of plan to share passwords in the event of a death, is vital to ensuring family and friends have space to grieve without going through the arduous process of recovering an account through a platform provider, or via the courts. A slew of Redditors agreed, saying it made the whole process so much easier. One shared the tale of how their mother’s own special way of storing passwords resulted in a treasured family investigation. After password books spent years as outcasts of the cybersecurity world, they’re now having a second moment in the sun. And while the leading minds in cybersecurity are busy working on ways to stop phisherfolk from hacking into your accounts, or rogue AI agents from doing the same, there’s still something to celebrate in the safeguards of yesteryear, both in life and death. ®

Canonical's fondness for AI and Rust is no secret. Now it is co-funding a three-year PhD project investigating whether the former can translate large C codebases into the latter. Engineering veep Jon Seager announced the investment on Ubuntu's Discourse forum. The PhD project will be conducted at the University of Bristol's Programming Languages Research Group. So don't panic. This is not an announcement that Canonical will turn the bots loose to rewrite all of Ubuntu as Rusty slop. (For a start, nobody can afford that many tokens.) Instead of burning dosh on bots, Canonical will pay a proto-boffin to spend several years investigating whether the idea can be made to work. That is welcome. In an industry overflowing with hype, the project should produce evidence about whether this approach can be useful. The Reg FOSS desk interviewed Seager last year, and he struck us as sensible and pragmatic, but not lacking in boldness. Under his guidance, Ubuntu 25.10 adopted a Rust implementation of sudo as well as the entirely separate Rust-based uutils coreutils. The sudo command did hit some problems but they were quickly fixed. Both uutils and sudo-rs were pre-existing independent projects, however. They are human-written replacements designed to reproduce the functionality of existing tools using entirely new codebases. Certainly, those human developers may have studied the original source code – that's one of the good things about FOSS, after all – but these are new implementations. The new project will investigate whether an LLM can take programs "comprising hundreds of thousands of lines of C" and decompose them into smaller components before using an LLM to rewrite those components in "safe, behaviourally correct and maintainable Rust." Seager's post runs to just over 1,000 words and addresses several objections that sprang to mind. For instance, existing tools attempt something similar, but their results leave much to be desired. As Seager puts it: "Traditional source-to-source translators can process substantial amounts of code, but often preserve the structure of the C too literally. The result may compile as Rust, but still rely heavily on unsafe operations, retain awkward C idioms and require significant manual work before it resembles code a Rust maintainer would choose to own." We suggest reading the post before attacking the idea. It sets out a relatively detailed and measured plan. One admirable aspect is its acknowledgment that mature codebases contain knowledge their programmers never consciously documented. Years of fixes and patches encode responses to real-world corner cases that nobody anticipated at the outset. This is the key argument of Joel Spolsky's 2000 essay: Things You Should Never Do, Part I. Such knowledge is rarely documented outside the code itself or, if you're lucky, a few comments. A machine translation might preserve some of that behavior, while a clean human rewrite based on the original design could miss it. The proposal names two specific tools that the effort intends to examine: snap-confine and AppArmor We may be excessively cynical, but openSUSE 16 replaced AppArmor with SELinux last year. Outside the Ubuntu family, enterprise Linux has largely consolidated around the more complex SELinux, although Debian and several smaller distributions continue to support AppArmor. A hardened Rust implementation could therefore benefit AppArmor's remaining users. Snap, of course, has a narrower constituency still. This is not a solo Canonical project. The company is co-sponsoring it with UK Research and Innovation, a public body sponsored by the UK's Department for Business, Innovation, Science and Trade. Seager will oversee the project alongside the University of Bristol's Professor Meng Wang and Dr Cristina David. Three years is a conventional duration for a UK PhD – provided it does not overrun, of course. It's not as if someone could go into a PhD program in 1998 and then get two decades' worth of comic strips out of it or something. The Reg FOSS desk remains staunchly skeptical of generative AI outside the narrow domain of translation between human languages. As such, we have grave doubts that this will prove viable. We suspect the difficult part will not be translating the code, but decomposing a large codebase into smaller components that bots can digest. The challenge recalls the long-running effort to divide arbitrary algorithms automatically into tasks that can be farmed out to parallel processes. Decades of research have produced useful techniques for particular cases, but no general solution. It may yet prove to be an incomputable problem, like the Halting Problem – and as that article says, if it could be solved, it would lead to solutions to the Busy Beaver function or even Goldbach's conjecture. As with much of generative AI, more evidence is needed, and producing it is exactly what a PhD research project should do. We salute Canonical for putting real money behind the question and would be delighted to have our skepticism proved wrong. As we speculated in 2024, automatic translation between programming languages could become immensely valuable for improving software reliability – not by fixing problems automatically, but by exposing previously unknown errors. ®

