The Register

KPMG has offered staff facing redundancy £100 to waive their legal rights to appeal or to bring claims over the loss of their job, according to sources. As The Register revealed last month, the consultancy is laying off around 4 percent of the workforce in the UK, equating to some 200 humans being forced to look for new employment in an uncertain job market before Christmas. Staff within the AI, Cyber, SAP, and Testing teams are at the sharp end of the corporate job cutting exercise. On October 1, those at risk were told the consultation process is over and KPMG will inform them of their fate by email on October 9. In a message to staff, seen by The Register, KPMG said: “The collective consultation process has now formally concluded. The firm's original proposal remains unchanged. Employees will be notified on Friday 9th of October 11-11:30am by email if they have been selected for redundancy.” In addition to exit terms that some at KPMG previously described to us as insulting, The Reg understands the consultancy is offering £100 in exchange for employees' existing contractual confidentiality and other post-termination obligations. Some staff see it as a gag order. “The firm's original proposal remains unchanged,” one insider claimed, “offering marginally above statutory terms alongside a nominal £100 consideration for employees to waive their legal rights to appeal or bring claims. “The prevailing consensus among staff [is that] this approach is arrogant and disgraceful, particularly given the decision to dismiss personnel right before the holidays while withholding their earned annual bonuses,” the source added. Annual bonuses are discretionary, performance-based, and typically paid in January. KPMG has not communicated with staff on bonuses for 2026 but we understand anyone under notice (given or received) is not eligible for a payout. KPMG refused to answer questions The Reg asked about the £100 payment or the bonus situation. Instead, a spokesperson re-sent its earlier statement: “As our market evolves, we are adapting where we are focusing and how we are set up to make sure we have the right skills in place to best serve our clients. To respond to these market dynamics combined with low levels of attrition, we are proposing reductions in some of our advisory client-facing teams and will support our colleagues throughout this process.” The latest job cuts come six months after KPMG expunged 600 roles across the UK org chart, and is indicative of the financial pressures facing the big four accountancy and consultancy giants. PwC UK is merging two of its advisory businesses in response to the disruption caused by AI. ®

Sweden is the home of bork, and what better way to mark it than with a huge example of digital signage that shows a GRUB message instead of an ad? Spotted by eagle-eyed Register reader Alexander, the GRUB in question is sadly not the delicious meatballs for which a well-known Swedish furniture maker is famed, but rather the bootloader familiar to Linux fans, usually when something is working as it should (since we can't imagine much call for selecting from multiple operating systems in the world of digital signage). GNU GRUB (the GRUB is short for Grand Unified Bootloader), according to GNU, "is responsible for loading and transferring control to the operating system kernel software." In this case, the user has some options. UEFI firmware is one. Some options for DMS are another. DMS might mean Dank Material Shell, unless you know better. It's an attractive enough desktop, but it seems like overkill for something that slings ads for the few seconds Stockholm commuters raise their gaze from their devices. Some enticing possibilities are on offer. Hitting "e" will edit the commands while "c" will drop the user into the command line. That's assuming, of course, that the user is carrying a keyboard to attach to the stricken screen, something we're sure every Register reader has in their backpack, but not something civilians would necessarily keep on their person. GRUB has featured in this column before, appearing on a UK bus after a filesystem became a bit poorly (motion sickness can clearly affect computers as well as commuters), but this is a wall-sized example of the breed. Judging by the passengers on the escalators opposite, it appears that nobody is rushing to the aid of the stricken signage. That's probably for the best – leave it for a few seconds (by the looks of things), and it'll have another go at booting. Still, it's always good to see GRUB putting in an appearance in a section often dominated by Windows and Android whoopsies. The era of the Linux desktop might not be upon us just yet, but judging by this Stockholm signage, bork is alive and kicking in the open-source world. ®

OpenAI is now offering API customers the option to watermark the output of certain models in order to comply with the letter of Europe's AI Act. And within a few weeks, the biz plans to apply a form of digital watermarking – like it or not – to eligible ChatGPT and Codex text output in the European Union. But unlike rival Anthropic, it isn't applying its watermarking technique to content generated outside of Europe, though API customers can choose to have the marks applied to content made anywhere. Moreover, OpenAI is using technology that misses a fair number of instances of AI-generated text and can easily be tricked. The EU AI Act requires generative AI service providers to make model output identifiable in a way that's readable by machines. Anthropic was the first major AI company to announce its approach to AI Act compliance, which involves the application of Google's SynthID-Text algorithm. It's doing so globally. Microsoft and Meta have been working on similar technology for image-based AI because the EU is not the only market in which people have raised concerns about content provenance and the potential to use AI-generated content to produce misinformation. OpenAI already applies other content provenance signals to AI-generated images, audio, and video. In keeping with the letter of the AI Act, OpenAI is now offering a text watermarking approach called textGrain [PDF] that "adds an invisible statistical signal to the model’s word choices. Our detector looks for that signal to assess whether a passage contains an OpenAI watermark." Large language models predict word tokens one at a time from a probability distribution of potential words. By picking a different word to emit periodically – ideally a synonym – the resulting text deviates from the statistical pattern produced by an unbiased model. A watermark detector, which OpenAI is making available only to approved researchers and organizations, should flag that vocabulary divergence. That's the theory, but in practice results vary. The target error rate is one percent, but in short passages of 200 words, the detector may only flag 80 percent of the watermarks, compared to 95 percent in a 400-word passage. The results are also worse in texts that are functional, like passages about math or code, as substitutions are more likely to introduce errors. The textGrain paper does not address the likelihood that a statistically altered word choice might alter the meaning of a given passage. That's a possibility if the algorithm were to alter some salient fact or figure. Anthropic's approach, which also replaces certain output words though using a different algorithm, brings similar risks. In any event, OpenAI's watermark is easily defeated by word substitution. In tests on 400-token passages, swapping out 10 percent of the words with synonyms reduced watermark detection from 92 percent to 66 percent. And when replacing 25 percent of words, the watermark signal could only be detected in 17 percent of cases. Now witness the firepower of this barely adequate but compliant AI detector… ®

When yellow dwarf stars like our sun die, they swell to an enormous size and destroy their inner planets by either crushing them with massive gravity or evaporating them with astounding heat. While these suns later settle down to become a white dwarf, the process is nearly always game over for planets in an inner solar system. Or maybe not, according to a paper published on Monday titled “Discovery of a second-generation planet candidate accreting onto a white dwarf.” As explained by NASA’s Hubble Mission team, the paper’s authors pored over Hubble Space Telescope observations of white dwarf star HS 0209+0832 from 1999. According to the paper, data from those scans “revealed the presence of carbon, aluminum, silicon, calcium, titanium, nickel and zinc in the atmosphere, as well as ~100 lines that could not be identified.” The authors of this paper used new techniques to identify the mystery matter, and concluded that there’s plenty of the element niobium present in what looks like a gas giant orbiting HS 0209+0832. In NASA’s post, co-author Nicholas Stone, a theoretical astrophysicist at the University of Wisconsin, pointed out that “Niobium and other elements heavier than iron are astronomically special because, unlike many common elements, they are not formed in the cores of stars by thermonuclear fusion.” “Instead, these heavy elements can only be synthesized in the exotic conditions that briefly emerge inside dying stars. The presence of niobium is a signpost of these ‘death’ throes, and the expulsion of the dying star's innards into space.” Finding niobium is also notable because, as the paper explains, “the canonical rock-forming elements [are] silicon and iron.” The paper explains that its authors compared the material observed at HS 0209+0832 with all known examples of rocky material around white dwarfs, and to all meteorites in our Solar System, but could find “no compositional matches.” That finding suggests that the niobium-rich gas giant at HS 0209+0832 represents both a secondary planet – which are rare – and a body composed of material we’ve not previously seen becoming a planet. It also shows that planets can form close to a star after the hellish red giant phase. Oh, and let’s not forget that HS 0209+0832 is also an exoplanet, a class of body humanity first spotted in only 1992. The paper suggests the niobium gas giant is about the size of Jupiter, but that the white dwarf is blasting away its atmosphere and perhaps creating “a comet-like tail of material that would form a disk around the white dwarf star and fall back on to its surface, leading to Hubble detecting the niobium when studying the star” Co-author Jamie Williams, from the Department of Physics at the University of Warwick, told NASA he thinks the niobium world will stick around. “If the second-generation planet is there, I think it is likely to survive. Eventually the white dwarf will cool and then maintain a consistent temperature, with the planet in its stable habitable zone for millions of years,” Williams said. ®

