AI News Archive: August 12, 2026 — Part 10
Sourced from 500+ daily AI sources, scored by relevance.
- Why Stream ring-maker Sandbar says the future of AI wearables is voice
AI notetaking hardware has taken off over the past couple of years, with credit-card-sized devices, pendants, pins, and even transcribing earbuds all promising to capture your meetings and turn them into summaries and action items. Now, a whole wave of wearables — rings especially — are betting people want to capture stray thoughts and ideas the same way. One of […]
Score: 30🌐 MovesAug 12, 2026https://techcrunch.com/video/why-stream-ring-maker-sandbar-says-the-future-of-ai-wearables-is-voice/ - Booksellers suspect AI firms are buying and then destroying rare books
AI firms quietly bulk buying rare books face resistance from booksellers.
- How SwishX Is Fixing Messy Back Office Sales Workflows For Pharma Companies With AI
When Swish Club raised $4.5 Mn in early 2024, it looked like yet another startup trying to solve enterprise device…
Score: 29🌐 MovesAug 12, 2026https://inc42.com/startups/how-swishx-is-fixing-messy-back-office-sales-workflows-for-pharma-companies-with-ai/ - Pivotal Advisors CEO Affirms Long-Term Confidence in AI Amid Growing Investor Scrutiny
Tiffany McGhee, CEO and CIO of Pivotal Advisors, expressed strong long-term confidence in the AI sector during a live financial news segment. While acknowledging the rapid evolution and transformative potential of AI technologies, she emphasized that the market is shifting from initial excitement and heavy spending toward a more critical evaluation of actual financial returns, including revenue and cash flow. She speaks with Romaine Bostick & Isabelle Lee on "The Close." (Source: Bloomberg)
Score: 28🌐 MovesAug 12, 2026https://www.bloomberg.com/news/videos/2026-08-12/pivotal-advisors-ceo-affirms-long-term-confidence-in-ai-video - Building Multimodal Workflows with a Local LLM
Image inputs and structured outputs with Gemma 4 and Ollama The post Building Multimodal Workflows with a Local LLM appeared first on Towards Data Science .
Score: 28🌐 MovesAug 12, 2026https://towardsdatascience.com/building-multimodal-workflows-with-a-local-llm/ - Virtual Tech Tools Fuel Higher Ed Robotics Research
The percentage of physical robots in industrial settings that are designed to work alongside humans has more than quadrupled since 2017, according to data from the nonprofit International Federation of Robotics. To investigate how collaborative robots can effectively communicate with people in manufacturing, medical and other settings, a number of colleges and universities have launched research efforts that involve virtual reality (VR), which may also help provide students with valuable workplace skills and experience. Vision-based training can be a powerful way to tap into the speed,…
Score: 28🌐 MovesAug 12, 2026https://edtechmagazine.com/higher/article/2026/08/virtual-tech-tools-fuel-higher-ed-robotics-research - The hard cases in speaker diarization: overlap, short turns, and noise
Analysis of challenging scenarios in speaker diarization.
- DB Group Investments Backs Blueye AI to Scale Advanced AI Solutions for Businesses and Enterprises
DB Group Investments Backs Blueye AI to Scale Advanced AI Solutions for Businesses and Enterprises USA Today
- Datamaxxing is the latest AI trend, and for once, it could be good for your health goals
Datamaxxing gives AI access to wearable health data so it can explain patterns your smartwatch leaves unexplained. Research suggests the idea has promise, but medical advice remains a dangerous line to cross.
- 5 custom Gemini Gems that I swear save me hours each week
5 custom Gemini Gems that I swear save me hours each week Tom's Guide
Score: 28🌐 MovesAug 12, 2026https://www.tomsguide.com/ai/google-gemini/5-custom-gemini-gems-that-i-swear-save-me-hours-each-week - GoodData.AI Cuts BI Migration From 18 Months to Weeks for DACH Enterprises, Opening the Fastest Route to Production AI
GoodData.AI Cuts BI Migration From 18 Months to Weeks for DACH Enterprises, Opening the Fastest Route to Production AI USA Today
- Law Firms Target AI Skills, First-Gen Connections in Summer Training Programs
Law firms are expanding pre-career training for law students, offering programs that range from early professional guidance for incoming 1Ls to courtroom-skills development for top 3Ls.
- The photographer who bet his business on the technology trying to replace him
For most photographers watching generative AI learn to render skin, fabric and light with unsettling accuracy, the instinct has been to defend the craft. Vincent Chow, however, did the opposite. The founder of Singapore-based product photography firm SnappyFly decided that if AI was coming for his industry, he would rather be the one driving it […] The post The photographer who bet his business on the technology trying to replace him appeared first on e27 .
Score: 28🌐 MovesAug 12, 2026https://e27.co/the-photographer-who-bet-his-business-on-the-technology-trying-to-replace-him-20260812/ - Agora voice agent with AssemblyAI Universal-3.5 Pro Realtime
Integrating Agora with AssemblyAI's Universal-3.5 Pro for voice agents.
Score: 28🌐 MovesAug 12, 2026https://assemblyai.com/blog/agora-voice-agent-with-assemblyai-universal-3-pro-streaming - Beyond algorithms: Why AI success is shaped by people, not platforms
Empowering employees to wield new technology effectively is fundamental for AI transformation in India. It's clear that technology alone isn't sufficient; rather, it's the people behind it who drive real change. Developing trust and awareness of AI's potential is essential for its successful integration. Organizations must focus on continuous learning and skill enhancement to adopt a human-centric approach, enabling AI to deliver meaningful decision-making and enduring benefits.
- Braneum launches Kara, a proactive AI chief-of-staff that works inside WhatsApp
Braneum has launched Kara, an AI chief-of-staff that helps businesses organize work, track commitments, and manage teams directly through WhatsApp. Kara went live on July 28, one day before Google announced Gemini Spark for India, as AI agents become the next focus area for global technology companies. While many AI assistants are built to answer […] The post Braneum launches Kara, a proactive AI chief-of-staff that works inside WhatsApp appeared first on CXOToday.com .
- Agentic Marketing Platform — Must-have Capabilities
Explores essential features for agentic AI marketing platforms to boost automation and personalization.
Score: 28🌐 MovesAug 12, 2026https://www.typeface.ai/blog/must-have-capabilities-of-an-agentic-ai-marketing-platform - Of course the ChatGPT dog cancer vaccine spawned a startup
The hyped AI-saved-my-dog story has reached its inevitable conclusion.
Score: 28🌐 MovesAug 12, 2026https://www.theverge.com/ai-artificial-intelligence/978671/ai-cured-dog-cancer-mrna-vaccine-startup-gamgee - The CMO's Guide to AI Marketing Platform Consolidation
Provides strategies for CMOs to streamline and consolidate AI marketing platforms for better ROI.
