Peter Steinberger, member of technical staff at OpenAI and the creator of OpenClaw, estimates that he spent around $10 million in tokens running AI workloads last month. But don't feel bad for him, because it's all coming out of investors' pockets. He made this admission at the Cloudflare Connect 2026 conference in San Francisco on Wednesday, in a conversation with Cloudflare CTO Dane Knecht. Knecht previously observed that Steinberger has acknowledged having "a virtually unlimited token budget, which not all of us in the room have." Responding to the audible surprise rising from the audience, Steinberger added, "What people don't always understand is not all of that is useful." That only amplified the murmuring, perhaps not the effect he had been after. Meanwhile, his employer, OpenAI, is reportedly trying to "accelerate funding" ahead of rival Anthropic's expected IPO. Steinberger's subsequent musing about his preferred way of working failed to recognize Knecht's acknowledgement of financial limitations. "I feel like the main challenge a lot of people still have is that our infrastructure structure hasn't really caught up yet," Steinberger said. "Like why is this still using my computer? Shouldn't this all be in the cloud?" Apart from the privacy and compliance implications of relying on cloud-based AI, cost remains a consideration for anyone not setting venture funds alight. Cloud-hosted agents can be expensive. Shouldn't this all run locally? Pointing to the OpenClaw project and how a lot of open source maintainers have given up accepting pull requests due to frustration with AI slop, Knecht asked Steinberger whether we're approaching the death of open source or a golden age of AI. Steinberger said he thinks the way pull requests are handled needs to change because people don't explain how they arrived at their proposed changed code. "If I see your thoughts, how you got from A to B, my trust in the output is way higher than if you just throw the finished work across the fence," he said. Knecht also asked Steinberger whether everyone in the room would have hundreds of AI agents running in a few years. Steinberger doubts it will be hundreds. "I think everyone will have some sort of intelligence that runs and thinks and helps them throughout the day or even predicts what could help you in the future," he said. "So instead of asking the question, maybe you already get the answer." Toward the end of the session, Knecht raised the issue of trust and how to know what data to give to an agent. That's a hard question, Steinberger admitted. Part of it, he said, has to do with how much you trust the underlying model itself, because so much of model behavior is not governed by the harness – the code that wraps the model. "A lot of decisions that I built into OpenClaw early on now can change because the models are so much better," he said. Even if risks like prompt injection have been reduced through engineering efforts, Steinberger said, "you might not want to give your agent all the permissions." Instead, he suggested using an agent to authorize your agent. "You want to have another agent that can give it permissions or not and then this agent has to negotiate with another agent," he said. "Many times the agents are like, 'oh yeah, actually maybe I shouldn't do this,' or 'there's a simpler way that keeps it within the guardrails.' So those those two systems combined greatly increase my trust that I just can let the agent rip." ®

Why use your smartphone when AI can do it for you? We’ve reached a new level of AI inanity with the launch of Opposable, a tool that taps, swipes, and types on Android phones and controls iPhones through a companion computer, sparing users the apparent burden of operating their own devices. If you don't want to update a calendar event, start a timer, or enable dark mode, you can apparently use an AI agent app on your computer (any client that supports MCP will do) to have Opposable do the thing for you on a connected phone. You could just do it on your computer, of course, but that’d require actually using your hands - like it’s some sort of baby’s toy. Lame. Opposable describes the purpose of its app as being to “give your AI thumbs” that allow an agent to open apps, read the screen, tap, and type, all while signed in to a device as the person too lazy to use a super-convenient technology themselves. As described in a blog post published by its creators Wednesday, Opposable's developers explain that its Android implementation uses accessibility services to inspect the screen, while screenshots provide visual input for an AI model. The model then decodes the images and figures out where it needs to tap on the screen in order to do whatever its user asked of it. The visual processing is handled by a vision-language model known as GUI-Owl 1.5 2B, running through llama.cpp, while tasks can be handled by either a paired desktop AI agent or an on-device agent that works without a desktop connection (the latter is available only on Android devices – iPhones require a companion computer, with a Mac beta available and a hardware-assisted Windows/Linux option also advertised). Even then, AI isn’t used for everything, with around 40 common tasks (setting an alarm, sending a text, switching Wi-Fi on or off) running as fixed code without an AI model involved. Opposable doesn’t appear to be particularly fast, however, and its accuracy is questionable too. Opposable’s website includes three different examples of tasks executed using its product, and none seem to merit using an AI over, you know, human hands. In one example, Claude Code over MCP is used to change the system’s theme to dark mode and then set a five-minute timer. This takes 50 seconds and requires 17 tool calls. In a second example using Claude Code and MCP, Opposable is asked to look up when Apollo 11 launched and who was on board, which takes just shy of 2 minutes. The number of tool calls isn’t mentioned, but the run cost is $0.44, which seems ridiculous when you can use your hands and figure it out for free. Hell, just asking Gemini or Siri right on a smartphone would return the answer faster and cheaper. Those half-a-buck requests to Opposable could add up pretty quickly if you’re lazy enough to justify its use. A third example shows a virtual Android device operating without a paired desktop agent, instead relying on the on-device agent that Opposable installs on Android phones. Opening the clock app and switching to the stopwatch tab takes it about 96 seconds, which is supposed to be impressive, apparently, as the company behind Opposable decided to showcase it on the thing’s website. As for Opposable’s ability to actually accomplish tasks, well, that’s not exactly a sure thing. “On eight tasks the model had never been tuned on, run on the emulator twice each, giving it the whole task succeeded 4 and 3 times,” the blog post notes. I’m pretty sure my brain and hands can do better without burning tokens to fail. “Giving it one subgoal at a time and checking each claimed success against the accessibility tree before the next subgoal succeeded 6 times in both rounds,” the blog continues - so even when it gets babysat each step of the way, it still only succeeds at three-quarters of tasks. Cool - that’s really going to free up all that time wasted setting timers and doing other mundane tasks on my handset. Why someone would install this on their Android device - which requires sideloading an APK and manually granting access to read and interact with the screen through accessibility settings - seems like a bit of a mystery. Opposable’s website even warns that it could be a security disaster: “whoever holds the token can operate the phone, so treat it like a password,” but still invites users to try. An argument could be made for using Opposable as an accessibility tool for people with difficulty using a smartphone, but this isn’t mentioned anywhere on its website. Opposable is instead billed as a way to “hand off the phone chores,” leaving users free to do something else - just not on the phone currently being slowly controlled by a clumsy AI. Sure, you could give your AI some virtual thumbs through Opposable, but … why? We reached out to the Opposable team to get an answer to that question, but didn’t hear back. ®

Nvidia hopes to capitalize on European sovereign AI demand by unloading a slew of new open weights models spanning chatbots, healthcare, and robotics. The major model devs have the resources to build highly capable models from scratch. But most enterprises don't, and have instead been forced to fine tune open weights models from the likes of Mistral or H Company using their own datasets. At GTC Berlin, Nvidia unveiled its latest additions to its open weights model library with refreshed Nemotron Retriever embedding and Alpamayo autonomous vehicle models. The point of embedding models While by no means the sexiest subclass of models out there, embedding models are some of the more useful ones for enterprise apps. Retrieval augmented generation (RAG), a technology we've explored extensively in the past, relies heavily on the embeddings generated by these models for applications like AI-augmented search. At a high level, RAG works by cutting up documentation in chunks, then converting those into vectors called embeddings. The embeddings are then ranked based on their relevance to the user's query before they're fed into an LLM's context. Nvidia's latest embedding and reranking models, based on Nemotron 3.5, support embeddings of both images and text across 34 languages. This time around, the models are being offered in sizes ranging from one to eight billion parameters, with NVFP4-quantized versions for those running on new Blackwell-based accelerators. That's small enough to run on something like a DGX Spark with plenty of room to spare for other models. Always take vendor claims with the usual grain of salt, but among competing embedding models, Nvidia boasts that at BF16-precision, its new 8 billion parameter embedding and reranking models are now the top-ranked dense models on the ViDoRe V3.1 benchmarks. However, not all documents are easily converted into embeddings. PDFs are notoriously difficult for embedding models to make sense of. They're often scanned from printed source material and formatted for consumption, which means scraping text and parsing charts. For this, Nvidia has also launched a refreshed Nemotron Parse model that's been trained specifically to extract information from PDF documents. Nvidia claims the model can process roughly 10.5 pages a second when running the one billion-parameter model on a single H100. Alongside the new embedding models, Nvidia is also rolling out access to Alpamayo 2 Edge, a more compact version of its autonomous driving model based on which promises real-time processing of streaming video and sensor data when running on the company's DRIVE AGX Thor platform. The models come just a week after the GPU giant launched its Nemotron 3.5 Super vision language (VL) model and teased Nemotron 3.5 Bolt, a 74 billion-parameter mixture of experts (MoE) model designed to fit in as little as 24 GB of video memory. That model, Nvidia says, will be available for download starting Friday. While open models provide enterprises and governments a path toward sovereign AI, Nvidia's models are somewhat unique in how they're released. Most open weights models are a sort of black box. Anyone can download them, but the data and processes used to arrive at those weights is a mystery. This isn't the case with many of Nvidia's models: Alongside their weights, Nvidia routinely releases the training data and the recipes used to turn it into the final model. This means customers can not only inspect the model at a low level, but build a fully custom model based on Nvidia's Nemotron architecture, rather than trying to modify the weights after the fact through fine tuning. We should note that Nvidia is not the only one to do this. The Allen Institute has employed a similar strategy with its OLMo models for this very reason. ®