Atlassian has told its users to patch its datacenter products, pronto, to prevent attackers accessing their files. The Australian collaborationware company on Monday sent users an email that opens with the words “Action required” and points to a security bulletin that explains CVE-2026-21589. The 9.3-rated arbitrary file access vulnerability is present in the datacenter versions of its Bitbucket, Confluence, Jira Service Management, Jira Software, Bamboo, Crowd, Crucible, and Fisheye products. Atlassian says the vulnerability “allows an unauthenticated attacker to access specific files within the web application root directory in affected versions.” That’s scary because Atlassian warns “In some configurations, there may be sensitive files present that increase your risk.” There’s also some good news in that attackers must know the exact filename and path to exploit the vulnerability, and the mess doesn’t allow anyone to see the contents of a directory. Another piece of good news is that Atlassian has updated its products – so users only need to find a change window in which to upgrade to a safe version of their software. Atlassian advised those who can’t patch ASAP to remove their instances from the internet, if possible. “Instances accessible to the public internet, including those with user authentication, should be restricted from external network access until you can take action,” the company warned. Its advisory also includes mitigations and advice on how to determine if your instances need the fix. Users who made the move from datacenter products to the Atlassian cloud have nothing to do, as Atlassian fixed the flaws in its own SaaS. That state of affairs rather vindicates Atlassian’s 2020 decision to stop developing its low-end server products and require users to shift into its cloud, and last year’s sequel in which it decided to discontinue its datacenter software, too. Atlassian admitted it hasn’t made that migration easy, because it somehow released a lift and shift tool that was worse than an earlier version. In March 2026, Atlassian axed ten percent of staff. The company’s share price was on a year-long slide at the time, as pundits suggested it might fall victim to the SaaSPocalypse, a theory that AI would replace business software. The price of Atlassian scrip has tripled since then, suggesting investors are more confident the company’s plan to use AI to power workflows represents a moat LLMs cannot cross. ®

Linux KVM, the hypervisor favoured by hyperscale clouds, apparently has a full VM escape bug. That nasty news came from security researcher Paulos Yibelo, who on X shared a screenshot of a bug bounty award he won for discovering what he described as “Full VM escape zeroday (guest>host root in industry standard hypervisors)!” The bug bounty Yibelo participated in is run by Vercel, a company that provides MicroVMs as sandboxes for AI agents to work inside. The company’s Sandbox uses Firecracker MicroVMs, a technology created by AWS, which relies on Linux KVM – the kernel level hypervisor in Linux. Vercel CEO Guillermo Rauch named KVM as the hypervisor identified by Yibelo. “We’ve confirmed a KVM 0day through our Vercel Sandbox bounty program. Affecting the industry’s gold standard solution for Linux virtualization,” he wrote. And that’s all the info that has made it into the public view at this time. The Register can find no chat on relevant mailing lists. We have asked Rauch and Yibelo for additional details. Hopefully, we don’t hear from either of them for days or weeks, for two reasons. One is that guest-host escapes are the nightmare virtualization scenario because they mean whoever runs a guest VM could take over an entire server, and perhaps gain the ability to control other guests. The other is that KVM is astoundingly prevalent: AWS and Google both use it to power their public clouds. Enterprise virtualization players Nutanix, HPE, and Proxmox also rely on KVM. And of course KVM is also in Firecracker, which is open source and could therefore be running in all sorts of places. Whatever Yibelo discovered therefore very much needs a responsible disclosure process, because if hints about the flaw emerge it could allow attackers to do a lot of damage. Once a fix is found, the next question is whether implementing it will require disruption or downtime. It’s possible to hot-patch KVM, and to migrate live VMs from vulnerable hosts to machines running a patched version of Linux. Hopefully those techniques will work. This might be the second nasty bug discovered in KVM this year, after the so-called Januscape flaw. Beyond the potential risks this bug created, observers have suggested the potential seriousness of the flaw means Yibelo’s reward should exceed the $50,000 available under Vercel’s bug bounty program. ®

Schneider Electric's market cap has more than doubled in the past four years as the AI boom has driven unprecedented demand for datacenter power and cooling infrastructure. Now the French multinational is cashing in on its good fortunes to cement its place as a key provider of industrial power and datacenter infrastructure with the $22.6 billion acquisition of American industrial software company PTC. Founded in 1985, PTC made a name for itself developing software for computer-aided design, plus lifecycle management tools for industrial products and applications. The all-cash deal, announced on Monday, positions Schneider Electric to capitalize on the proliferation of datacenters used to train and serve AI models at scale. Since ChatGPT kicked off the AI arms race in late 2022, datacenter design and power requirements have changed considerably. Forty-kilowatt air-cooled racks have given way to 250kW liquid-cooled systems that need much denser power delivery and thermal management systems, all of which need to work together. In light of the growing complexity and supply chain challenges facing datacenter deployments, Schneider has made several strategic acquisitions which expanded its role in bit barn development. In late 2024, not long after Nvidia unveiled its first all-liquid-cooled rack designs, Schneider invested roughly $850 million on a 75 percent stake in Motivair, one of the leading producers of cold plates, rack manifolds, coolant distribution units (CDUs) and rear-door heat exchangers required to support the widespread adoption of liquid-cooled servers. The company is expected to acquire the remainder of Motivair sometime in 2028. As Schneider's role in the datacenter has grown, users of its kit have more products to manage across their lifecycles. Buying PTC expands Schneider's lifecycle management offerings and introduces key design capabilities to its offerings. PTC isn't Schneider's only datacenter-centric acquisition this year. Back in June, the company paid $3.1bn for industrial data operations vendor Cognite. Speaking to investors on Monday, Schneider CEO Olivier Blum suggested that owning PTC and Cognite would enable the company to inject software-defined automation into power systems early in their development, improving their efficiency in the process. Datacenters are one of Schneider's initial targets. "We believe that at the design stage, it is super important that you can help your customer to design both product, machine, processes and energy system. This is where we see CAD engineering as a very critical part of the portfolio," Blum said, according to a transcript of his remarks to investors. That's assuming of course that Schneider doesn't run into any regulatory hurdles along the way. The deal, Schneider's largest ever, isn't expected to close until late next year. ®

The attackers just keep coming after Citrix NetScaler appliances, as both the feds and private security researchers warn that someone found - and has already exploited - yet another Citrix bug before it had a patch. This latest vulnerability, tracked as CVE-2026-88779, is a memory overflow bug that leads to denial of service attacks. It only affects NetScaler ADC and Gateway appliances configured as a SAML (Security Assertion Markup Language) service provider or identity provider, used for single sign-on authentication. Late Friday, amid exploitation reports, Citrix confirmed that it was investigating a “newly observed issue related to SAML authentication in customer-managed NetScaler deployments.” By Saturday night, the vendor released a security advisory for NetScaler ADC and NetScaler Gateway with patches, urging vulnerable customers to “install the relevant updated versions as soon as possible.” Citrix also posted a blog about the vulnerability, confirming that it has observed targeted attacks on unmitigated NetScaler deployments that can lead to denial of service. Citrix did not answer our questions about CVE-2026-88779 - including how many instances have been affected and what attackers are doing after exploiting the bug - but urged customers to "quickly apply" the fix to NetScaler instances. "We were recently alerted to a new issue that affects service availability for some NetScaler deployments," a Citrix spokesperson told The Register. "After we were alerted to this issue we immediately developed and published a mitigation while concurrently developing, testing and deploying a fix." On Sunday, the US Cybersecurity and Infrastructure Security Agency (CISA) confirmed CVE-2026-88779 was under active exploitation and ordered federal agencies to patch the bug by Wednesday. While the new vulnerability is not technically related to the earlier eight CVEs finally disclosed by Citrix on September 27 - weeks after miscreants began abusing two of these security holes (CVE-2026-88772 and CVE-2026-88771) - watchTowr researchers told us they suspect it has been used to purposefully crash machines, making exploitation of CVE-2026-88771 faster. “This vulnerability is incredibly simple to trigger, with a single specially crafted request being all that is needed to knock an appliance offline,” watchTowr’s head of threat intelligence, Jake Knott, told The Register. “Exploitation is already occurring in the wild, and disrupting an authentication gateway can prevent legitimate users from accessing the services behind it.” WatchTowr reproduced the vulnerability on Friday, and Citrix credited the attack-surface management company along with Bishop Fox with helping it address the issue. “Citrix provides an indicator-of-compromise script that teams can run to check exposed appliances for signs of compromise, though a clean result is not definitive proof,” Knott said. “Security teams should prioritize appliances configured as a Gateway or AAA virtual server with SAML authentication enabled, and affected organizations should apply the fixed build or Citrix’s interim mitigation if an immediate upgrade is not possible.”®