Score: 27🌐 MovesAug 12, 2026https://www.typeface.ai/blog/the-cmos-guide-to-ai-marketing-platform-consolidation - StoreEase and Sierra Partner to Bring Outcome-Driven AI to Self-Storage
StoreEase and Sierra Partner to Bring Outcome-Driven AI to Self-Storage azcentral.com and The Arizona Republic
- Najoom Al Thuraya unveils AI-powered fleetvision AI platform to transform telematics, road safety in UAE
Najoom Al Thuraya unveils AI-powered fleetvision AI platform to transform telematics, road safety in UAE Gulf News
- How AI-powered roleplay simulation is transforming sales training
By Gaurav Mishra, Founder and CEO, Proshort AI For decades, sales onboarding has followed a familiar pattern. New hires spend weeks studying product manuals, listening to recorded calls, and shadowing […] The post How AI-powered roleplay simulation is transforming sales training appeared first on Express Computer .
- How to Use LLMs for Powerful Automatic Evaluations
In this article, I discuss how you can perform automatic evaluations using LLM as a judge. LLMs are widely used today for a variety of… Continue reading on Towards AI »
- AI helping create popular mobile sports games
After the World Cup and before the domestic season, many fans still crave more football. Amateur coders are helping to fill this gap.
Score: 26🌐 MovesAug 12, 2026https://www.dw.com/en/ai-helping-create-popular-mobile-sports-games/a-78229226?maca=en-rss-en-all-1573-rdf - NEWSLETTER: American AI model makers smell an opportunity
NEWSLETTER: American AI model makers smell an opportunity reuters.com
- TCL CSOT Advances Into Professional Esports and Powers the Future of AI Computing at ChinaJoy 2026
TCL CSOT Advances Into Professional Esports and Powers the Future of AI Computing at ChinaJoy 2026 The Straits Times
- Opinion: Why Universities Need AI Accountability, Not Bans
A recent spike in fabricated citations, footnotes and sources in academic and professional reports is a sign that universities need to teach students to scrutinize AI outputs and be responsible for their use.
Score: 25🌐 MovesAug 12, 2026https://www.govtech.com/education/higher-ed/opinion-why-universities-need-ai-accountability-not-bans - AI-enabled stethoscope can miss the beat in pets
AI-enabled stethoscope can miss the beat in pets EurekAlert!
- Nvidia: Latest news and insights
More processor coverage on Network World: Intel news and insights | AMD news and insights With its legacy of innovation, Nvidia has become a dominant force in the AI market. Its GPUs have evolved from their gaming roots to power breakthroughs in scientific simulations, data analysis, and machine learning. Follow this page for the latest news, analysis, and features on Nvidia’s advancements and their impact on enterprise transformation. Nvidia’s $500B AI investment pool could impact enterprise chip pricing, availability Aug. 11, 2026 : Nvidia and six financial partners are creating a $500 billion investment pool to help Nvidia customers including frontier AI labs, AI clouds, and other enterprises buy its chips on credit . Nvidia moves to accelerate storage access, boost industry cooperation Aug. 4, 2026 : Nvidia is looking to speed access to AI storage and memory systems by open-sourcing its cuFile APIs and software stack. It’s also promoting a new 40-vendor initiative that aims to standardize GPU-driven storage advancements. Nvidia launches Open Secure AI Alliance with over 30 other AI companies July 27, 2026 : The Open Secure AI Alliance is new industry initiative to promote the creation of strong, safe, defensive AI cybersecurity tools built on open-source platforms. It’s an initiative of Nvidia with the backing of over 30 major AI makers and users, including Cisco, Databricks, Dell Technologies, HPE, IBM, Microsoft, OpenClaw, Palantir, Salesforce, SAP, ServiceNow, Siemens, Snowflake, and Thinking Machines. OpenAI is noticeably absent from the list of initial supporters. Nvidia unveils Spectrum-X networking platform designed to connect millions of GPUs July 22, 2026: Nvidia has introduced its next-generation Spectrum-X Ethernet networking platform , positioning it as a key building block for the next wave of “gigascale” AI factories designed to connect millions of GPUs while reducing power consumption and operational costs. Nvidia unveils Vera Rubin platform targeting AI, HPC infrastructure June 22, 2026 : Nvidia has formally launched the Vera Rubin platform , a combination CPU and GPU platform billed as a major step forward in the convergence of artificial intelligence and high-performance computing (HPC) for scientific research. Dell, Super Micro launch AI servers based on Nvidia Vera Rubin GPUs June 22, 2026 : Dell and Super Micro each unveiled new AI servers as part of Nvidia’s Vera Rubin rollout. The Dell PowerEdge XE8812, with its core Nvidia Vera Rubin NVL4 architecture, scales up to 144 GPUs per rack and will be at the heart of the Dell AI Factory with Nvidia preconfigured package of server, storage, networking, and software infrastructure. OpenAI weighs Nvidia-backed lease for 10 GW Ohio data center campus June 10, 2026 : OpenAI is reportedly in advanced talks to lease a proposed 10-gigawatt data center campus in southern Ohio in an arrangement that could include financial backing from Nvidia. The campus could cost at least $500 billion to build at current prices for chips, power, and construction. Startup Bolt Graphics promises 5x performance over Nvidia’s best GPU May 19, 2026 : It takes a brave company to go up against Nvidia in any market, let alone graphics performance. Intel tried and failed repeatedly, and AMD is barely hanging on. But Bolt Graphics thinks it has something in its Zeus GPU. Startup SPAN teams with Nvidia to put data center nodes in your backyard May 13, 2026 : A power-management company called SPAN has partnered with Nvidia and homebuilder PulteGroup to make use of spare electrical transmission capacity already available in many neighborhoods—something SPAN says its smart panels can detect. Rather than building massive new data centers with their own ZIP code, SPAN is proposing a network of small units, called XFRA nodes, installed outside of homes or in small commercial locations . These nodes are no bigger than an HVAC or power generator found outside of any home, according to SPAN. Nvidia’s ‘AI insurance policy’ balances immediate and future AI approaches April 27, 2026 : Cloud-hosted AI attracts extreme attention and investment. But what if the hype wave collapses? Nvidia’s challenge is to build interest in the boring aspects of future AI business cases when the market’s focus is elsewhere. Nvidia Rubin GPUs may be delayed, slowing the next phase of AI infrastructure April 9, 2026 : Nvidia’s latest generation of AI chips, the Nvidia Rubin GPUs, expected to ship later this year , may face supply delays amid ongoing geopolitical pressures and supply chain constraints. Nvidia’s SchedMD acquisition puts open-source AI scheduling under scrutiny April 7, 2026 : Nvidia’s recent acquisition of SchedMD, the company behind the Slurm workload manager, is raising concerns among AI industry executives and supercomputing specialists who fear the chip giant could use its new position to favor its own hardware over competing chips, whether through code prioritization or roadmap decisions. Nvidia overhauls the data center for OpenClaw era March 20, 2026 : Inference is the core data-center workload , and tokens are the new commodity of the AI era, says Nvidia. Recalling the classic data center during a keynote at GTC, CEO Jensen Huang said “it used to be … for files. It’s now a factory to generate tokens.” Nvidia joins push for data centers in space March 19, 2026 : Nvidia shared plans to bring AI and accelerated computing to space , joining a slew of