OPINION DVD is one of those technologies that dominated for barely a generation yet left a permanent cultural mark. It made high-quality video cheap enough for mass distribution, popularized the TV series boxed set, and encouraged its gluttonous consequence: binge viewing. Streaming and capacious storage have since pushed optical media toward irrelevance. There may no longer be a box, but the boxed-set mentality remains, and we still binge like Roman emperors. It is a genuinely new way of absorbing large quantities of information. Now it is time to recognize a further evolution: the bingeable tech conference. Streamed conferences are nothing new, but few reward being watched from beginning to end. They offer too many talks on too narrow a subject, too much marketing flimflam, or too much variation in quality. The Goldilocks zone of quality, range, relevance, and intelligence is rare, but worth seeking out. This July delivered a near-perfect example: Software Should Work 2026. The name supplies the plot arc. Software Should Work is about reliability in its broadest sense, because reliable software requires much more than making code that runs. It encompasses abstraction and formal methods, managing complexity, making the most of tools and processes, individual innovation, team culture, and the nuts and bolts of telling computers what to do. In just 13 videos, SSW creates a widescreen, high-definition, thought-provoking picture of the field today. You can watch them in a day, and once you start, you may not want to stop. It would be unseemly to give too many spoilers. In such a concise treatment of a multifaceted subject, the talks that grab you will depend on your own interests and experiences. Consider dependencies, part of the calculus of reliability. So far, so dry. But Richard Feldman's talk compares web stacks from 1996, 2006, and 2026, revealing what looks like an inexorable multiplication of abstractions and dependencies. He then compares eight contemporary language and runtime websites: TypeScript, Node.js, Python, Ruby, Rust, Zig, Go, and Roc. The result is one of the more illuminating graphs you'll see this year. Another thought-provoking contribution comes from Richard Hipp, creator of SQLite, one of the corest of the core technologies in everyday use. On its face, his account of how a codebase evolves in response to changing test methods sounds dry. It is also a personal journey, full of highs and lows, illustrating how reliability depends not only on the interplay between code and test design, but also on human motivation, creativity, flexibility, and doggedness. There is no shortage of technical detail either. This is why SSW 2026 works so well. Its speakers have deep experience of software reliability, but were free to choose their subjects. The resulting mix of opinions, experiences, and approaches ranges from the mathematics of formal analysis to the importance of daydreaming and the dangerous myth of the superhero engineer. Reports from the language wars, AI WTFery, and arguments over whether Nix partisanship is justified are all present, delivered by people working on the front lines. Most importantly, the conference explores the culture of reliability and what it means to say that software works. It feels like a proper symposium: an exchange of ideas about something that touches every part of the digital world yet is rarely considered as a whole. Commercial pressures can work against practitioners and customers alike, while the lack of a common language and agreed standards makes good practice harder to defend. Without reliability, everything falls apart. How to teach it, and where the discipline goes next, remain stubbornly difficult questions. SSW 2026 provides a splendid and much-needed focus on all of this. Its blend of narratives and subjects means there is always something you did not know you needed to see, accompanied by the pleasurable sense of time well spent. Curiosity is richly rewarded. If you see software reliability as an ongoing drama – and if you don't, how are you still here? – block out a bingeing window. Software should work, but so should conferences, and this one is well worth the popcorn. ®