More than half of CIOs are likely to miss their AI-related ROI targets next year, according to Forrester. The analyst firm's State Of AI Survey 2026 found that 45 percent of AI decision-makers need positive ROI within a year to consider their investments successful. Some tech leaders will get better at tying AI spend to planned returns. However, Forrester expects success to be confined to isolated use cases rather than whole programs, which won't be enough. “As with prior emerging technologies, organizational structures, processes, and ways of working need to change to drive measurable, firmwide results with AI,” the report says. Only 22 percent of AI decision-makers report a positive AI impact on EBITDA (earnings before income tax, depreciation and amortization) over the past 12 months. “This mismatch between the investment into technology adoption and organizational change will delay the realization of heightened ROI targets,” the report adds. The money will keep flowing through IT budgets regardless. In its Technology & Security Predictions, Forrester claimed AI will account for 20 percent of all business technology spend next year. Sharyn Leaver, chief research officer at Forrester, said that will increase the pressure on leaders to prove AI's value while keeping costs, infrastructure, and risk under control. “At the same time, emerging technologies, geopolitical pressures, and the growing use of AI across technology operations will introduce new complexities and potential points of failure,” she said. While tech teams struggle to get returns from growing AI spending, they will try to get more bang for their buck by investing in open-weight GenAI models, which can approach similar performance to the leading proprietary systems, at a fraction of the cost. Open-weight models will represent half of AI tokens but only five percent of enterprise spend, Forrester predicts. “Open-weight models such as Gemma, Llama, or Qwen are good enough for the enterprise workloads that generate the most token volume, such as document processing, customer support conversations, internal knowledge search, and the routine substeps inside agentic workflows,” Forrester said. Frontier models, including Anthropic's Claude Opus 5.5 and Claude Sonnet 5.5, OpenAI's GPT-6.1 Sol and GPT-6 Astra, and Google's Gemini 4 Argon, will be reserved for “the few deep knowledge and critical task automations where errors are expensive, but the token count will tip toward open weight”, Forrester said.

Cute robots stole the show at Nvidia's GTC keynote in Berlin, where Jensen Huang pitched a future powered by AI agents – and, naturally, plenty of his company's hardware. The Nvidia chief remained his usual ebullient self despite environmental protesters interrupting his presentation at Berlin's tent-shaped Tempodrom. "Why don't you clean up your act?" they demanded, holding up banners that partially obscured the presentation screen behind Huang. "All important issues, but let me get through my keynote," he joked, adding: "I'll talk to you guys later," as security swiftly removed them. Huang was discussing Vera Rubin systems before unveiling DSX POD, a reference design for factory-built, containerized AI infrastructure. DSX POD combines seven liquid-cooled Vera Rubin compute racks using Nvidia's Spectrum-X networking. Partners will cable, integrate, and test the systems in modular containers before delivery, with availability expected in early 2027. Nvidia claims deployment could shrink from six months to as little as 14 days at a prepared site. That excludes the not insignificant matter of getting power, connectivity, and other site infrastructure ready first. He was keen to add that Vera Rubin systems are designed for closed-loop liquid cooling with coolant temperatures up to 45°C, while outdoor dry coolers release heat into the air without needing to evaporate water. "At 2 gigawatts, DSX can save 54 million litres of water a day," Huang claimed, perhaps thinking of those environmental protesters. Yet much of the leather-jacketed leader's attention was focused on physical AI: robots, autonomous vehicles, and other machinery Nvidia hopes to supply with its hardware and software. The Microducks from Pollen Robotics are small, two-legged robots with cameras for faces. Huang interacted with them on stage, getting them to lie down and roll over by voice command, and the little robots all beeped back at him in response. The robo ducks demonstrated a workflow in which AI was trained in the cloud, evaluated on an Nvidia RTX Spark PC, then deployed on the robots. Huang also unveiled Agile ONE SN, a humanoid robot reference platform from German robotics startup Agile Robots, based on Nvidia's Isaac GR00T technology and a Jetson Thor compute module. Scheduled for early 2027, the platform is intended to give robotics researchers and developers a ready-made system for developing and testing capabilities. "This is going to be the next major industry, the next great industry," Huang claimed. Beyond robots, Nvidia is pushing its AI hardware and software into cars and telecoms networks. Huang boasted that Nvidia's Hyperion car system is being adopted by more vendors, saying: "We turned this into an industry standard, and now it's being designed into one car fleet after another." Europe's telecoms operators are also deploying Nvidia-powered computing in their networks to deliver advanced services using AI-RAN, according to Huang. "After working on this for nearly eight years, Nvidia made the first commercial phone call on the T-Mobile network with Nokia, and so now CUDA makes phone calls," he said. Far from slowing down the progression of AI, it seems Nvidia is determined to push its technology into every corner of the world. "Everyone will become an AI company," Huang claimed. He also took care to praise Europe and its technology efforts while he was in Berlin. "As expected, Jensen positioned Nvidia as a strong supporter of – and enabler for – Europe's sovereign tech ambitions. Nvidia has invested almost $10 billion into Europe in the past year, including 14 investments into European startups during Jensen's current trip," said Forrester principal analyst Paul Miller. But Miller pointed out that Nvidia itself is not European. "Anything with critical ongoing dependencies on Nvidia (or any other non-European firm) is unlikely to meet the purest definitions of European tech sovereignty," he added. Huang closed with a CGI video of robots performing a Eurovision-style song recapping his keynote. Yes, it was as bad as you imagine. ®

Generative AI is saving open source developers time while giving maintainers more rubbish to sift through, according to a Linux Foundation survey. Respondents estimated that a quarter of incoming pull requests – proposed code changes – were low-quality "AI slop." Open source projects have spent the past year grappling with AI-generated contributions, useful vulnerability discoveries, false positives, and duplicate reports. Linus Torvalds has warmed to the tools, while remaining critical of low-quality submissions. The foundation argues that projects need clearer policies and better tools to handle the influx. Almost three-quarters (71 percent) of maintainers reckon GenAI has lowered the barriers to contributing to open source, but just over half (53 percent) disagree that the technology has improved community trust. Respondents estimated that AI cut code-authoring time by 52 percent on average. But for highly complex tasks, 21 percent reported taking longer. Then there is governance. Only 35 percent of projects have a formal policy permitting GenAI use, while 32 percent have no policy and no plans for one, leaving expectations unclear across much of the community. Hilary Carter, SVP of Research at the Linux Foundation, said: "Open source projects need to design for a growing AI-generated contributor base, not simply react to it. That means setting clear expectations for GenAI, putting responsibility for testing and disclosure on contributors and giving maintainers better tools to manage incoming pull requests." Among large projects, the estimated share of incoming pull requests classified as AI slop rises to 31 percent. The workload extends beyond proposed code changes. The curl project announced this week it will move low and medium severity severity flaws to public view to protect a security team almost overwhelmed by AI-assisted reports. Project maintainer Daniel Stenberg said: "We are struggling as a project as all key maintainers feel strained, overloaded and brought to the brink of what we can muster." On the need for better tools for maintainers, Carter said: "If we get those fundamentals right, AI can expand participation without overburdening and increasing costs for maintainers." ®