other tech giants with out-of-this-world computing ideas. Nvidia CEO Huang talks up ‘tokenomics’ — the new currency for AI March 17, 2026: AI tokens are emerging as a kind of currency that will help in recruitment, budgeting and productivity, Nvidia’s CEO Jensen Huang said during a keynote address at the company’s GTC conference. Nvidia announces Vera Rubin platform, signaling a shift to full-stack AI infrastructure March 17, 2026 : Nvidia introduced its Vera Rubin platform , which combines compute, networking, and data processing into rack-scale deployments for large AI data centers, underscoring a shift in hyperscale environments toward more tightly integrated infrastructure. Why Nvidia’s DGX Rubin NVL8 runs on Intel Xeon 6 March 17, 2026 : Despite growing rivalry between the two chip makers, AI systems from Nvidia will use CPUs from Intel to maintain x86 continuity across data‑center workflows. Specifically, Nvidia has selected Intel’s Xeon 6 processors as the host CPUs for its Nvidia DGX Rubin NVL8 systems. Nvidia NemoClaw promises to run OpenClaw agents securely March 17, 2026 : In the few short weeks since OpenClaw became the biggest story in agentic AI, it has been dogged by concerns that it is not secure enough to be safely let loose in enterprises. This week at the Nvidia GTC conference, CEO Jensen Huang announced what he believes is the answer: NemoClaw . Built in consultation with OpenClaw’s creator, NemoClaw is based on Nvidia Agent Toolkit, part of the broader NeMo ecosystem for building AI agents. Nvidia targets inference as AI’s next battleground with Groq 3 LPX March 17, 2026 : Groq 3 LPX was announced at Nvidia GTC as part of an architecture comprising seven new chips and five racks meant to work together as “one big supercomputer.” The company says its new architecture marks a shift from training-focused infrastructure to systems optimized for continuous, low-latency enterprise AI workloads. Vendors tout Nvidia partnerships as Nvidia GTC 2026 kicks off HPE and Nvidia have boosted their partnership , adding a new server blade, GPU support, enhancements to HPE’s turnkey private AI package, and services targeting enterprise customers with growing AI workloads. Cisco extends its Secure AI Factory with Nvidia : The two vendors announced the expansion of their jointly developed Secure AI Factory with Nvidia, which melds Cisco security and networking technology, Nvidia DPUs and AI Enterprise software, and multivendor storage options. Lenovo bolsters hybrid AI platform with Nvidia GPUs : Lenovo is expanding its Lenovo Hybrid AI Advantage with Nvidia platform and positioning it as an end-to-end path for production AI inferencing. Palantir partners with Nvidia to streamline AI data center deployment : Two of the companies most synonymous with the AI revolution, Nvidia and Palantir Technologies, have linked arms to create an AI reference architecture operating system. Nvidia launches Nemotron 3 Super to power enterprise AI agents March 12, 2026 : Nvidia introduced a new reasoning-focused AI model that combines multiple neural network architectures in a bid to improve how enterprise systems handle complex tasks and automation. Nvidia partners with optics technology vendors Lumentum and Coherent to enhance AI infrastructure March 2, 2026 : Nvidia announced strategic partnerships with Lumentum Holdings and Coherent , which it said are designed to accelerate the development of advanced optics technologies used in AI data center infrastructure. Nvidia partners with telecom providers for open 6G networks March 2, 2026 : Nvidia has partnered with global telecom providers for a commitment to build 6G on open and secure AI-native platforms , bringing software-defined networking to telecommunications. Announced at the Mobile World Congress conference, the list of Nvidia partners is a who’s who of telecom. Nvidia plans a Windows PC SoC, setting up direct competition with Qualcomm, Intel, and AMD February 24, 2026 : Nvidia is developing a system-on-chip for Windows PCs , with Dell and Lenovo among the OEMs planning to build notebooks and desktops around the processor later this year. Nvidia lines up partners to boost security for industrial operations February 23, 2026 : Nvidia extended its collaborations with a handful of security vendors in an effort to improve real-time threat detection and response across operational technology environments and industrial control systems. Meta scoops up more of Nvidia’s AI chip output February 20, 2026 : AI’s insatiable demand for chips has already had an effect on the IT market, and it could be about to get worse: Nvidia has entered into a multi-year partnership with Meta to fill the social network’s new AI data centers with its cutting-edge processors. Nvidia claims 10x cost savings with open-source inference models February 13, 2026 : Nvidia has released analysis showing a 4X to 10X reduction in cost per token for AI inferencing by switching to open source models. The cost reductions were achieved by pairing Nvidia’s Blackwell GPUs with open-source models from Baseten, DeepInfra, Fireworks AI, and Together AI. Reports of Nvidia/OpenAI deal in jeopardy are overblown, says Nvidia’s CEO Huang February 5, 2026 : Nvidia CEO Jensen Huang told CNBC there was no validity to the rumors that he is reconsidering scaling back on his $100 billion investment in data centers for OpenAI , in sharp contrast to reports from the Wall Street Journal. Eying AI factories, Nvidia buys bigger stake in CoreWeave February 2, 2026 : Nvidia continues to throw its sizable bank account around . This time making a $2 billion investment in GPU cloud service provider CoreWeave. The company says the investment reflects Nvidia’s “confidence in CoreWeave’s business, team and growth strategy as a cloud platform built on Nvidia infrastructure. China clears Nvidia H200 sales to tech giants, reshaping AI data center plans January 29, 2026 : China has reopened access to Nvidia’s advanced AI processors , granting approvals to a small group of its largest technology companies. The companies include ByteDance, Alibaba, and Tencent, which are expected to collectively purchase more than 400,000 of Nvidia’s H200 accelerators Nvidia is still working with suppliers on RAM chips for Rubin January 26, 2026 : Nvidia changed its requirements for suppliers of the next generation of high-bandwidth memory, HBM4, but is close to certifying revised chips from Samsung Electronics for use in its AI systems, according to reports. RISC-V chip designer SiFive integrates Nvidia NVLink Fusion to power AI data centers January 19, 2026 : RISC-V pioneer SiFive has signed a deal with Nvidia to incorporate Nvidia NVLink Fusion into its data center products .The agreement means that SiFive will be able to connect its RISC-V CPUs to Nvidia GPUs and accelerators over a high bandwidth interconnect that lets multiple GPUs share compute and memory resources. Nvidia H200 chips in China: US says yes, China says no January 14, 2026: In what appears to be a case of diplomatic mind games in action, one day after the US government issued a regulation clearing the way for Nvidia to sell its H200 AI processors to Chinese companies on a case-by-case basis, a published report has revealed Chinese custom officers have been told not to let them into the country. Lenovo-Nvidia partnership targets faster AI infrastructure rollouts January 7, 2026 : Lenovo is pairing its liquid-cooled systems and networking with Nvidia platforms to deliver what it calls “AI cloud gigafactories ” designed to reduce deployment timelines from months to weeks. Top 10 Nvidia stories of 2025 – From the data center to the AI factory December 26, 2025 : For Nvidia, 2025 was not just about faster GPUs. It was about the fundamental re-architecture