China’s Manned Space Agency (CMSA) on Sunday cancelled a planned Monday moonshot launch. The mission plan for Chang'e 7 calls for it to send a lander to a region near the Moon’s South Pole, in a location thought likely to contain ice in deep craters. The lander carries two other vehicles: a wheeled rover, and a hopper. China wants to test the latter to see if it can jump over crater rims to look for ice, and then hop out again and continue exploring. The lander is also important to China, because it’s equipped with tech to enable a precision autonomous landing – a tricky feat given the chaotic terrain moon mappers believe awaits it. China has built the lander, rover, and hopper to survive multiple Lunar nights, the stretches of two weeks or more without light, during which the mercury can dip below a hellish minus 200 Celsius in the shadows. The mission also involves an orbiter to help the ground vehicles send data home and carry out its own observations. Richard de Grijs, a professor of astrophysics at Australia’s Macquarie University, last week wrote that the mission is significant because it represents “a transition from simply exploring the Moon to determining whether its resources can actually be used.” “Chinese scientists won’t be looking for signs of water; we already have compelling evidence for lunar polar water,” he wrote. “The question at the mission’s core is where exactly that water is located, in what form and concentration, and whether it could realistically be accessed.” Sadly, we won’t begin to get the answers to those questions until next year. CMSA and Chinese state media on Sunday published a brief statement that says “Chang'e 7 mission headquarters had reached a conclusion that the lunar expedition cannot be carried out within this year's scheduled launch window as it had failed to meet launch criteria.” “The decision was made in accordance with the principles of prudence, reliability and absolute safety.” State media reports of the mission’s cancellation point out that launch windows that allow flights to the South Pole come along only once or twice a year, and that none is available this year. Some Chinese outlets report that bad weather was the reason for calling the mission off, a plausible cause as state media today warns of torrential rain ahead of a typhoon in Hainan, the location of the launchpad for the mission. It is unclear why CMSA didn’t mention the bad weather. ®

WHO, ME? Welcome to another installment of "Who, Me?" – the reader-contributed column in which The Register opens the working week with your tales of IT woe. This week, meet a reader we'll Regomize as "Pascal," who once worked for a company that had just acquired its first Exchange server, home to around 600 mailboxes. "The IT manager was keen we kept on top of the performance of the Information Store and suggested we take the box offline once a week to defrag it," Pascal told The Register. That sounded like a sensible idea, so Pascal wrote a batch file to stop the Information Store, run eseutil /d to compact the database, delete the temporary files created during the process, and restart the service "It tested OK and did make a difference to the size of the Store, so we scheduled it to run on the Friday night and left for the weekend," he told Who, Me? Pascal didn't make it far into the weekend before his phone started buzzing with reports that Exchange had stopped working. "I drove into work to check and found the Windows NT 4 server wasn't happy. Hardly anything would run, and the services for Exchange had stopped," he confessed to The Register. Pascal quickly spotted the problem. His script contained the command DEL *.* but did not specify a directory. Running from the System32 folder, it therefore began deleting files there. "My script wiped most of the System32 files, and that meant Windows NT 4 was lobotomized and the server was beyond repair," Pascal wrote. "It needed a full rebuild and restore, which took me most of the weekend." Looking back, Pascal accepts that it wasn't his finest hour. It was at least a lucrative one: he received double time to fix it. Pascal deleted System32 without deleting his career. Have you deleted something important by mistake? Before your brain erases the memory of your shame, click here to send your story to Who, Me? Perhaps we'll restore it in full on a future Monday. ®

AI datacenters have become a major driver of rising electricity consumption, but AI doesn't just happen in a datacenter. The telecommunications networks connecting datacenters to businesses and consumers are themselves major electricity consumers. 5G and its successors mean ever higher energy demands. Newer comms architectures deliver more bandwidth and richer services, but they also require denser cells, and 5G infrastructure such as multiple-input multiple-output (MIMO) antennas consumes more power than earlier equipment. The International Energy Agency predicts total datacenter energy consumption will double between 2025 and 2030 to 950 TWh, representing three percent of global electricity demand, while AI datacenter consumption alone will triple over the same period. Telecom industry electricity consumption is rising in parallel. The IEA put telecom network consumption at 260 to 360 TWh annually in 2022, up to 1.5 percent of global electricity use at the time, with mobile comms accounting for two thirds of the total. This all plays out against surging electricity demand in general, as sectors such as transportation and industry shift away from fossil fuels. As the IEA puts it, the age of electricity has arrived. Electricity consumption is now projected to grow at least 2.5 times faster than overall energy demand over the five years to 2030, it says. The relationship runs both ways, the IEA notes. AI may prove critical for global industrial innovation and competitiveness. The agency says that "proven applications of AI could help firms in energy-intensive industries reduce their energy costs by three to 10 percentage points." The agency