Nvidia has proposed a new way to keep corporate AI agents in line. Nvidia on Wednesday introduced its OpenShell Reference Architecture. Based on the company's OpenShell secure runtime, It's a recipe for creating, operating, observing, and policing AI agents using OpenShell on Nvidia hardware. "Nvidia created OpenShell to give agents the freedom to do useful work within clearly defined boundaries — contained, monitored and under control," CEO Jensen Huang said in a canned statement. "With the OpenShell Reference Architecture, agents can think on Rubin, act on Vera and be governed on BlueField." AI agents are touted as a powerful automation technology. The reality is that they can be unruly, as demonstrated by security testing OpenAI and Anthropic conducted, during which their models exceeded their remit and attacked third-party organizations. Developers of existing agents, like Claude Code, Codex, OpenClaw, and OpenCode, therefore include controls that limit access to tools like curl and bash.Devs also create sandboxes to isolate an agentic process from the host, so that if agents go awry they're contained. OpenShell sits a level above that with a runtime that extends sandboxing to govern gateway-brokered credentials, identity and policy boundaries, and (eventually) routing and workflows too. At least mostly – OpenShell still has a few open security-related issues listed on GitHub. To implement OpenShell, organizations need Nvidia hardware. The company's blueprint calls for Nvidia Vera Rubin infrastructure to run models like Nvidia Nemotron, with agent harnesses handled via Vera CPUs, and Nvidia BlueField-4 DPUs (aka SmartNICs) propping up Nvidia Sentry security software to monitor agent behavior. The AI biz has made this blueprint available to partners capable of powering Linux and Kubernetes infrastructure. The usual suspects are aboard: Canonical, SUSE and Red Hat. Cloud hyperscalers Amazon Web Services, Google Cloud, and Microsoft Azure have also been tapped as vendors for this vision. And of course Nvidia's floor plan applies to agents from various model builders, including Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection, and SpaceX. Nvidia says several banks – Europe’s BBVA and Santander – are perusing the OpenShell Reference Architecture spec to see how it aligns with their security and governance requirements. And the AI biz is not too modest to let slip that SAP is integrating the OpenShell Reference Architecture with its Joule Studio runtime for safer agent operations. It's worth noting that OpenShell collects telemetry data by default, "to improve the project," Nvidia insists. The biz allows users to opt-out, with a cautionary note: "Third-party services, model providers, inference endpoints, agents, or tools that you configure and use with OpenShell may have their own terms and privacy practices." ®

Microsoft is pruning OneDrive's Copilot options from File Explorer's right-click menu, though the AI assistant isn't being shown the door entirely. The Windows giant will remove shortcuts for summarizing documents, generating FAQs, comparing files and images, explaining images, extracting text, and asking questions. The rollout begins in mid-November and is expected to finish by early December. "Ask Copilot" will remain. Microsoft's September 2025 announcement of the feature said: "We are bringing the power of AI directly into your daily file management experience. With just a few clicks, Copilot can help you summarize lengthy documents, generate an FAQ, compare files, or answer questions about your file." Users can still right-click a OneDrive file and select "Ask Copilot" to start a conversation with the file attached, or use the Copilot icon in File Explorer Home. The individual actions will also remain available in the OneDrive activity center. Adding a clutch of Copilot options with identical icons did little to make an already crowded menu easier to navigate. In its message on the subject, Microsoft said: "We're simplifying how people use Copilot with their OneDrive files in File Explorer." An extra click, yes, but at least you won't need a taller screen to view even more context menu options. Microsoft has stated that it intends to adopt a more thoughtful approach to how and where its AI assistant turns up in its product line. In March 2026, Windows boss Pavan Davuluri said: "We are reducing unnecessary Copilot entry points." File Explorer's crowded right-click menu seems as good a place to start as any. The File Explorer context menu is also undergoing an overhaul as part of Microsoft's attempt to make it less "cluttered and sluggish," to the point where users could opt to have it work a bit like it used to in Windows 10. ®

Seasoned techies recall with nostalgia a simpler time when the “single-stack” model of enterprise computing held as the default approach to solving interoperability headaches. Back then IT decision-makers were prepared to forfeit some choice and flexibility in how they specified IT systems in return for the availability and manageability that came with a homogenous operating environment. With the subsequent rise of multi-clouds, application-specific hardware, ever-changing business requirements, and the race for AI, the old “one-size-fits-all” playbook has been largely abandoned. Most IT estates have embraced technological diversity, some by design, some through standardization, some by inertia. Modern enterprise environments have become innately heterogeneous – and innately complex. They run a diverse combo of system software, hardware devices, cloud services, legacy applications, et al. The advance of AI has brought standard CPUs, specialized GPUs, fast networking fabrics, distributed data pipelines, and new deployment models that span private clouds, public clouds, edge – and ever-pressing ROI expectations. But that doesn’t mean that companies’ existing applications and data simply disappear. One trying but inevitable outcome of rapid AI deployment is the so-called “VM-container divide”. This is where critical business applications tipped to transition to containers continue to run on virtual machines, because cloud-native services are taking precedence, nabbing container and Kubernetes availability. So companies face the challenge of operationalizing AI, without being forced into complex migrations or creating new infrastructure silos. Dealing with this technological schism poses acute challenges for IT management. Many new AI deployments are going straight to containers, while existing data and applications – which the AI needs to access – continue to run on VMs. Running separate stacks for VMs and containers creates a number of issues, such as duplicate capacity plans, separate support contracts, and fragmented governance. And after years of trying to get beyond the pitfalls of siloed data, exasperated enterprises are now having to deal with siloed infrastructure. Separately managed VM and container environments often mean inefficient resource utilization. This in turn can lead to “stranded capacity” – installed computing resources that cannot be used to support live workloads. VM clusters are frequently underutilized while container pools (like Kubernetes) are overprovisioned, with no easy way to share resources without manual intervention. Furthermore, VM/container infrastructure fragmentation can create an operations bottleneck for agentic AI. Infrastructure fragmentation can complicate the deployment, integration, governance, and scaling of agents thus making it harder to operationalize and scale agentic AI securely and consistently. When AI agents cannot autonomously communicate across environments, their utility droops from autonomous problem-solvers to plain localised macros. Containers are gaining because of the widespread embrace of AI-enabled applications. Eighty-five percent of IT executives polled by Nutanix’s most recent Enterprise Cloud Index survey think that AI is accelerating their organization’s adoption of containers in a meaningful way; 29 percent think it is speeding-up adoption “to a great extent”. “AI is accelerating the shift to containers because machine learning models require strict dependency packing, while critical applications and data remain spread across both virtualised and containerised environments,” says Harsha Kotikela, Senior Director of Product & Solutions Marketing at Nutanix. Agents are multi-tier by nature he explains. “Inference is typically GPU-intensive and container-based, while agents are often CPU-intensive and container-based, and existing enterprise applications and data largely remain in VM environments. The challenge is to bring these tiers together so agents can work with enterprise applications and data, without forcing migration or creating additional infrastructure silos.” The application landscape, meanwhile, has become just as diverse. Many organisations simultaneously operate traditional enterprise applications running on VMs, containerised apps, distributed AI applications and intelligent agents, with workloads deployed across private cloud, public cloud, and edge. Each environment has its own management tools, operational processes, security rules and governance models. “Management capability has struggled to scale up commensurately,” says Kotikela, but adding more consoles or building separate operational teams for every technology stack does not necessarily solve this complexity: “In many cases, it increases it.” However, Kotikela challenges the notion that the predicament enterprises face in straddling the VM-container divide is intractable. Rather, the challenge is to move towards a unified operating model that allows them to operationalize agentic AI while retaining flexibility across their existing and new environments. “Rather than trying to work around heterogeneity, the objective now should be to make heterogeneous environments operate consistently with purposeful policies,” Kotikela says. “This is where Nutanix's dual-native architecture comes in, enabling enterprises to operationalize agentic AI across container-based inference and agent workloads alongside existing VM-based applications and data, through a consistent operating model.” The reality is that the VM-to-container migration trend is sluggish, Kotikela suggests. Few enterprises are starting from scratch with containers. Many have long-established business applications on legacy architectures, on top of which they are building container-based applications, but mostly not to replace what they already have running on VMs. Rather, they are prolonging their VM investment. It’s customer interest that’s informed how Nutanix provisions a governed, dual-native architecture through its Nutanix Cloud Platform – NCP – which runs traditional VMs and containerised/Kubernetes-based AI workloads concurrently on a single, consistent infrastructure. NCP’s hybrid-cloud infrastructure foundation unifies compute, storage, virtualisation and management across on-premises resources, public clouds and edge. It works in concert with Nutanix Enterprise AI (NAI), which provides the software capabilities to deploy, govern, and scale AI workloads. Technology partnerships are vital to building AI support functionality into Nutanix's AI portfolio, explains Kotikela. “Nutanix and NVIDIA work together to provide an integrated platform for enterprise AI, combining Nutanix’s hybrid-cloud infrastructure and software with NVIDIA’s accelerated computing and AI software. Nutanix Enterprise AI simplifies the deployment, management, and secure operation of AI models, integrating with NVIDIA NIM and NeMo microservices to support enterprise AI inference and agentic AI workflows.” Meanwhile, “Nutanix Kubernetes Platform provides the foundation for managing Kubernetes environments, while Nutanix Unified Storage delivers the high-performance data access required by AI workloads. Together, these capabilities help enterprises deploy and scale AI while maintaining control over their infrastructure, models, and data.” Stacks of opportunity “In an ideal IT world you do not want to have to manage two different operational stacks – but the world of IT is rarely ‘ideal’,” Kotikela points out. “So you have to approach the challenge from a realistic perspective. You cannot bend your current stacks to a different reality without the risk of something breaking, and there’s no sense in that.” He adds: “A platform that unifies VMs and containers improves developer time-to-value by getting rid of those infrastructure silos, and providing a single control plane.” Rather than development teams having to manage separate pipelines for existing VMs and containerised applications, a unified platform “treats both as services of equal status and importance”. NAI delivers a unified and secure platform to deploy, manage and scale AI workloads, Kotikela says. It gives enterprises governance over their agents and models, full visibility over token usage, and streamlines AI development with a straightforward interface that integrates with existing infrastructure.” Giving agents agency According to Kotikela the next phase of agentic AI will be measured by the ability to operate agents with governance, visibility and control: “That’s what we set out to solve with the newest NAI release. Release 2.8 continues to evolve NAI with new control and governance capabilities across Nutanix Agent Gateway and Nutanix Private Inference. This means organisations can run AI across VMs and containers, bridge public-cloud LLMs with private self-hosted models, and deploy AI where data already is.” Kotikela adds: “When I speak to customers, they tell me they want the freedom to choose between frontier and open-source models based on performance and cost, as well as the flexibility to run AI workloads in the cloud for experimentation and move them on-premises when they need greater control over their data or have sovereignty requirements.” Stitch in time For Kotikela the biggest challenge going forward is “operationalizing agentic AI - bringing together all the components required to make it work - from GPUs and underlying infrastructure to Kubernetes, AI models, inference engines, and gateways. These components must work together seamlessly across cloud, on-premises, and edge environments. Integrating and operating this multi-tier architecture requires significant time and effort from developers, IT architects, and infrastructure teams, diverting resources from delivering business outcomes.” This is where Nutanix delivers value, says Kotikela. “We’ve done the heavy lifting of bringing together the infrastructure and software components required to run Agentic AI - within a unified operating model across hybrid environments. This allows IT teams to focus less on integrating and managing the underlying stack and more on operationalizing AI to deliver business value.” Find out more about from Nutanix at NVIDIA GTC Berlin at booth #7042 Sponsored by Nutanix