of the enterprise data center from a storage/retrieval hub to a manufacturing plant for intelligence – or the AI factory. Nvidia moves deeper into AI infrastructure with SchedMD acquisition December 16, 2025 : Nvidia has taken a strategic step deeper into the AI software stack with its acquisition of SchedMD, the developer of Slurm, a widely used open-source workload manager for high-performance computing and AI clusters. Nvidia bets on open infrastructure for the agentic AI era with Nemotron 3 December 15, 2025: AI agents must be able to cooperate, coordinate, and execute across large contexts and long time periods, and this, says Nvidia, demands a new type of infrastructure, one that is open. HPE loads up AI networking portfolio, strengthens Nvidia, AMD partnerships December 2, 2025 : HPE unveiled a wide range of networking gear and software to help enterprise customers move efficiently into the AI networking era. The rollout includes new switches and routers as well as deepened HPE’s partnerships with AMD and Nvidia . Nvidia’s $2B Synopsys stake tests independence of open AI interconnect standard December 2, 2025 : Nvidia announced a $2 billion investment in Synopsys , a chip design software maker, placing the GPU giant in the unusual position of holding substantial equity in a company that serves on the board of an industry consortium developing technology that competes with Nvidia’s own interconnect technology. Nvidia chips sold out? Cut back on AI plans, or look elsewhere November 20, 2025 : Nvidia CFO Colette Kress’s claim that “The clouds are sold out, and our GPU-installed base […] is fully utilized,” may have thrilled shareholders listening to the company’s earnings call on Wednesday, but it’s bad news for CIOs and data center managers who were counting on Nvidia for increased AI computing capacity . Nvidia’s first exascale system is the 4th fastest supercomputer in the world November 17, 2025: The world’s fourth exascale supercomputer has arrived, pitting Nvidia’s proprietary chip technologies against the x86 systems that have dominated supercomputing for decades. Nvidia touts next-gen quantum computing interconnects November 17, 2025 : At this week’s Supercomputing Conference, Nvidia highlighted how it expects to accelerate computing to quantum processors in the future. Nvidia highlights considerable science-based supercomputing efforts November 17, 2025 : Nvidia announced that more than 80 science-oriented systems powered by its platform have been deployed around the globe in the last year, at a combined total of 4,500 exaFLOPs of AI performance. Next-generation HPE supercomputer offers a mix of Nvidia and AMD silicon November 14, 2025: Hewlett Packard Enterprise said its next-generation Cray supercomputing platform will offer a choice of processors from Nvidia and AMD , even though the chips aren’t available yet and the system is likely not going to be available until 2027. Cisco, Nvidia strengthen AI ties with new data center switch, reference architectures October 29, 2025 : Cisco and Nvidia are ramping up their partnership . The latest deliverables include the new Cisco N9100 series data center switch built on Nvidia’s Spectrum-X Ethernet switch silicon. The two vendors also unveiled reference architectures to guide customer AI implementations of Cisco and Spectrum-X networks. Nvidia looks to power AI factory networks October 28, 2025 : Networking technology took center stage at Nvidia’s GTC DC developers show as a range of networking products were introduced for high-performance AI inference processing and more. The spotlight was on the launch of the ConnectX-9 SuperNIC . This next-generation network interface card delivers 1.6T GPU, with advanced RDMA capabilities and PCIe Gen 6 support. Quantum Circuits brings dual-rail qubits to Nvidia’s CUDA-Q development platform October 28, 2025 : Quantum Circuits announced that its dual-rail Seeker quantum processing unit now supports Nvidia’s CUDA-Q programming language, a move designed to help developers combine quantum computing with AI and machine learning workloads as the two technologies increasingly intersect. Should enterprise developers care about Nvidia? October 21, 2025 : If you’re a Java developer at a bank or a JavaScript developer at a retailer, you’ve probably spent your career blissfully ignoring hardware.That’s what the cloud is for. And Nvidia? That was just for gamers , crypto miners, or those PhDs in the AI lab playing with massive models. Except, it’s not anymore. Nvidia, Infineon partner for AI data center power overhaul October 16, 2025: Infineon is teaming up with Nvidia to upgrade the outdated power architecture of AI data centers and replace them with a centralized high-voltage DC power setup. Nvidia’s DGX Spark desktop supercomputer is on sale now, but hard to find O ctober 15, 2025 : Nvidia’s “personal AI supercomputer,” the DGX Spark , may run fast but it’s been slow getting here. It finally went on sale today, five months later than the company initially promised, and early units are hard to find: Inside Nvidia’s ‘grid-to-chip’ vision: How Vera Rubin and Spectrum-XGS advanceAI giga-factories October 13, 2025 : Nvidia will be front-and-center at this week’s Global Summit for members of the Open Compute Project. The company is making announcements on several fronts, including the debut of Vera Rubin MGX, its next-gen architecture fusing CPUs and GPUs , and Spectrum-XGS Ethernet, a networking fabric designed for “giga-scale” AI factories. Nvidia and Fujitsu team for vertical industry AI projects October 6, 2025 : Nvidia has partnered with Fujitsu to collaborate on vertical industry-specific artificial intelligence projects. The partnership will focus on co-developing and delivering an AI agent platform tailored for industry-specific agents in sectors such as healthcare, manufacturing, and robotics. Nvidia and OpenAI open $100B, 10 GW data center alliance September 23, 2025 : OpenAI and Nvidia will create a strategic partnership to deploy at least 10 gigawatts of Nvidia systems for OpenAI’s next-generation AI infrastructure.The first phase is expected to come online in the second half of 2026 using Nvidia’s Vera Rubin CPU/GPU combination platform to train and run new models. Who wins/loses with the Intel-Nvidia union? September 22, 2025: Nvidia is dipping into its $56 billion bank account to acquire a 5% stake in Intel for $5 billion, making it the second largest shareholder of Intel stock after the federal government’s recent investment. The deal provides Nvidia greater access to the x86 ecosystem, important for the enterprise data center market, and provides Intel with access to GPUs that have demand and can move their CPU products as well. Nvidia reportedly acquires Enfabrica CEO and chip technology license September 19, 2025 : Nvidia has hired away the CEO and other staff of chip interconnect maker Enfabrica, and licensed its core technologies in a deal worth over $900 million, Behind the move is demand for computing capacity to power generative AI for the likes of OpenAI, Anthropic, Mistral, AWS, Microsoft, and Google. Intel will design CPUs with Nvidia NVLink in return for $5 billion investment September 18, 2025 : Intel will collaborate with Nvidia to design CPUs with Nvidia’s NVLink high-speed chip interconnect. Nvidia and Intel also agreed to “jointly develop multiple generations of custom data center and PC products,” they said in a joint statement. China’s strike on Nvidia threatens global AI supply chains, sparking enterprise concerns September 16, 2025 : China has accused Nvidia of breaching its anti-monopoly law , a move that could disrupt the chipmaker’s global operations and heighten risks for enterprises dependent on its GPUs as US-China trade tensions escalate. Nvidia