While OpenAI and Anthropic stoke fears over rogue agents, Canadian model dev Cohere promises enterprises a safer and more consistent experience. On Monday, the company said it revamped its North agent harness, which launched early last year, to improve collaboration, access control, token spend, and overall security of the platform. As you'd expect from an enterprise focused agent framework, Cohere's North 2 is designed to tie into existing knowledge bases and SaaS apps in order to automate various business processes. North 2 introduces a variety of features and functions that should be familiar to anyone who's used Claude Code or OpenClaw. Among them is the concept of skills, libraries, and automations. Skills are arguably the simplest, and are essentially standard operating procedures documenting how the agent harness should complete tasks. The skills are often generated by working with the agent to complete a task and then documenting which steps achieved the best results. With North 2, Cohere says that it has made it easier for these skills to be shared so that users aren't wasting time reinventing the wheel. Presumably, this should improve consistency by avoiding having the agent take different routes to achieving its goal, something that might miss key information or nuance. To facilitate the creation of new skills, North 2 also introduces the concept of automations, generic templates that can be customized using a drag-and-drop graphical interface, not unlike LangFlow, which we looked at earlier this year. Alongside skills and automations, Cohere's North 2 does something similar for shared knowledge and assets, which it calls libraries, to ensure that agents are working off of the same datasets. Finally, North 2 adds memory, a core feature of most modern agent harnesses, to ensure that agents can maintain context across multiple sessions. In some sense, these features are a bit like self-generating macros, but rather than being limited to Excel, they work across multiple applications and data sources. Sovereign AI remains a major talking point for Cohere, which says enterprises can deploy North 2 where they like and use their preferred models. If enterprises want to deploy the harness on-prem in an air-gapped environment, they can. If they prefer to run it in the cloud, that's also an option. Along with new features aimed at collaboration and automation, North 2 implements new guardrails designed to help enterprises get a handle on their token spend and access controls. Flow control functionality, meanwhile, promises greater visibility into which departments are hogging all the tokens, while access control functionality offers admins the ability to assign access control lists to various libraries and skills so that interns can't get their hands on proprietary data just because they asked for it. North 2 is only the latest in a slew of enterprise focused agentic AI platforms which have been rolled out over the last year. Essentially every major and minor cloud provider has introduced some kind of agent harness tailored to enterprise needs. The competition has only gotten hotter over the last few weeks as OpenAI, Meta and others have set their sights on the enterprise. ®

Apple’s attempt to integrate AI into its various operating systems hasn’t been entirely well received, nor has the decision to make Apple Intelligence increasingly difficult to disable completely. Thankfully, the open source community has come to the rescue. A recently released open source tool called RemoveMacAI does pretty much exactly what it says on the GitHub-hosted tin: Aggressively rips Apple Intelligence out of macOS, disabling most of its AI features and removing downloaded models following the launch of macOS 27 last month. Siri AI? Gone. ChatGPT and other AI extensions? Eliminated. Genmoji, random boxes offering to jazz up your writing, summaries of notifications and app content, and automatic photo clean-up? Never to be seen again. More importantly, RemoveMacAI can claw back the storage occupied by Apple’s AI models - potentially over 12 GB, depending on the Mac. You might just get confused and think it’s 2025 again. The impetus for RemoveMacAI, its creator Om Lahorey said in a LinkedIn post, is the fact that Apple made it practically impossible to remove Apple Intelligence features in macOS 27. “The settings are spread across about a dozen places, some of them only under Screen Time,” Lahorey wrote over the weekend. “Even with every one turned off, the models stay on disk. I wanted it gone in one step, so I wrote RemoveMacAI.” Unsurprisingly, Lahorey’s tool blew up on Hacker News, with plenty of people happy for an easy way to get rid of Apple AI tools. As of writing, the repo has over 1,500 stars, so it's clearly struck a nerve with frustrated macOS users. It’s simple to run, as its GitHub page explains: install it via a curl command to its GitHub repo, or with Homebrew, and then fire it up in the terminal. Selected features can be left enabled, or Apple Intelligence can be disabled completely, while a single command can revert the changes and restore the previous settings. RemoveMacAI works by installing a configuration profile, which the user has to approve, that applies Apple’s own restriction keys for Apple Intelligence features. Models are then removed via Apple’s asset service. Lahorey credits another open source tool called pared, and its developer, with mapping the asset service, model sets, and settings keys needed to remove Apple Intelligence. Pared functions similarly to RemoveMacAI – you can consider RemoveMacAI to be a more user-friendly wrapper for the other tool. As for whether it will persist, Lahorey explains that the config profile RemoveMacAI installs redirects the download for Apple Intelligence models to a closed local port, so updates to macOS shouldn’t result in the models being reinstalled. Also worth mentioning, Apple System Integrity Protection remains enabled with the models disabled and uninstalled, as no files under System are directly modified by the tool. Dictation should still work as well, as speech models aren’t removed by RemoveMacAI. That said, it’s certainly possible a future macOS update will break RemoveMacAI’s functionality. So if you’re interested in giving the tool a shot, there’s no guarantee this particular escape hatch will remain open forever. ®

The FBI and law enforcement partners have arrested “multiple” suspects as part of an investigation into a September hack allegedly involving data-theft-and-extortion group ShinyHunters, the bureau told The Register. “The bureau continues to aggressively investigate the recent cyber incident allegedly involving ShinyHunters, having already worked with partners to arrest multiple subjects and we will spare no resource in bringing each of the responsible individuals to justice,” an FBI spokesperson told us in an email on Monday. The FBI declined to comment on the specific arrests, including that of Saif al-Din Khader, a suspected ShinyHunters member who has reportedly been detained in Jordan, according to Reuters. Khader, who goes by the alias Rey, was reportedly brought into custody on September 29, and is said to be cooperating with the FBI to identify other members of the group. ShinyHunters did not respond to The Register’s inquiries about the arrest and Khader’s alleged involvement with the criminal group. Rey confirmed his real identity - Khader - to security journalist Brian Krebs last year. Krebs described Khader as the “technical operator and public face” of Scattered LAPSUS$ Hunters. “Rey got picked up finally,” said security sleuth Kevin Beaumont following news of the arrest. According to Beaumont, Rey was “one of the kids who got into JLR.” The Jaguar Land Rover (JLR) breach, which occurred in late August 2025, affected the company's IT systems and halted manufacturing operations. Dealer systems also went down, and suppliers faced canceled or delayed orders. In addition to crippling the carmaker’s business operations for months, the digital thieves stole personal payroll data belonging to thousands of JLR employees. The cyberattack, one of the most costly in UK history, was attributed to Scattered LAPSUS$ Hunters. Khader’s detention came two weeks after the Dutch National Police arrested a 24-year-old whom the FBI described as “one of the alleged leaders of ShinyHunters.” While Dutch cops have not named the suspect, Krebs and other reports say he is Pepijn van der Stap, who was convicted in 2023 for hacking and extorting numerous organizations and was on supervised release after three years in prison. Van der Stap also worked as a software engineer at the Amsterdam-based cybersecurity startup Hadrian and volunteered as a security researcher at the Dutch Institute for Vulnerability Disclosure (DIVD). In a video message following the arrest, Brett Leatherman, assistant director of the FBI's Cyber Division, had some advice for the “remaining members” of the data theft and extortion gang. “Arrests have a way of changing who is willing to talk, and seized infrastructure has a way of showing us who's left,” Leatherman said last week. “The longer you stay in this, the more we learn about you. You know how to find us, and we know how to find you. I suggest you reach out first while the choice is still yours.” At the time, the FBI declined to answer The Register’s questions about the video message, including whether it had seized any of the cybercrime group’s infrastructure, and whether any of ShinyHunters’ members had taken Leatherman up on his offer to “reach out first.” In late September, ShinyHunters hacked the FBIJobs.gov portal and claimed it stole sensitive personal details about current, former, and prospective FBI employees. A spokesperson told The Register that unlike most of its digital break-ins, this one was “NOT financially motivated … We want the FBI to correct or retract their statements they made, which included substantial false allegations.” Later, in an exclusive interview, the spokesperson told us the attention-grabbing hack was “fundamentally a public relations and marketing initiative for our business.” ®