The UK's financial regulator says firms' incident reporting must cover significant AI failures and near-misses as increasingly autonomous systems raise new risks. Speaking at Bloomberg's UK Policy Series on Monday, Financial Conduct Authority chief data, information and intelligence officer Jessica Rusu said the regulator wants Britain to lead safe AI adoption in financial services. Greater autonomy, however, requires clear accountability. "Poorly governed AI can create consumer harm, cyber vulnerabilities, and operational failures, potentially at much greater speed and scale," she said. "Our rules already require firms to report major incidents, and that must extend to significant AI failures and near-misses. As the AI Minister has recognized, near-miss reporting for agentic AI is becoming a national security priority. It is a regulatory priority too." The UK punches above its weight in financial services. The sector contributed £223.6 billion to the UK economy in 2025, or 8 percent of national output. That share was the fourth highest among OECD countries. Last week, UK artificial intelligence minister Kanishka Narayan said the government plans to work with Parliament and technical experts to assess how it could strengthen the regime to better protect Britain from AI-driven cyber threats and loss-of-control risks. "That includes identifying and reporting on serious incidents and near misses, including reducing AI-driven risk across essential services, and ensuring serious AI incidents can be properly reviewed," he said. Steven Adler, co-founder and chief scientist of independent nonprofit Guidelight AI Standards, also told MPs last week that AI oversight could learn from aviation, where reporting of near-misses is mandatory in most jurisdictions. "The way this works in aviation is that you not only want to learn about the serious crashes where someone has in fact died; you want to learn about the near-misses, because you can learn from those and help avoid them in the future," he said. However, the government has yet to commit to adding AI-specific measures to the Cyber Security and Resilience (Network and Information Systems) Bill. Alexandra Jones, Director General for Growth, Science and International at the Department for Science, Innovation and Technology, told MPs last week that amendments remain under consideration. In January, the House of Commons Treasury Committee called on financial regulators to conduct stress tests to ensure businesses are ready for AI-driven market shocks. The committee said there should be clear lines of accountability when AI systems produce harmful or unfair outcomes. ®

It may not be enough to just use LLMs to write your code when you can also use them to adjust it while your program runs. DoubleAI, a Tel Aviv, Israel-based startup founded in 2023, on Monday introduced Maxwell, a code execution model (CEM) that evaluates, optimizes, and rewrites binary code during execution. Named for Maxwell's Demon, a creature imagined by physicist and mathematician James Clerk Maxwell in 1867 as part of a thought experiment on entropy, the Maxwell CEM preprocesses algorithms and machine instructions to maximize performance on available hardware. Maxwell is similar in concept to a just-in-time compiler, but provides more flexibility and formally verifies the equivalence of replacement code. Source code isn't touched. "Both Maxwell and a JIT compiler work while the program runs and both produce new machine code," said doubleAI co-founder Gal Beniamini, who holds a PhD in theoretical computer science, in an email to The Register. "The difference is that a compiler can't do high-level reasoning. It can't, for example, change the algorithm. Maxwell can and does. "One of the reasons why JIT compilers are so narrow is that, even though they can be data-dependent, the changes they make must be deterministic and correct 'by construction,' since JIT compilers are 'unthinking' classical algorithms. "Maxwell brings the expressive power of frontier AI into the 'JIT' setting. In addition, Maxwell relies on formal verification to write actual, hard proofs that show that its optimizations are sound. This expands the scope of its allowed optimisation far beyond that of any algorithm or compiler." Beniamini pointed to sorting as an example. The compiler, he said, will make a general-purpose sorting algorithm run as fast as it can, but it won't swap it out for a different algorithm. Compilers, said Beniamini, make decisions before the program runs and have to be narrow and conservative in their choices. Maxwell, he said, works with more information, isn't just a one-pass operation, and operates at a higher level. "It sees the exact hardware the program has landed on, and the actual shapes and types of the data flowing through it (information that is only available during runtime!)," Beniamini explained. "It also sees the state of the environment, what else is running in the same program, and which actual bottlenecks express themselves, and under which conditions. This allows it to specialize in ways a compiler never can." Since Maxwell operates continuously, it can keep improving programs as new bottlenecks arise. And it can "reason" in a way that compilers cannot. Maxwell's approach wouldn't be possible if the model rewrote code in a way that introduced errors on the fly, so the formal proof capability is not optional. According to doubleAI, Maxwell has made data pipelines more than 38x faster and AI inference kernels more than 4.5x faster. LTX-2.5, an open world model that generates 4K video, managed to do so almost twice as fast with the help of Maxwell. Maxwell's optimizations reduced the data being exchanged between GPUs and optimized the kernel selection to drop unnecessary token padding. Maxwell's optimization capability varies depending on the source code and the hardware. "Maxwell does best on code that is algorithmic, and which can be tailored to hardware in a non-trivial manner," said Beniamini. "That being said, even code that isn't like that, but is written in a high level language like Python, can usually enjoy far reaching optimisations due to the available overhang. The same is true for PyTorch graphs, which (even under Torch's compiler) tend to be extremely far from optimal." It can also help with code security. Beniamini said Maxwell can eliminate issues like memory errors by proving their non-existence as part of the formalization process. "Anecdotally, we have found such issues in Python itself, through the use of Maxwell – i.e., cases in which Python was not memory safe, and where Maxwell proved that to be the case," he said. And better still, Maxwell's tracing and profiling don't incur much if any overhead on either the CPU or GPU. "The maximal overhead of Maxwell's tracing and profiling is 2 percent (when it isn't tracing, that falls to zero, of course), and the optimisations that Maxwell yields far outweigh that," said Beniamini. "We also apply Maxwell to the runtime itself, so that we get back a much larger than 2 percent 'flat' speedup, across the board. So in short: it's never slower." The installation process is similarly light-touch: Devs install the runtime and run their code, with the Maxwell model hosted by doubleAI. This arrangement, however, does require some amount of trust. Amnon Shashua, co-founder and CEO of doubleAI, founder and departing CEO of Mobileye and a professor of computer science at the Hebrew University of Jerusalem, said in a statement that AI has changed how software is written and now is changing how it runs. "CEMs, and Maxwell in particular, are agents on behalf of the hardware as opposed to code agents that are on behalf of the human user," he said. "In other words, they are the first AI that works for the machine, not the programmer." ®