adds, though, that "The energy sector as a whole is not yet taking full advantage of AI's potential... with lack of sufficient digital skills and data availability emerging as key barriers to adoption." That looks like a significant omission, given that traditional electricity generating infrastructure, in the shape of large central generators and national grids, is struggling to keep up with customer demand, particularly from the tech sector. Those same customers, ironically, already own substantial energy storage and even generating capacity of their own. Technology installations, particularly at the edge or in isolated areas, have always needed backup energy in the form of generators and batteries. These static resources are increasingly supplemented with renewable energy such as photovoltaic (PV) arrays and wind, which are critical, indeed the main option, for installations in remote areas beyond the grid. From idle assets to profitable energy nodes To date, those assets have remained dumb, sitting idle in wait of an emergency and, in an ideal world, never being used at all. Add smarter energy storage and AI, and that infrastructure could become a source of power in its own right, generate revenue for telecom operators, government, and enterprises and tech operators, and help stabilize the grid overall. The starting point is a full-stack energy storage system. As Kong Peng, vice president of ZTE Digital Energy, explains, conventional energy storage systems, have generally been assembled from components sourced across multiple vendors and reliant on heterogeneous comms protocols. This in turn leads to "compounded energy losses, inefficient joint debugging and ambiguous accountability". In contrast, ZTE has developed an integrated full-stack energy storage solution. Its core hardware includes a dual liquid cooling cabinet BESS (battery energy storage system) and a containerized BESS. The 261 kWh cabinet BESS is suitable for deployment at edge data centers and core sites. The containerized BESS can be cascaded to form a large-capacity energy storage system ranging from tens of MW to hundreds of MW and applied to AI datacenter, zero-carbon industrial parks, Grid-side independent energy storage and other scenarios. Around that is "native integration of core components including battery cells, BMS, EMS and PCS." (battery management system; energy management system; power conversion system). It also covers "maturing technologies" such as liquid cooling thermal control and cluster-level management. AI ties the whole thing together through ZTE's intelligent energy management system, ingesting data from the generation and storage infrastructure alongside external signals in real time, such as weather forecasts and energy prices. The VPP system tracks real-time electricity prices while AI algorithms work out the optimal charge and discharge strategies to maximize trading. Beyond creating revenue opportunities, the system stabilizes supply and makes better use of available PV energy. ZTE's complete energy storage system has an overall efficiency of over 90 percent, which can completely replace traditional backup power sources such as diesel generators in some scenarios. The infrastructure also opens up multiple revenue-generating possibilities. One route is "arbitrage via peak-valley electricity price differences, grid demand response and frequency regulation subsidies." Another involves using sites for local PV power absorption "to gain revenue from green energy certificates and CCER carbon trading." Or operators could turn to energy storage asset leasing for predictable long-term cash flow. The result is that operators, or rather the AI, can determine when it makes sense to pull energy from the grid, when to store it, and when to supply it back. ZTE also offers a one-stop service to get customers up and running, Kong says, "covering site survey, grid connection, construction, asset custody and carbon asset development." Adapting to the environment There is no one-size-fits-all approach for how operators can use the system to generate revenue. As Kong points out, Southern Europe is "abundant in photovoltaic resources, prioritizes integrated PV-storage base stations and scales up installations after verifying economic returns." Northern Europe, by contrast, "features volatile power prices and a mature frequency regulation market, where operators focus on revenue from energy storage auxiliary services." The technology is particularly relevant for telecom operators, who, as Kong puts it, have "massive base station resources with untapped load regulation potential, delivering win-win outcomes for both power grids and telecom carriers." This can include peak-valley arbitrage as well as "demand response and frequency regulation auxiliary services, substantially cutting electricity expenses." Using vacant land at sites lets operators generate power while aligning more closely with Europe's renewable energy policies. Datacenter operators also have substantial energy storage infrastructure and are increasingly looking at PV and other renewable, behind-the-meter energy options. Türkiye Telecom has put the technology to work in a 128MWp solar power plant covering 130 hectares at Sivas in central Anatolia. The plant uses N-type PV panels and 350kW inverters, and when complete will produce 196GWh of energy, which amounts to 15 percent of the operator's energy consumption, and