Microsoft Windows 11 26H2 has arrived, and, unfortunately, it's already picked up a known audio issue within weeks of release. The glitch also affects Windows 11 25H2 and 24H2, which share the same code branch. It relates to decoding AC-3 (Dolby Digital) audio and can cause the application using the decoder to crash abruptly or fail to start at all. Identified after the September 22 Windows update, the issue mostly impacts older applications, according to Microsoft. "Many modern applications do not use Windows for this type of encoding and therefore will not encounter this issue. "The code component involved in this issue is most commonly used in legacy applications that still rely on Windows built-in audio decoding." In summary – if your app has the temerity to use Windows built-in audio decoding for AC-3, you might have a problem. If it includes its own audio decoding component, you should be okay. "Affected applications may vary," explained Microsoft, "and can include games, media players, and certain productivity applications." There's also no fix as yet, so uninstalling is the only way to address it for now. To be clear, the AC-3 malfunction is not specific to Windows 11 26H2; the update preceded the release of the operating system's latest version by a week. It joins other known issues in 26H2, including other ongoing audio woes that have lingered in the list despite workarounds being available from Microsoft Support for Business. Microsoft currently has no mitigation for the AC-3 bug and offers little comfort to affected users: "We are working on releasing a resolution for this issue in a future Windows update." Acknowledging the latest trouble continues a run of misfortune for Microsoft. September's Patch Tuesday resulted in an out-of-band update to deal with RDP and Hyper-V issues. The update did not, however, fix all the audio problems it introduced, and the September 22 preview update has compounded the situation. Microsoft did not provide a timeline for resolving the latest issue. ®

Sam Altman-backed AI chip startup Volantis has a plan to break through the memory wall using light, and this time it's not just an IP play. Founded in 2022, the San Francisco-based startup is developing an AI accelerator that will use a photonic interposer to dramatically increase memory bandwidth without compromising on capacity. Volantis believes it can pack upwards of 10 TB of memory at speeds up to 240 TB/s onto a single massive package. This, it claims, should be enough to serve a 20 trillion parameter model at 10,000 tok/s per user. To do this, Volantis is taking a somewhat unusual approach. Rather than developing some novel compute architecture or memory technology, the company is focusing its efforts on the optical interposer. Everything else it plans to license from IP providers. "The differentiation is primarily coming from the optical interposer. So why don't we license absolutely everything we possibly can, even if it's not the highest performance in the world, because we're already biting off a lot of risk with the interposer and the integration," CEO Tapa Ghosh tells El Reg. The basic idea behind the chip, which Volantis is calling A-1, is that by using optical wave guides rather than electrical interconnects it can dramatically increase the amount of high-bandwidth memory per package. Because of the speeds at which most AI accelerator memory, like HBM, operate, and the size and number of data pins used, it needs to be as close to the compute as possible to minimize power consumption and maximize signal integrity. As a result the maximum capacity and bandwidth of modern GPUs is limited by the compute die's shoreline. That's the physical perimeter around the chip. For now, that means the cap is right around 432 GB on AMD's MI455X. Photonic interconnects can extend that much farther. Instead of a few millimeters the memory can be located a few centimeters away, eliminating the shoreline as a bottleneck. While HBM is useful as an example, we'll note Volantis hasn't said what memory tech its chip will use just yet. Volantis aims to achieve this using optical waveguides integrated into an interposer. Compute and memory designed or supplied by its partners will be packaged on top. If any of this sounds familiar, Volantis isn't the first to propose using silicon photonic interposers in this manner. As our sister site Next Platform has previously reported, both Lightmatter and Celestial AI have toyed with the idea in the past. In fact, we strongly suspect the idea of HBM-at-a-distance contributed to Marvell's decision to acquire Celestial for $5.5 billion. However, unlike either Celestial or Lightmatter, Volantis isn't trying to license its tech to other chipmakers. At least not yet. Instead it aims to develop a fully fledged inference chip. "It is, at least on day one, a way easier problem to demonstrate a solution that ships," Ghosh says. Another differentiator is the technology on which the startup's interposer is based. Rather than relying on external lasers, Volantis' design uses micro vertical-cavity surface-emitting lasers — micro-VCSELs for short — and means the light source is integrated directly into the interposer. VCSELs aren't new by any means, but have faced criticism over their reliability. However, Volantis CTO Roy Meade tells us the technology has come a long way in recent years and by giving up some speed, significantly higher reliability can be achieved, a philosophy he describes as wide and efficient. "People are still constrained to this, what I would call a fat pipe point-to-point mentality, and naturally, if that's the case, you're going to go towards SerDes," he says. "If we were to go back and take a look at the benefits of a wide-and-efficient approach. HBM for example. Taking that same playbook makes a lot of sense." What Meade is getting at is that interconnects can achieve higher bandwidth in a couple different ways. The first is by increasing the link speed, and the second by using more links. For example an 800 Gbps link can be achieved using four 200 Gbps links or eight 100 Gbps links. In his example of HBM, higher bandwidth is achieved by using an extremely-wide bus but slower data rate per pin relative to other memory technologies. The same concept applies to Volantis' optical interposer-based accelerator. Power efficiency remains an open question that's dogged many photonics startups, but this may be less of a problem for Volantis because it's not trying to license its first-gen interposer tech. For its first product the company is targeting 1 picojoule per bit per 24 Gbps lane and a total system power of around 20kW. Volantis still has a ways to go before it tapes out its first production-ready chip, and will need to raise a fair bit more in venture capital to do that — far more than the $88 million it raised this week with the help of Lachy Groom, Abstract Ventures, and others. Ghosh is well aware of this. He says the funding will enable the company to scale from small-scale optical demonstrators to something more closely resembling the final product. The interposer alone is challenging enough, but then there is the complexity of actually integrating the IP they license. "Aggressively, we should aim for an MVP product in 12 to 18 months as our target," he tells us, noting that's to have a working prototype not volume production. ®

After adding text-based ads to its chatbot earlier this year, OpenAI announced on Monday that it was adding a new “visual ad format” to ChatGPT, which is precisely what it sounds like: Visual ads, like so much of the advertising content online, will now be part of the ChatGPT experience. OpenAI will initially test these ads during image generation. “Through images showing product inspiration, product usage, or the experiences they make possible, people can discover a new brand, understand what makes an offering distinctive, and decide what to buy,” OpenAI explained. “For advertisers, this creates new opportunities to tell their story through images.” If you build it, someone in the ‘States will inevitably find a way to cram an ad into it, which is why testing of OpenAI’s new ad format will begin sometime in October in the US, given its status as the advertising hellscape capital of the world. Visual ChatGPT ads will be initially tested during image generation in ChatGPT, though like much of the scheme it’s not particularly clear how that’ll work from OpenAI’s announcement. The company didn’t detail much of anything about the process, so ads might appear as a de facto loading screen while ChatGPT conjures up whatever visual slop users ask of it, or they might remain on screen after the image is generated. Or not - OpenAI told us it would have more to share once the ads enter testing. It’s also entirely unclear whether the ads will be generated by ChatGPT on the fly or pulled from a library of ads submitted by companies. OpenAI also didn’t say how it will target visual ads, which has one advertising firm CEO questioning whether the new format will be worth the expense for companies looking for captive eyeballs. "Search ads work because they meet people at the moment they're looking for something,” Donatas Smailys, CEO of creator marketing firm Billo, told The Register in an email. “Someone generating an image in ChatGPT is not shopping." OpenAI has explained how it matches ads to ChatGPT conversations, Smailys added, but given it hasn’t mentioned how it plans to target visual ads, he remains unconvinced that they’ll be an effective advertising tool. That, and they’ll be intrusive - it’s bad enough getting targeted text ads or being invited to chat with an advertising AI after clicking an ad - now visual clutter will be part of ChatGPT, too. Free users can opt out of ads without paying, but doing so disables tools including image generation; if you want to generate images without ads, Plus starts at $20 a month. ®