rolls out new GPUs for AI inferencing, large workloads September 9, 2025 : Nvidia has taken the wraps off a new purpose-built GPU along with a next-generation platform specifically targeted at massive-context processing as well as token software coding and generative video. Cadence adds Nvidia to digital twin tool for data center design September 9, 2025 : Cadence has updated to its Cadence Reality Digital Twin Platform library with the addition of digital twins for Nvidia’s DGX SuperPOD with DGX GB200 systems. Nvidia networking roadmap: Ethernet, InfiniBand, co-packaged optics will shape data center of the future September 4, 2025 : Nvidia’s networking roadmap is based on data centers evolution into a new unit of computing, from a focus on CPUs to GPUs as the primary computing units and from the distribution of functions across different components to support the infrastructure for AI workload Nvidia’s new computer gives AI brains to robots August 25, 2025: Nvidia CEO Jensen Huang sees a future where billions of robots serve humans, bringing in trillions of dollars in revenue for the company. To meet that goal, Nvidia on Monday unveiled a new computing device that will go into high-performing robots that could then try to replicate human behavior. Nvidia turns to software to speed up its data center networking hardware for AI August 22, 2025 : Nvidia wants to make long-haul GPU-to-GPU communication over Ethernet faster and more reliable, and hopes to achieve that with its new Spectrum-XGS algorithms, software protocols baked into Nvidia’s latest Ethernet gear. . Nvidia: ‘Graphics 3.0’ will drive physical AI productivity August 15, 2025 : Nvidia has floated the idea of “Graphics 3.0” with the hope of making AI-generated graphics central to physical productivity. The concept revolves around graphics created by genAI tools. Nvidia say AI-generated graphics could help in training robots to do their jobs in the physical world or by helping AI assistants automate the creation of equipment and structures. Nvidia launches Blackwell-powered RTX Pro GPUs for compact AI workstations August 12, 2025 : Nvidia announced two new professional GPUs, the RTX Pro 4000 Small Form Factor (SFF) and the RTX Pro 2000. Built on its Blackwell architecture, Nvidia’s new GPUs aim to deliver powerful AI capabilities in compact desktop and workstation deployments. Nvidia’s new genAI model helps robots think like humans August 11, 2025: Nvidia has developed a genAI model to help robots make human-like decisions by analyzing surrounding scenes. The Cosmos Reason model in robots can take in information from video and graphics input, analyze the data, and use its understanding to make decisions. Nvidia patches critical Triton server bugs that threaten AI model security August 5, 2025 : A surprising attack chain in Nvidia’s Triton Inference Server , starting with a seemingly minor memory-name leak, could allow full remote server takeover without user authentication. China demands ‘security evidence’ from Nvidia over H20 chip backdoor fears August 4, 2025 : China escalated pressure on Nvidia with the state-controlled People’s Daily publishing an opinion piece titled “Nvidia, how can I trust you?” — a day after regulators summoned company officials over alleged security vulnerabilities in H20 artificial intelligence chips. Nvidia to restart H20 exports to China, unveils new export-compliant GPU July 15, 2025: Nvidia will restart H20 AI chip sales to China and release a new GPU model compliant with export rules, a move that could impact global AI hardware strategies for enterprise IT teams. Nvidia has applied for US approval to resume sales and says that the government has indicated licenses will be granted and deliveries could begin soon. Nvidia GPUs are vulnerable to Rowhammer attacks July 15, 2025: Nvidia has issued a security reminder to application developers, computer manufacturers, and IT leaders that modern memory chips in graphic processors are potentially susceptible to so-called Rowhammer exploits after Canadian university researchers proved that an Nvidia A6000 GPU could be successfully compromised with a similar attack. Nvidia hits $4T market cap as AI, high-performance semiconductors hit stride July 11, 2025 : Nvidia became the first publicly traded company to surpass a $4 trillion market capitalization value, 13 months after surpassing the $3 trillion mark. This makes Nvidia the world’s most valuable company ahead of Apple and Microsoft. New Nvidia technology provides instant answers to encyclopedic-length questions Jul 8, 2025: Have a question that needs to process an encyclopedia-length dataset? Nvidia says its new technique can answer it instantly. Built leveraging the company’s Blackwell processor’s capabilities, the new “Helix Parallelism” method allows AI agents to process millions of words — think encyclopedia-length — and support up to 32x more users at a time. Nvidia doubles down on GPUs as a service July 8, 2025: Nvidia’s recent initiative to dive deeper into the GPU-as-a-service (GPUaaS) model marks a significant and strategic shift that reflects an evolving landscape within the cloud computing market. Nvidia, Perplexity to partner with EU and Middle East AI firms to build sovereign LLMs June 12, 2025: Nvidia and AI search firm Perplexity said they are joining hands with model builders and cloud providers across Europe and the Middle East to refine sovereign large-language models (LLMs) and accelerate enterprise AI uptake in local industries. Nvidia: ‘Sovereign AI’ will change digital work June 11, 2025: Nvidia executives think sovereign AI has the potential to change digital work as generative AI (genAI) aligns with national priorities and local regulations. AWS cuts prices of some EC2 Nvidia GPU-accelerated instances June 9, 2025 : AWS has reduced the prices of some of its Nvidia GPU-accelerated instances to attract more AI workloads while competing with rivals, such as Microsoft and Google, as demand for GPUs and the cost of securing them continues to grow. Nvidia aims to bring AI to wireless June 6, 2025 : Nvidia hopes to maximize RAN infrastructure use (traditional networks average a low 30% to 35%), use AI to rewrite the air interface, and enhance performance and efficiency through radio signal processing. The longer-term goal is to seamlessly process AI traffic at the network edge to create new monetization opportunities for service providers. Oracle to spend $40B on Nvidia chips for OpenAI data center in Texas May 26, 2025: Oracle is reportedly spending about $40 billion on Nvidia’s high-performance computer chips to power OpenAI’s new data center in Texas, marking a pivotal shift in the AI infrastructure landscape that has significant implications for enterprise IT strategies. Nvidia eyes China rebound with stripped-down AI chip tailored to export limits May 26, 2025: Nvidia plans to launch a lower-cost AI chip for China in June, aiming to protect market share under the US export controls and signal a broader shift toward affordable, segmented products that could impact global enterprise AI spending. Nvidia introduces ‘ridesharing for AI’ with DGX Cloud Lepton May 19, 2025 : Nvidia introduced DGX Cloud Lepton , an AI-centric cloud software program that makes it easier for AI factories to rent out their hardware to developers who wish to access performant compute globally. Nvidia opens NVLink to competitive processors May 19, 2025: Nvidia kicked off the Computex systems hardware tradeshow with the news it has opened the NVLink interconnect technology to the competition with the introduction of NVLink Fusion. NVLink is a high-speed interconnect born out of its Mellanox networking group which lets multiple GPUs in a system or rack share compute and memory