cut carbon by 88,000 tons. This marks a crucial step in the cell operator's energy transition, and sets a replicable model for similar projects. Other telcos have adopted the technology. In Italy one provider is building out energy storage systems at its base stations and then accessing peak-valley arbitrage and demand-side response revenues, while also providing third-party storage for domestic and commercial customers. Other European partners are implementing the system in countries including Austria, Romania and Finland, amongst others. Not every site is viable, Kong explains. Regions with flat electricity prices offer little chance of arbitrage revenue, while stable grids reduce the need for backup power. Grid approval mechanisms, broader energy storage regulations, and carbon trading policies will all come into play. Operators must of course ensure they comply with EU regulations around grid stability, telecom infrastructure, and renewable energy obligations. They also need to make sure installations are in line with more mundane rules around construction standards and fire safety. This AI-powered integrated intelligent energy storage system is not limited to the telecommunication world. Its strong compatibility and intelligent capabilities, means it can be deployed in a wide range of industrial and livelihood applications. In industrial parks, the system connects with photovoltaic and wind power equipment. With the same AI algorithms able to optimize energy scheduling, and balance peak and off-peak power demand. It can help cut high electricity expenses, ensure stable power supply for precision equipment, and facilitate the zero-carbon transition of industrial parks. Likewise, in commercial buildings and urban complexes, it adapts to fluctuating building power loads to store energy during off-peak hours for electricity savings. Meanwhile, it can participate in grid demand response to create new revenue streams. Energy storage delivers a clear value in industries which operate in remote locations. Smart farms can adopt PV-energy storage integrated systems to address unstable and costly power supply, ensuring steady operation of agricultural equipment and underpinning the development of smart agriculture. In the mining sector, intelligent energy storage can replace energy-intensive traditional diesel generators. Tailored to the intermittent power consumption of mining operations, it can reduces costs and carbon emissions while securing reliable power for high-risk work. And in areas with weak power supplies such as off-grid villages, islands and remote islets, the integrated PV-Energy storage solution enables the construction of independent, clean microgrids, freeing them from reliance on conventional fossil fuel generation. Continuous advances in energy storage technologies, AI algorithms and supportive new energy policies are paving the way for large-scale adoption of all-scenario intelligent energy storage. Moving forward, the industry will focus on exploring technical adaptation, implementation challenges, profitable business models and compliant operation for diverse application scenarios. Though each operator faces a distinct environment, one thing is clear. We may be in the age of electricity, but electricity, at least from the grid, is not a limitless resource, partly because of the demands of AI. Combined with modern electrical infrastructure and renewables, AI can give telecom and datacenter operators the opportunity to do more than tick the sustainability box while offsetting soaring power prices: they can make their services more resilient, cut their power bills sharply, and help stabilize local and regional power ecosystems.

The long legal battle over ownership of Linux is closer than ever to ending, after a panel of three judges ruled a claim against IBM and Red Hat isn’t valid, and that time has expired for further action. This story starts in 1998 when IBM decided the world needed a single version of UNIX capable of running on multiple processor architectures. To make that happen, Big Blue allied with a company called the Santa Cruz Operation (SCO) which made a version of UNIX for x86 CPUs. Intel and Sequent also signed up for the multi-architecture effort, called “Project Monterey.” The alliance didn’t work, largely because Linux came along and delivered a *Nix that ran on multiple processors (and introduced a new way to develop software). IBM decided to add some of the code developed during Project Monterey to Linux, leading SCO and its legal successors to claim they owned that code and therefore had some sort of legal claim over Linux. That’s quite a potential prize given that Linux runs in billions of devices. To understand why, consider that Huawei owns patents that earn it $0.50 for every device that uses its Wi-Fi 7 IP. If it’s possible to score 50c for Wi-Fi alone, the royalties available from Linux could be rather more. In 2021, an heir to SCO settled with IBM for $14.25 million – a sum that reflects the fact SCO had for years failed to produce strong evidence to back its claims. Another of SCO’s legal successors, Xinuos, filed a new claim that IBM should be on the hook because Big Blue knew it did not own the code it contributed to Linux but instead had a non-exclusive license to use it. Xinuos argued that when IBM contributed Project Monterey code to Linux, it breached that license. Xinuos eventually took that argument