The age of LLM-assisted security vulnerability discovery is really starting to bite. Debian's latest Linux kernel security update comes with a formidable reading list of 1,313 CVE identifiers. The DSA-6528-1 Linux security advisory, published on September 29, covers kernel package version 6.12.111-1 for Debian 13, codenamed Trixie. Debian 13.7 was released on September 12, before upstream kernel 6.12.111 arrived nine days later. The Debian security tracker links to descriptions of the individual issues. We have not examined every entry. If we had, this article would not appear until after Debian 13.8 (which is likely to appear later in October), or possibly at some point in 2027. Several checked at random also affect older kernel versions, so the list should not be read as a tally of bugs introduced in 6.12.111. The Linux kernel project became a CVE Numbering Authority (CNA) in February 2024. Back in February this year, kernel maintainer Greg Kroah-Hartman described the Linux CVE assignment process in some detail. He said kernel development averages around nine changes an hour, with a feed of known bug fixes averaging about 30 changes a day providing the basis for the CNA team's review. The kernel team's policy is to assign CVEs automatically after fixes have reached a stable kernel tree. It takes a deliberately cautious approach because the security implications of a bug may not be apparent when it is fixed. A CVE identifier alone therefore says little about severity or exploitability. We strongly suspect that this number of CVEs is due to LLM bots doing the bug hunting, and quite possibly doing the bug fixing as well. Linux is not an anti-AI project, and neither is Debian. AI-assisted bug hunting is already swamping the Linux security mailing list, as The Register reported in May. Kroah-Hartman released kernel 6.12.112 on October 3. Its detailed changelog runs to more than 27,000 lines. With both the rates of change and the sizes of the changes getting so large, it is hard to deny that LLM bot assistance must be very useful to the hard-pressed maintainers. Whether coding bots constitute a net benefit to the projects, to software, or to humanity as a whole remains at best an open question. ®

There is a question about AI data sovereignty you would ask HPE or NVIDIA if nobody was recording. It is usually specific to your own workload and your own regulator, and it usually ends with whether the sovereign option on the slide survives contact with a procurement team. On Thursday 29 October you can ask it. The Register is hosting a dinner in Washington with HPE and NVIDIA, our editor-in-chief Matt Rosoff in the chair, and nothing said on the record. HPE sends Thierry Pienaar, HPE Fellow and Chief Technology Officer for HPC and AI Worldwide Sales. NVIDIA sends Bert Condensa, Vice President for Enterprise AI Factory Segment Sales. They are there to take questions, and neither brings a deck. Both spend their working lives at the point where AI infrastructure gets bought, which is where sovereignty requirements are either met or quietly worked around, and the Chatham House Rule means they can say which. That rule is what makes the evening worth the trip. What gets said at a table like this does not get said on a webinar: what a bank had to rebuild, for instance, or which sovereign cloud option somebody paid for and would not buy again. Everyone is there as a peer, the vendors included, and the chair steps in the moment anyone starts to sell. The questions on the table are the practical ones: which options actually hold a boundary, what each costs in capital, capability or elasticity, and whether any of it still holds once the pilot is running for real. It is one table, and places are limited. The seats are for technology, data and risk leaders in healthcare, financial services, research, pharmaceuticals, government and higher education who are making these decisions now. Dinner runs from 6.00pm to 8.30pm at Fiola Mare. Places go by invitation. Request one here. Government and public sector guests: where your agency’s ethics rules require it, you can pay your own share and we will provide a receipt. Please say so when you request an invitation.

Elon Musk says SpaceXAI, the company formed by merging SpaceX and xAI, will be renamed SpaceXSI after President Donald Trump ordered US federal agencies to replace "Artificial Intelligence" with "Super Intelligence" in all comms. Trump's September 29 executive order directs federal agencies to switch from AI to SI in official communications and other non-statutory documents. Musk endorsed the change, posting "No more AI. SI. It's better," and confirmed plans to rename SpaceXAI: "Yes, we will make that change." He gave no date for the rebrand. At the time of writing, the spacexsi.com domain was listed for sale. SpaceXSI sounds somewhat like a performance car of the 1990s, or perhaps a family wagon with sporting aspirations. This writer remembers putting his driving license at risk in a Peugeot 106 XSi, occasionally finding that ambition exceeded ability. We asked OpenAI whether it plans to rebrand as OpenSI. The company has yet to respond. Trump's order does not require private companies to change their names. Tech leaders including Mark Zuckerberg already use "superintelligence" to describe their ambitions. Trump's order goes further by applying the label to technologies covered by the existing federal definition of AI. Musk has form when it comes to rebranding. Twitter became X in July 2023, nine months after he bought it. "SpaceX is a super intelligence company," Musk posted. Bad news for NASA, which probably thought it had something to do with rockets. ®

Games Workshop is on the hunt for a head of IT to guide the organization through its technological saga for the next two years. The UK-based tabletop games retailer and manufacturer of the globally popular Warhammer figures said that under the fixed-term contract, the heroic individual would take overall responsibility for its global IT and customer technology capability, and report to the Chief Operating Officer. “The role has a broad remit spanning our core technology estate, information security, major systems change and the technology behind our customer-facing digital platforms, including warhammer.com,” the job ad said, although sadly it closed for applications over the weekend. “We need someone who can understand what Games Workshop needs now, anticipate what we will need in the future, and make good decisions about how technology supports the long-term success of the business,” it said. Candidates might want to be aware that, if they are lucky enough to be offered the job and they accept the mission, the battle will not be an easy one. The head of IT will take responsibility for core IT environment and information security, advising senior management on technology, security, risk and investment. They will also support the teams to keep “significant technology programs moving, openly tackling the challenges that come with major change, and ensure that we continue to make rapid progress.” The Nottingham-based retailer and manufacturer is in the midst of ERP and supply chain management systems overhauls that were first mobilized years ago and have been beset with delays and over-runs. Back in January 2021, The Register reported that Games Workshop delayed its Microsoft Dynamics-based ERP rollout and appointed a new integrator, having already hoped to go live in 2020. By September that year, it said project issues were impacting the development of new features in online shopping. In January 2022, it added another £500k to the project budget with no end in sight. A year later, it said the business faced ongoing challenges to integrate new IT systems when it was still heavily reliant on working with its legacy tech systems. “Finishing these projects and upgrading our systems continue to be a key area of focus. We are making some progress but it's costing us more time and money – we will remain cost conscious," it told investors. Around four years later, challenges remain ongoing, according to the latest annual report, published in July, for the year ending in May 2026. It said the group’s main financial system is still being supported by a number of legacy ERP applications for which “we have identified control deficiencies in our previous audits.” These still exist, according to a review of IT systems. The report added that the responsibility for delivering IT had moved to its operational manufacturing and supply chain director, under the COO from June, a decision which “helped bring all of the internal departments together.” “Our Systems Improvement Plan (SIP) is on track with very clear milestones for us all to support and deliver,” it added. Games Workshop is famous for the phenomenon of Warhammer 40,000 — also known as WH40K or 40K — thought to be the most popular miniature wargame in the world. Battles are set in the distant future, leaving the technology department to hope the end of its long-running ERP saga comes before that date. ®