resources, thus making many GPUs appear to the system as a single processor. AMD, Nvidia partner with Saudi startup to build multi-billion dollar AI service centers May 15, 2025 : As part of the avalanche of business deals coming from President Trump’s Middle East tour, both AMD and Nvidia have struck multi-billion dollar deals with an emerging Saudi AI firm. The deals served as the coming out party for Humain , a state-backed artificial intelligence (AI) company that operates under the Kingdom’s Public Investment Fund (PIF) and is chaired by Crown Prince Mohammed bin Salman. Nvidia, ServiceNow engineer open-source model to create AI agents May 6, 2025 : Nvidia and ServiceNow have created an AI model that can help companies create learning AI agents to automate corporate workloads..The open-source Apriel model, available generally in the second quarter on HuggingFace, will help create AI agents that can make decisions around IT, human resources and customer-service functions. Nvidia AI supercluster targets agents, reasoning models on Oracle Cloud April 29, 2025 : The move marks the first wave of liquid-cooled Nvidia GB200 NVL72 racks in OCI data centers , involving thousands of Nvidia Grace CPUs and Blackwell GPUs. Nvidia says NeMo microservices now generally available April 23, 2025 : Nvidia announced the general availability of neural module (NeMo) microservices , a modular platform for building and customizing gen AI models and AI agents.NeMo microservices integrate with partner platforms to provide features including prompt tuning, supervised fine-tuning, and knowledge retrieval tools. Nvidia expects ban on chip exports to China to cost $5.5B April 16, 2025: Nvidia now expects new US government restrictions on exports of its H20 chip to China will cost the company as much as $5.5 billion . Incomplete patching leaves Nvidia, Docker exposed to DOS attacks April 15, 2025: A critical race condition bug affecting the Nvidia Container Toolkit, which received a fix in September, might still be open to attacks owing to incomplete patching. Nvidia lays out plans to build AI supercomputers in the US April 14, 2025 : There was mixed reaction from industry analysts over an announcement that Nvidia plans to produce AI supercomputers entirely in the US. The company said in a blog post that, together with its manufacturing partners, it has commissioned more than one million square feet (92,900 square meters) of manufacturing space to build and test Nvidia Blackwell chips in Arizona and AI supercomputers in Texas. Potential Nvidia chip shortage looms as Chinese customers rush to beat US sales ban April 2, 2025: The AI chip shortage could become even more dire as Chinese customers are purportedly looking to hoard Nvidia chips ahead of a proposed US sales ban. According to inside sources, Chinese companies including ByteDance, Alibaba Group, and Tencent Holdings have ordered at least $16 billion worth of Nvidia’s H20 server chips for running AI workloads in just the first three months of this year. Nvidia’s Blackwell raises the bar with new MLPerf Inference V5.0 results April 2, 2025: Nvidia released a set of MLPerf Inference V5.0 benchmark results for its Blackwell GPU , the successor to Hopper, saying that its GB200 NVL72 system, a rack-scale offering designed for AI reasoning, set a series of performance records. 5 big takeaways from Nvidia GTC March 25, 2025: Now that the dust has settled from Nvidia’s GTC 2025 , a few industry experts weighed in on some core big picture developments from the conference. Here are five of their top observations. Nvidia wants to be a one-stop enterprise technology shop March 24, 2025 : After last week’s Nvidia GTC 2025 event , a new, fuller picture of the vendor emerged. Analysts agree that Nvidia is not just a graphics chip provider anymore. It’s a full-stack solution provider, and GPUs are just one of many parts. Nvidia launches AgentIQ toolkit to connect disparate AI agents March 21, 2025 : As enterprises look to adopt agentic AI to boost the efficiency of their applications, Nvidia introduced a new open-source software library — AgentIQ toolkit — to help developers connect disparate agents and agent frameworks. The toolkit, according to Nvidia, packs in a variety of tools, including ones to weave in RAG, search, and conversational UI into agentic AI applications. Nvidia launches research center to accelerate quantum computing breakthrough March 21, 2025 : In a move to help accelerate the timeline for practical, real-world quantum applications, Nvidia is establishing the Nvidia Accelerated Quantum Research Center . “Quantum computing will augment AI supercomputers to tackle some of the world’s most important problems,” Nvidia CEO Jensen Huang said. Nvidia, xAI and two energy giants join genAI infrastructure initiative March 19, 2025: An industry generative artificial intelligence (genAI) alliance, the AI Infrastructure Partnership (AIP), on Wednesday announced that xAI, Nvidia, GE Vernova, and NextEra Energy were joining BlackRock, Microsoft, and Global Infrastructure Partners as members. IBM broadens access to Nvidia technology for enterprise AI March 19, 2025: New collaborations between IBM and Nvidia have yielded a content-aware storage capability for IBM’s hybrid cloud infrastructure, expanded integration between watsonx and Nvidia NIM, and AI services from IBM Consulting that use Nvidia Blueprints. Nvidia’s silicon photonics switches bring better power efficiency to AI data centers March 19, 2025: Amid the flood of news from Nvidia’s annual GTC event, one item stood out. Nvidia introduced new silicon photonics network switches that integrate network optics into the switch using a technique called co-packaged optics (CPO), replacing traditional external pluggable transceivers. While Nvidia alluded to its new switches providing a cost savings, the primary benefit is to reduce power consumption with an improvement in network resiliency. What is Nvidia Dynamo and why it matters to enterprises? March 19, 2025: Chipmaker Nvidia has released a new open-source inferencing software — Dynamo, at its GTC 2025 conference, that will allow enterprises to increase throughput and reduce cost while using large language models on Nvidia GPUs. Nvidia, xAI and two energy giants join genAI infrastructure initiative March 19, 2025: AI Infrastructure Partnership (AIP) announced that xAI, Nvidia, GE Vernova, and NextEra Energy joined the AIP . But given that no financial commitments or any other details were released, will it make a difference? HPE, Nvidia broaden AI infrastructure lineup March 19, 2025: HPE news from Nvidia GTC includes a new Private Cloud AI developer kit, Nvidia AI blueprints, GPU optimization capabilities, and servers built with Nvidia Blackwell Ultra and Blackwell architecture. Cisco, Nvidia team to deliver secure AI factory infrastructure March 18, 2025: Cisco and Nvidia have expanded their partnership to create their most advanced AI architecture package to date, designed to promote secure enterprise AI networking . Nvidia’s ‘hard pivot’ to AI reasoning bolsters Llama models for agentic AI March 18, 2025: The company has post-trained its new Llama Nemotron family of reasoning models to improve multistep math, coding, reasoning, and complex decision-making. The enhancements aim to provide developers and enterprises with a business-ready foundation for creating AI agents that can work independently or as part of connected teams. Nvidia details its GPU, CPU, and system roadmap for the next three years March 18, 2025: Nvidia CEO Jensen Huang shared previously unreleased specifications for its Rubin graphics processing unit (GPU), due in 2026, the Rubin Ultra coming in 2027, and announced the addition of a new GPU called Feynman to the mix for 