to the US District Court for the Southern District of New York – and failed to convince it that IBM and Red Hat had a case to answer. Xinuos appealed, and on August 10th the United States Court of Appeals for the Second Circuit decided [PDF] not to revisit the District Court’s decision, agreeing that the original legalese governing Project Monterey means it’s too late to re-litigate the matter. The Appeals Court also agreed that Xinuos tried to frame the case as a licensing issue but failed, instead arguing that the issue was really about ownership. That ain’t all, folks, because Xinuos intends to file a petition to have the case re-heard by the full bench of the Court of Appeals. That hardly ever happens, unless the court finds significant errors or major legal issues that make a rehearing worthwhile. Law firm Kaplan says the Second Circuit has allowed reviews of less than 0.03 percent of the cases it has handled. So perhaps this matter is now close to a final resolution. ®

Developers recently burned by GitHub’s system outages should take note that there are other ways to manage Git at scale. One approach, recently put into action by Cursor, is to build the distributed version control system on object storage. A recent post from Cursor principal systems engineer Vicent Martí explains how the SpaceX subsidiary worked through its scaling issues with the notoriously fickle Git distributed version control system. The post explains how Cursor arrived at an architecture for its own Git-based repository service called Origin, which is powered by an internal engine called Continuity. A beta of the service is available with paid Cursor plans. “Agents have fundamentally changed the way we work with software, and in many ways they've made this situation worse. More code, more PRs, more CI runs. Version control is at the core of all of this, and it is possibly the hardest thing to change overnight,” Martí wrote. Martí speaks from experience, having worked at GitHub through much of the last decade when the company arrived at its own current architecture for managing Git. Synchronization is a bitch Git creator Linus Torvalds designed his software to work as a content-addressable data store, where all the objects are stored and indexed by the SHA-1 hash of their contents. Git sees a repository as a directed acyclic graph (DAG), with each commit being a node in a graph of nodes all connected by pointers. A Git server can look up an object directly by its SHA, but if it doesn’t have the SHA, then it “must actually walk the DAG step by step,” Martí noted. A client may just want to fetch or clone a consolidated packfile of the repository, or even get a list of recent changes, but to fulfill these requests, the server must traverse the entire graph only to assemble the necessary objects. Now, imagine providing such a service for over 400 million repositories, and you’ll get an idea of the scale at which GitHub operates (or struggles to do). After some fiddling about, GitHub engineers landed on what they called Spokes, which basically involves keeping at least three tightly synchronized copies of every repository on speedy NVMe disks. The approach became an industry standard, though over time its limits became apparent; the chief one is that the more replicas you make, the longer the synchronization takes. And Git does not play well with “eventual consistency,” Martí explained. Plus, these days agents bring their own mayhem. “When agents work with Git repositories at scale, they often operate outside of a monorepo by creating vast numbers of small repositories, many of them throwaway, and most of them barely touched,” Martí wrote. Object stores to the rescue When Cursor set out to build its own Git repository, it turned to object storage. Unlike file or block storage, object storage gives each chunk of bits its own unique identifier and files it with all the others in a single namespace (no directories). The most popular object storage today is Amazon Web Services’ Simple Storage Service (S3), which is increasingly used as a foundational layer for enterprise software such as databases, container registries and message brokers, thanks to its low cost, built-in redundancy and – for all practical purposes – endless scalability. With Origin, pushes are uploaded into S3 in a write-ahead log (WAL), capturing all changes as immutable objects. Wherever possible, changes are bundled together for faster throughput. Simultaneously, the pushes are written to the local “reference” copy of the repository (usually an NVMe disk). Once both actions complete, other replicas of the repository can download the changes as needed. “With the only requirement of having to synchronize the reference transaction with a single local repository instead of a quorum of replicas, we have a system that can ingest pushes as fast as our disk allows,” Martí wrote. Git will still have to do DAG traversal for many operations, but it is better to do it locally on a speedy solid-state drive than over a network. “Where does every repository live? The answer is ‘anywhere’. It doesn't matter! We treat repositories like a warm cache on disk, but the source of truth is always the write-ahead log,” Martí noted. We will see how well this approach plays out as Origin moves into being a production service. But if we don’t see stories about Origin outages, then Git managers will know to give object storage a serious look. ®