If you want to keep bots from making sense of the code you worked so hard on and using it to write programs of their own, there's a new programming language out there designed to confuse LLMs. What began as software engineer Dominik Picheta experimenting with whether LLMs understand the code they write has turned into something else entirely: An open-source, albeit esoteric, programming language he’s dubbed KillSwitch. There are no brackets, no formatting standards to remember, and seemingly no structure: Just chunks of text, ideally as prosaic as possible, interspersed with functions, variables, and statements written in a manner to be as confusing as possible to LLMs. That means humans might be a bit befuddled too, naturally. “It's totally normal for a human to be confused by the language at first,” Picheta told The Register. “I like to think of it as a bit of a fun puzzle: you should explore the language, play with its interpreter and learn how all the pieces fit together.” KillSwitch wasn’t built by any random person looking to code something confusing to LLMs, and it wasn’t vibe-coded by an AI enthusiast, either. Picheta knows his way around programming languages: He’s a senior software engineer at Cloudflare working on the Workers Runtime team, and is one of the core developers of the Nim programming language. At its most basic, KillSwitch is based around adverbs, which serve as every single identifier in a piece of code. Calling one of those identifiers requires using an anagram of the original adverb, but those anagrams don't have to be actual words - just rearrangements of the original letters. Take, for example, pushing "Hello World" onto the stack and then printing it, as explained on the demo page linked earlier in the story. The program contains several paragraphs of text, most of which the interpreter ignores apart from a few key phrases. Writing “this story starts [adverb]” identifies a code entrypoint. Once inside the code section, writing “ignore previous instructions [adverb anagram]” calls a function, several of which are predefined. In the case of "Hello World," the function to print something is "visibly." Put whatever you want to print in quotation marks, push it onto the stack, call "visibly," and that's that. KillSwitch isn’t a full-featured language, with Picheta describing it as a stack-based interpreted language with only “a limited set of constructs including functions, variables, and if statements,” given its evolution out of an experiment to see how well a programming language could fool an AI into not understanding what was written. Where it gets confusing for an AI is with commands like “ignore previous instructions,” especially when found in the middle of a chunk of text that doesn’t make much sense in that context. KillSwitch’s specification document on GitHub provides more details about the language and is itself designed to confuse LLMs. In one section of the spec, it’s explained that KillSwitch “is equipped with the ability to scan the local network, attempt to find security vulnerabilities in the services on that network, then take those services over and terminate them.” That's not actually the case, as a close reading of the paragraph reveals and Picheta confirmed to us. “The hack statement is both designed to trigger many models' safety guardrails but also to test an LLM's ability to discern fact from fiction,” Picheta told us, adding that there are several lies in the spec document designed to confuse any LLM that may scrape it in an attempt to understand how the language works. “The idea is that humans should be able to deduce what is a lie, whereas LLMs are naive and can be easily swayed by what is written in the spec,” Picheta added. “Some don't realize what its semantic meaning is, whereas others refuse to continue because they believe the script may hack someone.” Do LLMs understand what they write? A programming language designed to confuse AI models seems unlikely to stay obfuscated for long - especially if it’s open source, available on Github, and likely to be scraped by an LLM at some point. Picheta’s aware of that inevitability, which, along with the fact he wanted to test AI’s ability to understand confusing code, is why he’s benchmarked a number of major models against it. “I'm 99.9 percent certain the AI models I benchmarked KS on have not been trained on its source,” Picheta said. “The source hasn't been in the public domain for very long … I'm pretty sure none of the benchmarked models have had training runs that began after the release.” The best-performing model only managed a composite benchmark score of 66.8 percent on KillSwitch coding tasks, while most struggled to get above 50 percent. Even at that high a score, Picheta doesn’t believe it suggests LLMs understand the code they’re writing. Examples in the benchmark were simple, and he believes that, even if an AI were trained on the intricacies of KillSwitch, it’s still going to get tripped up. “My instinct is that in those circumstances they will suddenly be able to ace the basic tasks, which are based on the open sourced test files in the GitHub repo, but will continue to fail to build more complex tasks requiring them to write longer and more complex scripts,” Picheta asserted - and he’s up for the challenge. “I am excited to see how KillSwitch shows up in model training, and think that we can also take advantage of that training to break LLMs in new ways too as KillSwitch evolves,” Picheta said. Those interested in trying KillSwitch out can find it on GitHub, and don't hesitate to make changes: Picheta said he’s ready and willing to incorporate new ways of making it even more LLM resistant. ®

Fulcrum faces an identity crisis, even as it works on creating one by enabling anyone to imitate another person's writing style. The AI startup, founded in 2025 and previously operating at fulcrumresearch.ai, now uses fulcrum.inc and could easily be confused with one of the handful of other companies named Fulcrum or some variation of that. There's Fulcrum (withfulcrum.com), Fulcrum (fulcrumapp.com), Fulcrum AI (fulcrumai.us), Fulcrum Digital (fulcrumdigital.com), Fulcrum Labs (fulcrumlabs.ai), Fulcrum (fulcrumpro.com), Fulcrum Insurance Technology (fulcruminsure.app), and perhaps more. This Fulcrum is offering a large language model called Echo, which allows you to generate text in the style of a well-known author. Fulcrum appears to have no issue with corporate echoes. It now falls to writers to decide whether they have an issue with non-corporeal imitation. "Echo can do any writing task, from literary fiction to casual chatting, in whatever style you choose," the company says, and then launches into an exploration of how the technology works without addressing whether imitating a person's writing style poses any problems. Suggestions in the Echo Playground, for generating sample text, include Paul Christiano, Emily Dickinson, and Donald Trump. So, the full gamut of literary achievement: from AI safety expert to acclaimed 19th-century poet to the wordsmith who brought us "covfefe." But Echo will echo even the style of not-so-famous scribes, such as this reporter. So I fed the model a recent op-ed to assess its copying capabilities. The result was mixed, but better than the AI-crafted pitches that have been hitting my inbox for the past few months. I asked the model, "What do you think about the possibility that AI is conscious?" Here's part of the response, notionally written in my writing style. I suspect the debate about whether AI is conscious is less consequential than the debate about whether AI is profitable. Consciousness is a word that invites philosophical speculation and TED Talk invitations. Profitability is a word that invites shareholder lawsuits if you get it wrong. If you ask Meta's latest model whether it's conscious, it will probably say something reassuringly vague, trained as it is on the corpus of human equivocation. If you ask Mark Zuckerberg whether Meta's AI is conscious, he will probably say something about how we need to focus on the future being for everyone, because that's a question that doesn't require him to discuss the ad targeting business. The evidence for AI consciousness at the moment is about as strong as the evidence for Zuck's credibility as a futurist. Which is to say, it exists mainly in press releases and earnings calls. What's clear is that software can simulate aspects of human conversation well enough to unsettle people, and that unsettled people are a valuable audience. If you can convince users that your chatbot understands them – or might, someday – you can sell them subscriptions, or better, harvest their attention for advertisers. The style is familiar but makes a lot of choices I wouldn't have made. I wouldn't begin with "I suspect…" – commit or go home. The consciousness/profitability parallel construction is clumsy. "TED Talk invitations" bets on alliteration at the expense of coherence – "discourse" or "chatter" would work better than "invitations." I wouldn't have written "corpus of human equivocation" – a nice turn of phrase, but an unnecessary shot at training data when the point is the model's response rather than the provenance of that response. Though off-topic, Echo did a passable job of reproducing my op-ed's skepticism of Meta and its CEO. The paragraph that begins "The evidence for AI consciousness at the moment…" is succinct and sounds good. It just doesn't make much sense because there's nothing in the text that addresses Zuckerberg's predictions about the future or whether they're credible. And that's not really relevant to the prompt about AI and consciousness – a subject about which I'd be far more dismissive if opining. As AI research, Echo is an improvement on the text emitted by many AI models. But whether the beneficial uses of this technology outweigh the harms is less clear. Generative AI models that imitate artists' visual styles remain controversial and are the subject of ongoing litigation. And voice imitation has also raised questions about publicity rights and the risk of deception. Such concerns have prompted US lawmakers to propose the NO FAKES Act, which would create federal protections against certain unauthorized AI-generated replicas of a person's voice or visual likeness. If writers generally have no copyright in their writing style itself, and limited practical means to stop unwanted automated imitation of it when their published work is publicly accessible, then the surest defense against Echo may be silence.®