Anthropic's IPO prospectus leaked this week, and boy was it filled with interesting details. You can listen to the latest episode of The Kettle on this page, Spotify, Apple Podcasts, or YouTube. You can also subscribe on those platforms to be notified when a new episode goes live For starters there's the fact a third of it is all about existential AI risks, then there's the fact that, despite some impressive revenue growth the company still has yet to turn a profit. Combine that with massive capex predictions, and you have an IPO with massive valuation, yet a long, hard road toward anything resembling earnings. Join host Brandon Vigliarolo, El Reg editor in chief Matt Rosoff, and enterprise IT reporter Lindsay Clark on this week's episode of The Kettle where they discuss the revelations, what they mean for the state of the AI industry, and just how much distraction from its actual problems Anthropic can eke out by keeping that doomsday narrative front and center. A lightly edited transcript is below: Brandon (00:04) Hi everyone, and welcome to another episode of The Register's Kettle podcast. I'm your host, Brandon Vigliarolo, and this week we're gathered around the teapot to discuss hey, what else, AI? Specifically, the reported leak of Anthropic's IPO prospectus, which was apparently filled with all the doom, gloom, and to the moon promises for investors you'd expect from the world's most disruptive industry. With me to talk about the revelations are our editor-in-chief Matt Rosoff and enterprise IT reporter Lindsay Clark who brought us news of this whole mess. Thank you both for joining me. Lindsay Clark (00:36) Thank you. Matt Rosoff (00:37) Thanks for having us. Brandon (00:38) Yeah. So Lindsay, I know we didn't break this story. This was, I believe, a Financial Times exclusive, but what was reported about those IPO docs, in the story you wrote and just how apocalyptically optimistic is this whole thing? Lindsay Clark (00:53) Well we have this scenario where a company, Anthropic, is trying to sell itself by telling people that there's an existential risk to its technology, which I think is a bit of a weird way to go about it. But nonetheless, that's in the prospectus according to the Financial Times. And they also offer some insight into the numbers that are going around. Anthropic is still massively loss-making as far as we know. Brandon (01:22) There was what, I think you said, $8 billion loss on $4.6 billion of [revenue] or something? Lindsay Clark (01:28) That's right, and it says it plans to spend $500 billion or so on cloud computing over the next few years. So I don't know what you make of those numbers, but it's hard to get them to add up in my head. Brandon (01:40) I mean it's interesting just how it's still not a profitable company. But Matt and I were discussing earlier, I was saying that I think it's interesting that Anthropic is pushed ahead with this IPO despite OpenAI pausing it, because I feel like Anthropic has long had this reputation as sort of being the more safety-conscious, the more conservative AI company. But here they are with OpenAI kind of saying, Well, we're gonna pump the brakes on our IPO until we can make sure our models are safe, but Anthropic is plowing ahead. And Matt, you had an interesting take on that and it was more to the financial end, right? So what were you saying to me about that earlier? Matt Rosoff (02:21) I think that if you probably were to look at OpenAI's numbers, they would be even worse than Anthropic's. And I can't overstate based on the leaked prospectus information that we've seen. It seems as if the more money they earn, the more they have to spend. So the margins seem pretty bad. Now there was a report that they had actually turned profitable on an adjusted basis [in] Q2 this year, but that's if you strip out all of the training costs. So I guess if it's just inference alone, without the training costs and without certain other costs, they have decent margins. So maybe there is a path to profitability. Now it looks as if their operating loss last year was almost double their revenue. So if their revenue continues to increase at the rate that it is, which is, it seems like it's going up 4x, 12x per year, something like that, depending on the period you're looking at. If they can't get the expenses under control, the business doesn't seem to make a whole lot of sense. Nonetheless, I think they are growing faster because they were earlier to enterprise and they have become more essential to more companies than OpenAI did. And I think OpenAI is still sort of adjusting its business model and maybe not growing at quite the same rate. So that's my suspicion is that those numbers would look even worse. Brandon (04:01) Yeah, so the safety argument is probably just some spackle slapped on top of the bad financials to try to kinda hide it, I guess. If what you're saying is correct. Matt Rosoff (04:12) That's my speculation. It doesn't seem to make a whole lot of sense to say, well, we're so worried about this amazing thing that we've built that we don't want to show you our books and we're not gonna go public yet until we have that all figured out. That seems like a strange argument to make. Brandon (04:31) Yeah, I would have to agree with that. So speaking of economic impact, Lindsay, you were talking a bit more in your story about some of the broader sort of economic scope of the AI industry. And I'm curious, could you go more into some of the numbers you discussed in your story? Lindsay Clark (04:51) Yeah, sure. So it was a few months ago, I think in August, ING. Well, I'll go back a bit. So in 2025, there was an economist, Jason Furman, who estimated that 92% of the US GDP growth in the second half of 2025 was down to AI investment. Now that figure has come down, ING said in August it was about a third and I've just seen a more recent figure saying it's about a fifth. So, nonetheless, a lot of US growth is coming from AI spending and I mean the forecast for spending, there's one estimate from Bain, capex on AI datacenters will be 6 trillion by 2031. And then PWC said capex again would be 31 trillion or so by 2050. So God knows what will be happening by 2050, but a lot of money spent on datacenters according to PWC. Matt Rosoff (06:03) What's not clear to me is where is all this money going to come from? Where do you find $31 trillion to subsidize what is still essentially a loss-making industry? Brandon (06:15) Right. Lindsay Clark (06:16) This is the trillion-dollar question, if you excuse me. I spoke to Gartner earlier this year and Gartner is still very optimistic about AI and tech spending. Most of the tech spending is coming from AI investment by tech companies, so about two thirds of the total tech spending in the world is coming from tech companies themselves investing in AI. Surely that must be passed on in some way to tech companies' customers. And that's true, a lot of these models will be embedded in the applications and the products businesses use. What concerns me and what is an open question at the moment, I think, is we're seeing cheaper models flood into the market. Open source, open weight models and kind of commoditized models from China. And everyone I speak to in development, datacenter companies, which is my beat, say that internally, they're completely trying to cut costs by using cheaper models wherever they can and not going to the most expensive model first, as it were. And yet these AI companies we've just mentioned, they're ploughing billions and billions of dollars into building bigger and better models. I don't know how OpenAI and Anthropic, how their forecasts depend on still being able to charge top dollar to access their models if the market is going to commoditize very quickly. Brandon (08:01) Right. I mean, it's most business use cases don't even require these models, right? You can make do with a smaller local model. I mean, it's interesting, right? We saw with the whole Jev boom recently. A lot of business AI uses can be done by these decision models, right? You don't need a full LLM doing all this stuff. And, Jev was already what 0.042 cents per million tokens going in and free going out? And then Cloudflare just comes out with an open weight version that can be installed on local hardware. If you run it on their hardware, it's a bit more expensive thinking by a factor of six, which is still cheap, right? But it's available on Hugging Face, and they're like, if you got the hardware to run it, go ahead and download it and run it yourself. Right. It's like there's a lot of ways to trim AI costs. And I think the pipeline for these big companies is....I don't know if it's super clear, right? They're banking on a lot of what-ifs to hit profitability. Matt Rosoff (09:03) Right. I think probably the idea Anthropic had is that they will subsidize the costs initially, get enterprises hooked on the product, and then gradually raise prices when companies don't really have an alternative. But they do have alternatives and I think that's probably something they hadn't anticipated working out quite the way that it did. Brandon (09:24) That's probably why Deepseek was such a bomb in the industry when it first came out, earlier this year, right? Everyone was like, Well, the age of Anthropic and OpenAI is over if these are what these open weight models can do. And it doesn't seem like it's gotten a lot better, right? And now we're staring down the barrel of this massive percentage of the American economy being dedicated to this. I mean, Lindsay, you mentioned in your story that it was comparable in a lot of ways to the dot com boom, but maybe hyper inflated. And so do you have any idea relative to the dot com boom what kind of economic impact this could have on the global economy? Like what share of the economy is tied up in AI investment right now versus dot com? Lindsay Clark (10:06) I don't have a direct comparison. It's worth remembering that although the dot com boom and subsequent crash, billions and billions of dollars were wiped off the stock market value, the actual real economy recovered quite quickly from that. Contained within the tech industry, there was a huge write off of telecoms investments. So a lot of telecoms companies put fiber in the ground. They were very leveraged to do that, financed by debt. And then the demand was not there as soon as they thought, and they were unable to service the debt. And a lot of these companies went bankrupt. They invested about $500 billion in that period. So that's the story there. And there was a real effect of the dot com crash. In some ways there was a technical recession and we weren't going great guns in the way we had done in the late '90s. But the effect on the real economy was not all that dramatic and the crash was relatively contained within the tech sector. And you have to remember that investment, that telecoms investment is being used eventually. It was used eventually. It was worth putting in the ground. It's just a lot of companies went bust in the meantime. Brandon (11:34) I guess which begs the question whether or not all these datacenters will end up being used if the AI bubble bursts, right? I got a feeling that maybe not. Maybe there'll be some massive redistribution of hardware. Lindsay Clark (11:44) I think in the long term, yes, but it just depends if the same thing happens In the meantime, a bunch of companies go bust. I mean, the thing that would concern me is how much of these circular investment strategies are going on where you see OpenAI is promising money to Oracle, but $300 billion to Oracle to build datacenters, but OpenAI doesn't have the money yet and this whole thing becomes quite circular. There's a number of these deals which are dependent on stock market price and so on. Brandon (12:18) Yeah, it's a real potential mess. Speaking of Oracle, right? I think you mentioned in your story that they're having some issues getting financing together for this New Mexico data center. And I think after you published that, it's come out in the past couple of days that they're also having similar problems at a Wisconsin facility they're trying to build. I know a lot of people were saying that Oracle might be the canary in the coal mine for all of this. Lindsay Clark (12:43) Kind of yes and no. It's worrying. Oracle seems to be borrowing a lot of money. And I don't know if they're borrowing more than the other companies. I know Meta's borrowing as well. It's all to do with getting regulatory clearance for the power that goes into these data centers. That's why the force majeure was issued with the New Mexico data center to do with the developer. Brandon (13:09) Okay, so it has more to do with utility issues... Lindsay Clark (13:12) Yeah, but it shows the practicality of actually going from, hey, let's build a load of data centers to actually turning them on. I mean, can take longer than certainly longer than Oracle thinks. Matt Rosoff (13:23) Right, it's very easy to write in a prospectus or to say, "We're gonna build a bunch of data centers." It's much harder to do it. It's the sighting and the power and the other resources