2028. Oracle, Nvidia partner to add AI software into OCI services March 18, 2025: Nvidia’s AI Enterprise stack will be available natively through the OCI Console and will be available anywhere in OCI’s distributed cloud while providing enterprises access to over 160 AI tools for training and inference, including NIM microservices, the companies said in a joint statement at Nvidia’s annual GTC conference. Nvidia GTC 2025: What to expect from the AI leader March 3, 2025 : Last year, Nvidia’s GTC 2024 grabbed headlines with the introduction of the Blackwell architecture and the DGX systems powered by it. With Nvidia GTC 2025 right around the corner, the tech world is eager to see what Nvidia – and its partners and competitors – will unveil next. Cisco, Nvidia expand AI partnership to include Silicon One technology February 25, 2025 ; Cisco and Nvidia have expanded their collaboration to support enterprise AI implementations by tying Cisco’s Silicon One technology to Nvidia’s Ethernet networking platform. The extended agreement is designed to offer customers yet another way to support AI workloads across the data center and strengthens both companies’ strategies to expand the role of Ethernet networking for AI in the enterprise. Nvidia forges healthcare partnerships to advance AI-driven genomics, drug discovery February 14, 2025 : Through new partnerships with industry leaders, Nvidia aims to advance practical use cases for AI in healthcare and life sciences. It’s a logical move: Healthcare has the most significant upside, particularly in patient care, among all the industries applicable to AI. Nvidia partners with cybersecurity vendors for real-time monitoring February 12, 2025 : Nvidia partnered with leading cybersecurity firms to provide real-time security protection using its accelerator and networking hardware in combination with its AI software. Under the agreement, Nvidia will provide integration of its BlueField and Morpheus hardware with cyber defenses software from Armis, Check Point Software Technologies, CrowdStrike, Deloitte and World Wide Technology . Nvidia claims near 50% boost in AI storage speed February 7, 2025: Nvidia is touting a near 50% improvement in storage read bandwidth thanks to intelligence in its Spectrum-X Ethernet networking equipment, according to the vendor’s technical blog post. Spectrum-X is a combination of the company’s Spectrum-4 Ethernet switch and BlueField-3 SuperNIC smart networking card, which supports RoCE v2 for remote direct memory access (RDMA) over Converged Ethernet. Nvidia unveils preview of DeepSeek-R1 NIM microservice February 3, 2025 : The chipmaker stock plummeted 17% after Chinese AI developer DeepSeek unveiled its DeepSeek-R1 LLM. Last week, Nvidia announced the DeepSeek-R1 model is now available as a preview Nvidia inference microservice (NIM) on build.nvidia.com. Nvidia unveils preview of DeepSeek-R1 NIM microservice January 31, 2025: Nvidia stock plummeted 17% after Chinese AI developer, DeepSeek, unveiled its DeepSeek-R1 LLM. Later the same week, the chipmaker turned around and announced the DeepSeek-R1 model is available as a preview Nvidia inference microservice (NIM) on build.nvidia.com. Nvidia intros new guardrail microservices for agentic AI January 16, 2025 : Nvidia added new Nvidia inference microservices (NIMs) for AI guardrails to its Nvidia NeMo Guardrails software tools. The new microservices aim to help enterprises improve accuracy, security, and control of agentic AI applications, addressing a key reservation IT leaders have about adopting the technology. Nvidia year in review January 10, 202 5: Last year was Nvidia’s year. Its command of mindshare and market share was unequaled among tech vendors. Here’s a recap of some of the key Nvidia events of 2024 that highlight just how powerful the world’s most dominant chip player is. Nvidia launches blueprints to help jumpstart AI projects January 8, 202 5: Nvidia recently issued designs for AI factories after hyping up the idea for several months. Now it has come out with AI blueprints , essentially prebuilt templates that give developers a jump start on creating AI systems. Nvidia’s Project DIGITS puts AI supercomputing chips on the desktop January 6, 2025 : Nvidia is readying a tiny desktop device called Project DIGITS , a “personal AI supercomputer” with a lightweight version of the Grace Blackwell platform found in its most powerful servers; it’s aimed at data scientists, researchers, and students who will be able to prototype, tune, and run large genAI models. Nvidia unveils generative physical AI platform, agentic AI advances at CES January 6, 202 5: At CES in Las Vegas, Nvidia trumpeted a slew of AI announcements , with an emphasis on generative physical AI that promises a new revolution in factory and warehouse automation. “AI requires us to build an entirely new computing stack to build AI factories, accelerated computing at data center scale,” Rev Lebaredian, vice president of omniverse and simulation technology at Nvidia. Verizon, Nvidia team up for enterprise AI networking December 30, 2024 : Verizon and Nvidia partnered to build AI services for enterprises that run workloads over Verizon’s 5G private network. The new offering, 5G Private Network with Enterprise AI, will run a range of AI applications and workloads over Verizon’s private 5G network with Mobile Edge Compute (MEC). MEC is a colocated infrastructure that is a part of Verizon’s public wireless network, bringing compute and storage closer to devices and endpoints for ultra-low latency. Nvidia’s Run:ai acquisition waved through by EU December 20, 2024: Nvidia will face no objections to its plan to acquire Israeli AI orchestration software vendor Run:ai Labs in Europe, after the European Commission gave the deal its approval today. But Nvidia may not be out of the woods yet. Competition authorities in other markets are closely examining the company’s acquisition strategy. China launches anti-monopoly probe into Nvidia amid rising US-China chip tensions December 10, 2024 : China has initiated an investigation into Nvidia over alleged violations of the country’s anti-monopoly laws, signaling a potential escalation in the ongoing tech and trade tensions between Beijing and Washington. Nvidia Blackwell chips face serious heating issues November 18, 2024 : Nvidia’s next-generation Blackwell data center processors have significant problems with overheating when installed in high-capacity server racks, forcing redesigns of the racks themselves, according to a report by The Information. These issues have reportedly led to design changes, meaning delays in shipping product and raising concern that its biggest customers, including Google, Meta, and Microsoft, will be able to deploy Blackwell servers according to their schedules. Nvidia to power India’s AI factories with tens of thousands of AI chips October 24, 2024 : Nvidia plans to deploy thousands of Hopper GPUs in India to create AI factories and collaborate with Reliance Industries to develop AI infrastructure.. Yotta Data Services, Tata Communications, E2E Networks, and Netweb will lead the AI factories — large-scale data centers for producing AI. Nvidia added that the expansion will provide nearly 180 exaflops of computing power. Nvidia contributes Blackwell rack design to Open Compute Project October 15, 2024 : Nvidia contributed to the Open Compute Project its Blackwell GB200 NVL72 electro-mechanical designs – including the rack architecture, compute and switch tray mechanicals, liquid cooling and thermal environment specifications, and Nvidia NVLink cable cartridge volumetrics –. As global AI energy usage mounts, Nvidia claims efficiency gains of up to 100,000X October 08, 2024 : As concerns over AI energy consumption ratchet up, chip