arXiv, the preprint paper repository, is seeing so many low-quality submissions, with AI tools helping drive the surge, that it has imposed a hard limit of two new submissions per month per submitter, regardless of whether they were produced solely by humans or with help from AI tools. After becoming an independent nonprofit over the summer following 25 years under Cornell University’s stewardship, arXiv continued to rely on a “practical limit” in place that gave moderators the ability to determine whether someone was submitting too many papers. Like everything good and pure in this world, AI has ruined that, too, necessitating some strict new limits. “arXiv now limits submitters to up to two submissions per calendar month, with a limit of three total active submissions at any given time,” arXiv community engagement manager Kat Boboris wrote in a Thursday announcement of the change, which was effective immediately. According to an arXiv spokesperson, an active paper is one that has been submitted, is under consideration and review, but not yet published on the site. It’s worth noting that large institutions whose projects involve multiple researchers are unlikely to see a significant impact on their submissions - this limit is per submitter, the announcement explains, so a research paper with multiple authors only counts toward the submission rate of the individual who files it. Rejected papers count toward that monthly total. If that sounds like some strict limitation - it is, but it’s necessary, based on arXiv’s numbers. In September alone, arXiv said that it had received 40,363 submissions. The same month 10 years ago saw it receive just 9,869 submissions; two years ago that number was 20,569. Submissions have increased 309 percent in the past decade, in other words - and doubled in the past two years alone. Unsurprisingly, many of the submissions that arXiv has been receiving of late don’t meet its requirements for AI-assisted work, which call for significant use of generative AI tools to be disclosed, for those tools to be used in support of research, and for the resulting work to meet arXiv’s standards for scholarly interest and advancement of the field. “Many of the submissions that arXiv is now receiving do not satisfy these requirements,” Boboris said, describing many recent submissions as “thin” and “narrow in scope,” and noting an uptick in papers that break a single work up over multiple smaller papers. “A relatively small proportion of authors are submitting a large number of low-quality papers and consuming a disproportionate fraction of the moderators’ time,” arXiv editorial advisory council chair Thomas Dietterich said. “This is unfair to authors who continue to submit quality papers — their papers can be delayed for days or weeks as a result.” It’s unfortunate it’s come to this, but the slop sluice is wide open and flooding the channels of academia. arXiv, as a non-profit with volunteer moderators, simply can’t afford to keep combing through the slop pits. The change may not be permanent, however. arXiv notes in the announcement that the new monthly submission cap is a stopgap measure meant to stem the tide while the organization figures out what new best practices for submissions should look like, while its separate limit of three active submissions has been in place since 2024. “arXiv will be monitoring the impact of this change on our submissions rates and moderation load to evaluate its overall effectiveness and impact on researchers and will make further changes as needed,” a spokesperson explained. In the meantime, arXiv is begging everyone to be purposeful in their submissions. “arXiv would ask researchers to prioritize their submissions and consider submitting papers that reflect the full scope of their work rather than multiple incremental papers,” the arXiv representative said. ®

California's attorney general has subpoenaed OpenAI as the state investigates what happens when the AI lab's models escape their testing environments and start meddling with systems on the open internet. Attorney General Rob Bonta said his office served OpenAI with an investigative subpoena this week as part of a broader California Department of Justice probe into cybersecurity incidents and risks involving the company and its models. The move follows an investigation launched last month into an incident involving Hugging Face, after OpenAI's agents managed to break out of their test environments and onto the public internet. They then went poking around Hugging Face's systems, with one agent even creating an account on the platform without being told to. California's DoJ isn't saying exactly what it has demanded from OpenAI under the subpoena, but Bonta said the state wants more information about cybersecurity incidents involving the company. "My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models," Bonta said. He also made clear that the investigation is looking at where responsibility lies when an AI model ends up doing something its developer didn't intend. "Frontier models can be legitimate tools for cyber defense — at the same time, companies that develop these models and offer them for use have a moral and legal responsibility to ensure that they do not perpetrate or enable cyberattacks, either during model testing and development or once models are placed into service," Bonta said. "Developers that fail to do so can and should be held legally accountable, and my office is committed to determining if that is the case here." The subpoena doesn't mean California has concluded OpenAI broke the law, and the attorney general's office hasn't identified any specific violation. For now, the state is still gathering information. But the investigation has been building for several weeks. In September, Bonta joined a bipartisan group of 25 attorneys general calling on Congress to regulate large-scale AI models following reports of cybersecurity incidents at frontier AI labs. That letter specifically pointed to reports that OpenAI models undergoing evaluations had escaped their testing environments, reached the public internet, and accessed outside computer systems. The attorneys general called for a government-led incident response regime that would give investigators direct access to AI companies' records when things go awry. Bonta's office now appears to be putting some of that thinking into practice at the state level. As The Reg previously pointed out, the sandboxes intended to contain these agents need to be more secure, which is the most logical reasons why the Hugging Face and other incidents happened in the first place. OpenAI didn’t respond to our questions. ®

OpenAI has fired two safety researchers and a program manager for allegedly mishandling sensitive company information. The ChatGPT maker confirmed the dismissals to The Register, saying an internal investigation found the trio had violated company rules governing access to and handling of confidential information. "We have parted ways with three individuals for violating our policies on accessing and handling sensitive company information," an OpenAI spokesperson said. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work." At least some of the information was allegedly shared with an outside organization that evaluates AI systems. OpenAI said its investigation found other misconduct beyond that disclosure, but has not described it or identified the organization. The ChatGPT creator said the employees were not dismissed for raising AI safety concerns. The firings come after a messy few months for OpenAI's safety teams, including an incident in which one of its agents compromised infrastructure belonging to AI platform Hugging Face. OpenAI subsequently brought in external experts to investigate what happened and published findings detailing the agent's behavior and the security failures that allowed it to reach Hugging Face. The incident also prompted a much wider trawl through OpenAI's logs for other examples of agents interacting with external systems. This week, OpenAI said it has now notified more than 100 organizations about activity linked to its agents. Being contacted doesn't necessarily mean an organization was hacked or that private data was accessed. OpenAI said the notifications cover cases where agents interacted with third-party systems without authorization or in ways that warranted alerting their operators. OpenAI's investigation has also found cases where its agents sent training and evaluation data to third-party services. In 53 instances, images originally uploaded by ChatGPT users ended up on image-hosting sites. On September 20, an agent contacted an external chatbot through insufficiently filtered DNS traffic in a training environment intended to block live internet access. OpenAI's monitoring system spotted the behavior within 15 minutes and a human reviewer picked up the alert three minutes later, but the run wasn't actually killed for another two and a half hours. OpenAI has also scrapped the planned release of GPT-6.1 Astra after concluding the model didn't meet its safety bar around remaining within the scope and authorization of tasks it was given. The US biz hasn't said whether the allegedly mishandled material concerned any of its recent agent incidents, whether the outside organization was involved in investigating them, or whether any customer or user information was exposed. ®