that need to be used and getting enough GPUs to put them in and letting the advanced networking technology to reduce latency. It's an incredibly complicated thing to do that is just sort of glossed over. We're gonna build some data centers. I mean these are orders of magnitude bigger than the ones that we were building twenty years ago to sort of power the first big generation of the internet. It's non trivial. Brandon (14:01) Yeah, no, absolutely. Lindsay Clark (14:01) Yes, I would agree with that. Although Oracle to some extent is writing this in Red Ink, its capex forecasts exceed its revenue for next year. Oracle has very deep pockets. It's got a huge number of enterprise customers that can't walk away from that software anytime soon. And so it has this cash cow of enterprise licenses that it will carry on milking for years to come. So it has very secure revenue streams. I'd hate to be the one to predict anything too bad happening to Oracle. Matt Rosoff (14:36) If anything, they can probably withdraw from some of these commitments and their core business will be fine. And that's the other thing. Google is doing a lot of build-out, Meta is doing a lot of build-out. These companies have an incredibly strong bottom line with huge, huge, huge margins. So to some degree, this is an affordable expense for some of them. It's just at some point the borrowing gets ahead of the revenue to a point where I wonder if some lenders will start to call debts in and that's when things can reverse in an interesting way. Brandon (15:12) Yeah, and for Meta and Google, right, and even Oracle, as you said, Lindsay, that might not be as big a problem, right? Because they have other things to fall back on. For OpenAI and Anthropic, on the other hand, where their entire business is based around AI buildouts, they don't have an alternative product to pull money from, redirect some funds. It's all kind of dependent on this. It's also interesting. Lindsay, when I was reading your story, I think it came out before that whole meeting that the Trump administration had with AI leaders this week. And we write a lot, and as you alluded to in your story, we write a lot about the fear-mongering around AI, right? About how this is all maybe just one big marketing ploy. It's a regulatory capture push. And I found it interesting that one of the sort of big headlines that made it out of there, aside from this voluntary AI pledge that is probably going to mean absolutely nothing, was the fact that a lot of these leaders in these meetings were really kind of laying into Dario Amodei for sounding the alarm on all this, right? Like they seemed upset that he was so aggressive on the apocalyptic warnings. What is that? I mean, what do you guys think that means for the AI industry? Are they all less scared than they let on? You know, is Amodei really pushing that hard to the point where it's become a joke kind of amongst people following this? What does it mean for the industry? Matt Rosoff (16:38) I sort of vacillate back and forth between "do they really believe this stuff" and "is this just a cynical marketing ploy"? Because what a great way to say, hey, this company is worth more out of the gate than any other IPO in history. It's a two trillion dollar value, but it's absolutely world-changing. It's akin to the invention or the discovery of fire or electricity or something like that. If you believe that, then that valuation is justified. And one way to get people to believe that is to say, this thing is so powerful, we don't even know what it can do. It's out of our control. It takes action on its own. There are internal documents and there's stories that leak about how people within Anthropic actually seem to refer to Claude as an independent living entity. There was a story in the New York Times recently about them consulting with religious leaders, asking for some guidance on how to make Claude act ethically, and in some of these conversations they were hinting that they thought that perhaps Claude might be an entity that is alive. I mean, this stuff is really fringy. So I had been in the camp that I thought it was kind of a cynical marketing ploy, but I'm not as sure. There was an article in the Wall Street Journal, I think this week, about sort of the history of where Amodei comes from and he comes out of the effective altruism movement, which is, in my opinion – Brandon (18:23) Oh boy. Matt Rosoff (18:24) – a sort of fringe philosophy based on doing the most good for the most number of people in some sort of indeterminate long term future. And as I read about it, a lot of times it comes across as a convenient excuse for sociopathy, right? You can say, well, I can take this horrible action now, but it's going to be justified by things that happen in the future. At any rate, apparently he and some other folks, adherents to this philosophy, were all living in a house in the Glen Park neighborhood of San Francisco, which is actually kind of near where I live, and talking about these kinds of things. And Sam Bankman-Fried, who later went on to found a crypto company and then went to prison for fraud was part of that same philosophy. Brandon (19:14) Yeah. I feel like he's been he's been kind of the poster child for effective altruism as sort of like a hey, this might not be too legitimate. Matt Rosoff (19:22) Yeah, Anthropic is kind of distancing itself from this philosophy. But if you read this article, they were all kind of traveling in these circles and having these sort of deep philosophical debates. And apocalypse came up frequently before AI had been invented. So if you're looking for the end of the world everywhere already, and then you discover this really interesting new way of making computer systems operate in somewhat unpredictable fashion and allowing them to self-improve, maybe you're gonna project your end-of-the-world fantasies onto this particular technology. So maybe they really do believe it. I'm actually not sure. I just think it's crazy, it's nonsense, and I think I find myself in rare agreement with President Trump who sort of seems to be saying this is silly, stop talking like this, we have to continue developing this full speed ahead. Brandon (20:24) Yeah. Matt Rosoff (20:25) I don't know if I believe that either. It does seem like we need to look at the very real safety effects and really very real things that are happening today and why some of these models are not in control and how guardrails can be placed on them and so forth. But to sort of have this, "It's gonna end the world and there's nothing we can do about it" is a really strange way of going about things. Brandon (20:47) Yeah, it is, especially when these companies are all the ones that really control the off switch, right? Matt Rosoff (20:54) Yeah. Brandon (20:54) I don't think we've had any AI models yet escape from their confines and replicate themselves somewhere on some mystery server that stays on no matter what happens. It's all perfectly within our control, right? The most critical systems in the government are air gapped. They probably can't be accessed by these things, right? Like I think there's some very real risk here too, right? Like we've seen some of the bad things these AI can do when you prompt them. Again, when you remove their guardrails and you give them problems that they cannot solve based on their own constraints, they're going to solve those problems and figure out a way to do it outside their constraints, right? They can act autonomously in bad ways to a degree. But I think going from that to, well, you might as well give us all the money you can so we can stop this thing from ending the world. It's just I don't see how A turns into B here, really. Matt Rosoff (21:49) Yeah. Lindsay Clark (21:49) And it's kind of a smoke screen for the nitty gritty, practical legislation that would be required to ensure companies take responsibilities for the technologies they've produced. And I've written about liability in using AI agents in business. And it's a completely unknown. The lawyers I spoke to said, we're just going to have to rely on test cases as they come through, certainly in Europe. The European Union's trying to introduce, well has introduced, legislation on AI. In the US not so much. So I think the whole doom talk avoids that conversation, or at least puts it off, which is what they all want to do until the IPOs are out of way and the money's in their pocket. Matt Rosoff (22:35) Right. What if I'm using an AI and it does something unintended and causes harm, whose fault is that? Am I liable for the prompt I gave it or is the company that created it liable? Yeah. These are all really good, interesting real world today questions... we don't have to think about the apocalypse. Brandon (22:53) Yeah, I mean we need to be asking, or like I think we've written on more than one occasion, I think maybe Tom has written this in his stories about OpenAI's AI agents going rogue and stuff. It's like if a human did any of these things, we would hold them legally accountable. Like when are we going to start holding these companies legally accountable for basically putting the entire internet downrange of their tests and saying, Hey, let's see what we can get this thing to do. Lindsay Clark (23:18) Yeah, and it's much more prosaic stuff AI is being used to summarize your HR history. AI is being used to make financial decisions and financial recommendations. And so, who's accountable for that? Is it the people who provide the model or is it how you train the model? And it could be both, but how do you decide whose real responsibility it is? And that's a complete unknown. Matt Rosoff (23:41) There was an article Tom Claburn wrote this week. He went to an AI conference here in San Francisco. And he was talking about how the industry is really hot on the idea of agentic commerce, right? So allowing these agents to make spending decisions for you, but the technology and the reality just isn't there because you can't have an unpredictable, indeterminate program deciding how to spend real money. It doesn't work. And what's more, n a practical matter, they're not capable of doing certain things, passing certain checks. There's a concept of card present, right? You have to present the card to say, yes, I give authorization for this transaction to happen. And an agent literally cannot do that. So it's again these really complicated here and now real world questions that need to be resolved about who controls and who's responsible and who's liable. And I think when you have an IPO prospectus that has paragraphs and paragraphs about end of the world far-fetched scenarios, it's a bit of a distraction from the real stuff we have to figure out first. Brandon (24:52) Yeah, absolutely. Absolutely. I mean it'll be interesting to see kind of what happens with this IPO. Was there any date, any time frame in the IPO leak, Lindsay? Was there information on that when this might come out? Lindsay Clark (25:05) November. Brandon (25:06) Okay, so we're soon, right? So we're looking at possibly before the end of the year. Lindsay Clark (25:09) Soon. Yeah, they hope, yeah. Matt Rosoff (25:12) Yeah, I've seen I've seen Fall. So before before the end of the year, yeah. Brandon (25:17) What happens next year, once they're public, if they can be held accountable, if anyone's gonna bother holding them accountable, I guess is the other question, right? Full speed ahead on AI seems to me like, well, let's just hope nothing serious happens. Matt Rosoff (25:31) And and on a practical matter, again, these are leaked documents, Reuters and the FT are the two that I've seen, but they didn't have detailed numbers for 2026. So we don't yet know how fast this growth is continuing or what the expense profile looks like this year. Again there was a report that Anthropic put out there saying that they were profitable on an adjusted basis in Q2, but we don't know what that means. When the actual IPO document drops, and everybody can look at it, and it's got quarter-to-quarter numbers going all the way up to the most recent quarter, and we can see exactly what the spend looks like, that'll be interesting because then you'll have investors taking really close looks at things and deciding whether this makes sense and whether there is a path to massive profitability and to making this a $2 trillion company. We saw the SpaceX IPO earlier this year and that core business of launching rockets and satellites particularly looks okay. the AI component of the business looked less good, but it was good enough that it went public and the stock is a little bit above where it debuted. So people aren't too skeptical about that. So we'll just have to see how this IPO even flies. Brandon (26:57) Yep. I'm sure it's gonna be an interesting time once it actually becomes public and we have a chance to dig into it. And we will definitely dig into it on The Register and more than likely dedicate an episode of the kettle to it. So we really hope that you will tune in to hear more about what we have to say about that when it actually becomes available. Thanks for listening.®

Ubuntu 26.10, “Stonking Stingray,” has entered beta testing, slightly later than planned – it is in fact scheduled to be released on October 15. While many people will stay with the LTS version, there’s some interesting stuff in there. There are already some release notes, but like the OS itself, they aren’t final yet, and there are still some gaps. Canonical does note that the "oxidization" process – exchanging core Unix tools written in C with new Rust replacements – goes further in this edition: now, the stock OpenPGP encryption support is via the Rust-based Sequoia-PGP. Perhaps one of the most daringly different new features is Canonical’s promised LLM-powered Myna speech-to-text tool. This isn’t installed by default, but you can add it, and a choice of speech-recognition models. We found various sets of instructions, including the unofficial “Ubuntu Handbook”. Our geriatric testbed was a little old for this, though, and we couldn't try it. A couple of weeks ago we took at look at the beta of what will become Fedora 45. Both Ubuntu 26.10 and the default workstation edition of Fedora will use GNOME 51, but since the Fedora beta, the final release of the latest GNOME is here: GNOME 51 “A Coruña”. The version in the Ubuntu beta isn’t working completely smoothly. We saw a persistent error about missing themes, and although the OS offered to install whatever was missing automatically, this failed. Since it doesn’t actually say what is missing, we can’t install it ourselves. The problem is only cosmetic: just some missing icons in applications, very similar to what we saw in Alpine Linux 3.16 four years ago. Sadly, problems around GNOME themes are a long standing issue. As we reported earlier this year, the Linux Mint developers are still struggling with them. The occasional missing icon aside, though, Stonking works pretty well. We tried both in a Virtualbox VM and on bare metal on our old testbed Thinkpad W520, and it ran well in both. The new installation program that arrived with the 2024 LTS version works fine, and now makes it much clearer which control is currently selected with an orange highlight. We do wish the window was resizable, though. The testbed machine has two SATA drives and we had to do a lot of scrolling to choose the right partitions. It also failed to use our dedicated swap partition, although we selected it during installation. Although GNOME 51 removed their support for old Nvidia devices in this development cycle, it still picked up and used the machine’s Nvidia-controlled Displayport output. It didn’t work perfectly: GNOME couldn’t report on the model or properties of our Quadro 1000 GPU, or the CPU-integrated Intel one for that matter, but at least the second screen worked. More surprisingly, it even worked in the Fallback session, even though that uses the Metacity X11 window manager – and while logged into to a Fallback session, it could identify both GPUs. GNOME 51 isn’t a very feature-heavy release, but the Settings app has some useful additions, such as the ability to automatically disable the trackpad when an external mouse is connected. In our testing, it picked up and let us configure mouse, trackpad, and trackpoint, each in a separate tab. GNOME 51 no longer supports Wi-Fi networks with the old WEP encryption system, though. Modern Linux continues to move away from being an OS for old and low-end hardware. If the machine is online during installation, the installer fetches any package updates as it works, so the first update post-install is much quicker. However, this doesn’t apply to Snap packages: the .deb package phase of our first update was very fast, but we got various new snaps. One detail we haven’t noticed before is that while a snap application is updating, its dock icon temporarily gets a progress bar to tell you it’s being updated. We did a full install with all the optional extras, so this applied to Thunderbird as well as Firefox. A full installation took a fairly hefty 13 GB of disk space, although to be fair, that included 4GB of swap space – so in real money, it’s about 9 GB, which isn’t too bad for 2026. According to the htop command, it idles at about 1.2 B RAM, which is quite high, but then GNOME is no lightweight. Linux kernel 7.3 is still in development, but at the time of writing it’s up to RC5, and our testing ran without a hitch. The final version may be out before “Stonking” is completed, but if not, users will get it very soon afterwards. We also gave the Kubuntu desktop a very brief try. In this release, it uses Plasma 6.7.5, and ran well – it was fast, smooth, and happily drove both screens of our testbed machine; helpfully, on our first login, it asked where the second display was relative to the first. If you want modern Wayland niceties such as different dots-per-inch settings on different screens, but you also want a traditional Windows-style layout, KDE is currently the way to go. Plasma 6.8 is scheduled for later this month, and totally drops X11 support. It’s likely to be the desktop in Kubuntu 27.04, making it Wayland-only – so 26.10 is probably the last ever version of Kubuntu in which displaying your desktop thanks to X.org is a supported option. We installed X11 support, which is just one command: sudo apt install -y plasma-session-x11 This brings in multiple extra dependencies, which is why we specified the -y switch, but it took just seconds. We logged out, logged back again choosing “Plasma (X11)”, and… Everything looked absolutely identical. Which is how it should be, of course. Both screens were working normally, thanks to the Nouveau FOSS NVIDIA driver. The only visible differences are the classic “wobbly window” effect when moving windows around, and that the title-bar context menu regains the option to “shade” a window: roll it up, leaving only the title bar visible. You can’t do this in a Wayland environment. Under Wayland, htop reports about 1.2 GB in use; under X.org, about 1.3 GB. The resurrected Oxygen theme from KDE 4 is built in, if you want a bit more visual bling, but we preferred the flatter grayer look of the default Breeze theme. The only freeze that we saw in all our testing of the Stonking beta was after logging out of a Plasma (Wayland) session and trying to log back in using Plasma (X11). To be fair, we suspect most people wouldn’t chop-and-change this much. Ubuntu Stonking is shaping up well. Aside from the small theme-related glitches, GNOME worked fine, and KDE didn’t even have those minor issues. The equivalent beta versions of all the official flavors are also available, including both Ubuntu Unity 26.10 and Ubuntu MATE. Despite a hundred other distros, Ubuntu remains the go-to desktop Linux for most people. While Canonical is getting slightly more experimental with interim releases again, this is a mature product now and they’re a pretty safe bet. ®