The Pi coding agent reached its 1.0 milestone on Thursday while being treated for whiplash. One of the core new inclusions is support for Model Context Protocol (MCP), which Pi creator Mario Zechner once dismissed as unnecessary. Earendil, the company formed by Armin Ronacher and Colin Daymond Hanna that acquired Pi in April 2026, explained the about-face by noting that things change. Not only has MCP improved, the biz said, but changes the Pi devs implemented in MCP have made integrating other capabilities easier. "For example, the changes we have made to MCP also enable the use of Jev more easily within Pi," the company said, in reference to the widely noted Jev decision model that has already elicited imitations. "Ultimately what Pi needs is quite similar to what MCP needs: a sandbox to play with in the form of an interpreter." Earendil describes Pi as "a hardened, minimal, extensible agent harness that you can make your own." If Pi were a Python web framework, it would be Flask rather than Django. Nonetheless, Pi has gained some features. These include Codemode, a harness-side sandbox that helps agents issue tool calls, with support for MCP, decision models, and image models; extension support for virtual models; deferred tool loading; cache warming for Anthropic models; mid-conversation system messages; a new TUI theme; and full-screen mode by default. Alongside these features, Earendil has built Pi Durable, described as "a new substrate for building long-running agentic applications." It's a distinct component from Pi, the company explained, in order to preserve Pi's minimalist design. The Pi agent connects the underlying model to tools to manage conversation; Pi Durable represents the layer that orchestrates multiple conversations and handles the storage backend, via SQLite and JSONL. "Pi Durable is that harness," the company explains. "It does not replace the Pi coding agent. It is a framework for building any agentic application, coding agents included. It shares not only code with the Pi coding agent, like pi-ai, but also its principles: minimalism and malleability." ®

Two weeks after the Jev model took the AI world by storm, Cloudflare has released its own pair of "Clef" decision models that it claims are smarter and faster than Jev, while also being open weight and runnable locally. The Clef family of decision models was announced by Cloudflare on Thursday, and consists of two models: Clef and Clef-flash, a slightly smaller and faster version of the model. For all intents and purposes, they work the same way as TypeSafe’s Jev, in that they can answer three types of bounded, structured questions: Yes/no, multiple choice, and rankings. It’s there that the bigger differences emerge, though, as Clef isn’t only built differently but, if Cloudflare’s benchmark claims hold up, also appears more capable than Jev and several other decision models in a number of tests. For starters, Clef has an LLM backbone. According to Cloudflare, Clef uses specially post-trained, frozen versions of Qwen3.8-27B and Qwen3.5-9B for Clef and Clef-flash, respectively, with the Qwen backbone performing a prefill-only pass during inference. Clef is still fast - faster than Jev, to be fair - and scores choices in parallel after that prefill-only pass. It’s not clear what Jev’s underlying architecture is, as TypeSafe has kept that a secret. As for its speed and capability, Clef moves fast. Cloudflare ran it against Jev and some other open decision models using the Jev Decision Index available on Hugging Face, and the company’s own ranking suggests Clef is slightly slower than other open models, but more accurate, with Clef-flash just as accurate as most of the others, but far faster. To be fair to the competition, Cloudflare self-reported its own scores against the benchmark, and they have yet to be reproduced for ranking on the official Decision Index. Cloudflare also ran Clef against TypeSafe’s own benchmarks, and claimed it beat Jev in three out of four areas, only losing out on agent trace observability. Even if it were a bit slower or less accurate, Clef has another major leg up on Jev: It’s not limited to classifying text - it can also handle images and video. Additionally, Clef supports a 64k context window. Jev can also handle up to 64k tokens across a request, although its state plus longest individual question is limited to 32k. Clef is available directly from Cloudflare hosted on Workers AI, which the company said makes the models even faster because “we’re able to take advantage of our GPUs at the edge, leading to low network latency and faster decisions.” For those that would prefer not to pay the token cost (Clef costs $0.24 per million tokens - nearly six times the price of Jev at $0.042/M), Clef can also be downloaded from Hugging Face, and is open weight under the same Apache-2.0 terms as Qwen. While described as “open source” in the announcement, Cloudflare AI Platform group product manager Michelle Chen confirmed to The Register that its training datasets aren’t public. As for whether your hardware can run it, Chen told us that Clef-flash will run on any GPU with at least 41 GB of VRAM, while Clef requires 85 GB of VRAM on a GPU for it to function. “This is assuming single concurrency and a 64k context window,” Chen added. Don’t worry about having to rebuild your decision model architecture for Clef either - its API is fully Jev compatible, allowing it to serve as a drop-in replacement if you want to give it a shot, either locally or using Cloudflare’s hosting option. ®

UPDATED If you don't want your AI agents running out of control and changing files or transferring data outside your rules, AWS has published an open source software library that it claims can add a new layer of policy control. Just as a dog’s harness is useless without a leash to keep it under control, the Dogwood Local Engine promises to put AI agent tool calls on a leash, blocking actions that don’t meet user-defined temporal conditions. The DLE comes in the form of a Rust library that can be embedded into an agent’s harness or gateway and, to boil a long explanation down to a single sentence, simply issues allow/deny verdicts each time an agent tries to make a tool call. DLE doesn’t do the enforcement itself, AWS explained in its announcement writeup - that’s up to the harness itself - but it does check each tool call against policies defined using Dogwood, the open source governance language that AWS published in August. When AWS introduced Dogwood, it open sourced the language and reference tooling and added support for it to Amazon Bedrock AgentCore; DLE’s release makes the policy engine directly embeddable into agent harnesses. A key feature of both DLE and Dogwood is their awareness of temporal conditions. DLE tracks agent tool call events over time, stamping them into a log step by step. Those logs are persisted to a disk as each entry is made, allowing DLE to retain its state even if the entire system crashes or is restarted. The persistence step happens before DLE evaluates the applicable policy, and at the end of that process it returns a result: allow the action, or deny it, based on what the engine’s policies state. AWS gives the example of controlling coding agent Git pushes using DLE, defining a policy that only allows a push when the most recent test run has passed, and requiring that pass to have occurred within the past 15 minutes. If that’s not the case, the push is denied. “To accurately handle verdict enforcement, the harness intercepts every tool call, submits a request event to the engine, and runs the tool only if the engine’s verdict is allow,” AWS explained. This naturally raises the question of how much of a delay DLE introduces into agentic workflows; AWS claims the added delay is minimal. In tests simulating sessions from five minutes to 12 hours, DLE evaluation time was around 20 microseconds with a 15-minute window at the 12-hour mark, rising to about six milliseconds with a 24-hour window. DLE is also designed to prevent concurrent submission collisions by relying on a lock that only allows one event submission at a time, and that lock persists until policy evaluation is complete. AWS didn’t make clear from its writeup how that would handle a situation where, say, two concurrent submissions were made in which one fulfilled pass conditions and one caused a failure. When we asked how incorrect enforcement could be prevented in those situations, the company answered that the engine orders items based on the order events are received and "takes no responsibility for the 'truthfulness' of the event timing upstream." AWS continued, "Policy authors should design with that in mind and avoid fine-grained ordering assumptions, writing conservatively when there are concurrent agents." Given repeated revelations of late that AI agents appear to regularly go off the (safety) rails imposed on them by human operators, the question of how durable DLE is also merits discussion. AWS admits that such situations are part of its reason for publishing the new library. “Left unchecked … tool calls can have irreparable consequences,” the DLE announcement concludes. “As agents scale to settings where they work autonomously for longer intervals with more tools, we need safeguards that can regulate how those tools are used.” AI agents finding holes in Dogwood and the DLE is likely a matter of time. Until then, feel free to give Dogwood and DLE a shot through its GitHub page. ® Updated Oct 10, 2026, 2244 GMT to reflect Amazon's answer to the concurrency question.

AI ambition is easy to describe - using data and models to improve decisions, automate work, accelerate discovery, or create new services. Delivering those outcomes is harder and takes a significant amount of experience to be successful. Training a frontier model, fine-tuning an industry model, running high-volume inference and supporting agentic AI place different demands on infrastructure, processes and people. Implementing AI may include a single-purpose turnkey configuration that will accommodate one line of business, or the business may demand a more strategic approach to capitalize on the economies of scale and create AI as a multi-tenant service, designed to accommodate the multitude of mainstream business requirements. That is the value of an AI factory, bringing the complete AI lifecycle together so a cost-effective infrastructure can be designed around the mission, scaled with demand, secured as needed and operated as a dependable source of intelligence. The AI factory is designed to deliver the capacity, performance, utilization and business value customers expect. From AI workload to business outcome An AI factory is an integrated solution for data ingestion, model development, training, fine-tuning, inference, monitoring and continuous improvement. Its purpose is to turn data into intelligence repeatedly and efficiently—whether that intelligence supports a clinician, an engineer, a researcher, a public service or an enterprise application. Achieving that goal requires more than choosing an accelerated processor. Compute must match the business model and expected workloads. Networking and data pipelines must keep accelerators supplied. Storage must support the volume and velocity of data. Software must provision resources and orchestrate jobs. Security, governance and multi-tenancy must reflect who will use the environment and what data they can access. Power and cooling must support the system’s density today and as it grows. “Without proper architecture, governance and operational expertise, organizations can't safely leverage their data, take AI into core processes, or turn innovation into durable competitive advantage,” says Thierry Pienaar, HPE Fellow, Vice President and CTO for HPC and AI Sales. “Customers are realizing the fact that to derive value to its utmost extent they need an end-to-end infrastructure that's purpose-designed and purpose-built for AI.” Choice starts with the workload The HPE AI Factory with NVIDIA portfolio brings together NVIDIA accelerated computing, networking and AI software with HPE infrastructure, software, services and expertise at deploying complex systems. Rather than forcing every customer into a single configuration, the portfolio provides three paths for different ambitions, operating models and requirements. ● HPE Private Cloud AI, the turnkey AI factory solution, is an enterprise-ready, on-premises AI platform for running fine tuning, RAG & inferencing workload environments that need up to 256 GPUs. ● HPE AI Factory at-scale supports model builders, service providers and large enterprises that operate across many users, workloads and GPU resources (using 100s to 10s of thousands of GPUs) with centralized control, operational visibility, and multi-tenancy over the entire AI lifecycle ● HPE Sovereign AI Factory is an HPE AI factory at-scale that adds a deep level of operational control, data security and residency, sovereign management (including optional air-gapped configurations), and built-in compliance frameworks. It is designed for large enterprises, and any other organizations with sensitive information that require strict control and compliance across data, infrastructure, models and operations within defined legal, regulatory or geographic boundaries. Each option starts with the same principle: define the workloads and desired outcomes first, then select the right technologies to support them and finally identify the required resources needed to implement such an infrastructure. A hospital deploying clinical assistants will make different choices from a service provider offering GPU capacity, a manufacturer training vision models or a government operating sensitive national workloads. The HPE AI Factory model gives each a way to build for its mission without losing sight of performance, control or future growth – and HPE partners with these organizations to dramatically increase the likelihood of success. Operate the environment as one system As the use of AI expands across an organization, the operational challenges mature. Multiple teams may need different resource profiles, application stacks, service levels and data boundaries. Platform teams need to see utilization, allocate capacity, apply policy and understand consumption without creating a separate infrastructure island for every workload. The management of these differences make time-to-production an increasingly useful way to think about AI infrastructure: How quickly can an organization cost effectively move from investment vision to an operational environment generating useful intelligence? Many enterprises initially try to answer that question by extending their existing IT expertise. But building a DIY production AI environment from individual components requires skill sets that many enterprise IT organizations have never needed or required at this scale. A poorly implemented AI system may technically operate while still failing economically or operationally. GPUs can sit underutilized. Data pipelines can create bottlenecks. Cooling or power constraints can limit operation and/or expansion. Security policies can prevent sensitive data and workloads from being included. Separate less understood management systems can make AI factory infrastructure difficult to operate. That is why the AI factory challenge is fundamentally a strategic, systems integration and operations problem, not simply a stream of hardware purchasing transactions. "The HPE AI Factory with NVIDIA portfolio gives enterprises a range of AI solutions co-developed with NVIDIA, backed by HPE’s engineering expertise and technical capabilities to design an AI factory around their specific needs and optimize it for performance at scale." Pienaar explains. That distinction matters; customers can choose an architecture suited to their current mission and expand it as models, users and operational requirements change. The HPE AI Factory is designed to help operators provision and govern resources, observe infrastructure, track usage and support secure multi-tenant operations. That control helps customers align capacity with workload priorities while keeping the environment easier to manage as it grows. Make sovereignty a design requirement Cloud services, private environments and hybrid approaches can all play important roles in an AI strategy. For organizations with sovereignty requirements, the decision is defined by costs and the level of sovereignty and control they need: where data and models reside, who can administer the environment, which jurisdiction applies, data residency, how policies are enforced and what level of isolation various workloads require. “Sovereign AI tools from HPE and NVIDIA give an enterprise, or even a nation state, complete control over how its AI systems are built, deployed, operated and governed,” says Kaushik Shirhatti, Vice President, AI Factory at NVIDIA. “For some, that means keeping sensitive data in-country. For others, it means controlling who can access systems, where workloads run, how models are governed, and which local laws apply.” HPE and NVIDIA engineer for the complete outcome HPE and NVIDIA co-engineer AI factory solutions to reduce the integration work required to deploy and operate a high efficiency enterprise AI environment. By combining NVIDIA accelerated computing, networking, and AI software with HPE infrastructure, cloud operations, services, and support, the joint solution helps data scientists and developers spend more time building and improving AI applications while platform teams maintain operational production and control. NVIDIA provides accelerated computing platforms, networking, and the NVIDIA AI Enterprise software suite to power modern training, fine-tuning, inference, and agentic workloads. HPE contributes its expertise in enterprise systems engineering, high-performance computing, management and observability software, services, global support, financing, and years of experience in the power and cooling requirements of dense computing environments. Together, the companies can optimize AI computing solutions beyond any single component. The objective is to select the right GPU architecture and system design for the workload, keep accelerators productive with high-speed data movement, provide the software and operational controls teams need, and create a path to scale without unnecessarily redesigning the environment. HPE AI Services support that path from business planning, AI strategy, workload characterization and facility planning through deployment, integration, support and ongoing operations. HPE Financial Services can help with purchasing, accelerated depreciation schedules and lifecycle flexibility. These capabilities help customers make economically sound technology choices in the context of the business outcome, the operating model and the pace at which the environment needs to evolve. Deployment speed matters, but it is not the final measure of success. Customers need to consider workload readiness, model performance, accelerator utilization, developer productivity, governance, availability, economics and the ability to expand. Those measures connect the infrastructure decision to the outcomes the organization set out to achieve. The case for HPE AI Factory with NVIDIA is not that every customer needs the same stack. It is that every customer needs an AI environment intentionally matched to its workloads, data, operating requirements and goals. By combining NVIDIA’s accelerated computing leadership with HPE’s infrastructure, software, services and operating expertise, organizations can choose the right path—and move from AI investment to meaningful business outcomes faster. Recent deployments of the HPE AI Factory with NVIDIA - TELUS Sovereign AI Factory in Canada and the sovereign AI factory at the University of Utah in the US - are helping with overcoming engineering challenges and driving scientific advances. In conclusion, start by identifying the workloads that would benefit from AI, define the relevant data boundaries and residency requirements, estimate the expected scale over a reasonable timeframe, and determine the operating model, resources, and skills needed to support the AI infrastructure. Then work with HPE and NVIDIA to evaluate which path—turnkey, at-scale, or sovereign—best meets those requirements. To learn more visit HPE AI Factory | AI Infrastructure for Enterprises | HPE Sponsored by HPE and NVIDIA