maker Nvidia is defending what it calls a steadfast commitment to sustainability . The company reports that its GPUs have experienced a 2,000X reduction in energy use over the last 10 years in training and a 100,000X energy reduction over that same time in generating tokens. Accenture forms new Nvidia business group focused on agentic AI adoption October 4, 2024 : Accenture and Nvidia announced an expanded partnership focused on helping customers rapidly scale AI adoption. Accenture said the new group will use Accenture’s AI Refinery platform — built on the Nvidia AI stack, including Nvidia AI Foundry, Nvidia AI Enterprise, and Nvidia Omniverse — to help clients create a foundation for use of agentic AI. IBM expands Nvidia GPU options for cloud customers October 1, 2024 : IBM expanded access to Nvidia GPUs on IBM Cloud to help enterprise customers advance their AI implementations, including large language model (LLM) training. IBM Cloud users can now access Nvidia H100 Tensor Core GPU instances in virtual private cloud and managed Red Hat OpenShift environments. Oracle to offer 131,072 Nvidia Blackwell GPUs via its cloud September 12, 2024 : Oracle started taking pre-orders for 131,072 Nvidia Blackwell GPUs in the cloud via its Oracle Cloud Infrastructure (OCI) Supercluster to aid large language model (LLM) training and other use cases, the company announced at the CloudWorld 2024 conference. The launch of an offering that provides these many Blackwell GPUs, also known as Grace Blackwell (GB) 200, is significant as enterprises globally are faced with the unavailability of high-bandwidth memory (HBM) — a key component used in making GPUs. Why is the DOJ investigating Nvidia? September 11, 2024 : After a stock sell-off following its quarterly earnings report, Nvidia’s pain was aggravated by news that the Department of Justice is escalating its investigation into the company for anticompetitive practices. According to a Bloomberg report, the DOJ sent a subpoena to Nvidia as part of a probe into alleged antitrust practices . Cisco, HPE, Dell announce support for Nvidia’s pretrained AI workflows September 4, 2024 : Cisco, HPE, and Dell are using Nvidia’s new AI microservices blueprints to help enterprises streamline the deployment of generative AI applications. Nvidia’s announced its NIM Agent Blueprints, a catalogue of pretrained, customizable AI workflows that are designed to provide a jump-start for developers creating AI applications. NIM Agent Blueprints target a number of use cases, including customer service, virtual screening for computer-aided drug discovery, and a multimodal PDF data extraction workflow for retrieval-augmented generation (RAG) that can ingest vast quantities of data. Nvidia reportedly trained AI models on YouTube data August 4, 2024 : Nvidia scraped huge amounts of data from YouTube to train its AI models, even though neither Youtube nor individual YouTube channels approved the move, according to leaked documents. Among other things, Nvidia reportedly used the YouTube data to train its deep learning model Cosmos, an algorithm for automated driving, a human-like AI avatar, and Omniverse, a tool for building 3D worlds. Can Intel’s new chips compete with Nvidia in the AI universe? June 9, 2024 : Intel is aiming its next-generation X86 processors at AI tasks , even though the chips won’t actually run AI workloads themselves.mAt Computex, Intel announced its Xeon 6 processor line, talking up what it calls Efficient-cores (E-cores) that it said will deliver up to 4.2 times the performance of Xeon 5 processors. The first Xeon 6 CPU is the Sierra Forest version (6700 series) a more performance-oriented line, Granite Rapids with Performance cores (P-cores or 6900 series), will be released next quarter. Everyone but Nvidia joins forces for new AI interconnect May 30, 2024 : A clear sign of Nvidia’s dominance is when Intel and AMD link arms to deliver a competing product. That’s what happened when AMD and Intel – along with Broadcom, Cisco, Google, Hewlett Packard Enterprise, Meta and Microsoft – formed the Ultra Accelerator Link (UALink) Promoter Group to develop high-speed interconnections between AI processors. Nvidia to build supercomputer for federal AI research May 15, 2024 : The U.S. government will use an Nvidia DGX SuperPOD to provide researchers and developers access to much more computing power than they have had in the past to produce generative AI advances in areas such as climate science, healthcare and cybersecurity. Nvidia, Google Cloud team to boost AI startups April 11, 2024 : Alphabet’s Google Cloud unveiled a slew of new products and services at Google Cloud Next 2024, among them a program to help startups and small businesses build generative AI applications and services. The initiative brings together the Nvidia Inception program for startups and the Google for Startups Cloud Program. Nvidia GTC 2024 wrap-up: Blackwell not the only big news March 29, 2024 : Nvidia’s GDC is in our rearview mirror , and there was plenty of news beyond the major announcement of the Blackwell architecture and the massive new DGX systems powered by it. Here’s a rundown of some of the announcements you might have missed. Nvidia expands partnership with hyperscalers to boost AI training and development March 19, 2024 : Nvidia extended its existing partnerships with hyperscalers Amazon Web Services (AWS), Google Cloud Platform, Microsoft Azure, and Oracle Cloud Infrastructure, to make available its latest GPUs and foundational large language models and to integrate its software across their platforms. Nvidia launches Blackwell GPU architecture March 18, 2024 : Nvidia kicked off its GTC 2024 conference with the formal launch of Blackwell, its next-generation GPU architecture due at the end of the year. Blackwell uses a chiplet design, to a point. Whereas AMD’s designs have several chiplets, Blackwell has two very large dies that are tied together as one GPU with a high-speed interlink that operates at 10 terabytes per second, according to Ian Buck, vice president of HPC at Nvidia. Cisco, Nvidia target secure AI with expanded partnership February 9, 2024 : Cisco and Nvidia expanded their partnership to offer integrated software and networking hardware that promises to help customers more easily spin up infrastructure to support AI applications. The agreement deepens both companies’ strategy to expand the role of Ethernet networking for AI workloads in the enterprise. It also gives both companies access to each other’s sales and support systems. Nvidia and Equinix partner for AI data center infrastructure J anuary 9, 2024: Nvidia partnered with data center giant Equinix to offer what the vendors are calling Equinix Private AI with Nvidia DGX, a turnkey solution for companies that are looking to get into the generative AI game but lack the data center infrastructure and expertise to do it.
Score: 25🌐 MovesAug 12, 2026https://www.networkworld.com/article/3562856/nvidia-latest-news-and-insights.html - AICC Platform Demonstrates Businesses Can Deploy AI Customer Service Agents for Under $50 Per Month
AICC Platform Demonstrates Businesses Can Deploy AI Customer Service Agents for Under $50 Per Month azcentral.com and The Arizona Republic
- Ray-Ban Meta AI glasses are scaring me and my homegirls | Opinion
Ray-Ban Meta AI glasses are scaring me and my homegirls | Opinion USA Today
- Smart biometric IDs remove login friction
Smart biometric IDs remove login friction Healthcare IT News
Score: 25🌐 MovesAug 12, 2026https://www.healthcareitnews.com/video/emea/smart-biometric-ids-remove-login-friction - The New Rules of Digital Sovereignty: Architecture, Control, and Competitive Advantage
The New Rules of Digital Sovereignty: Architecture, Control, and Competitive Advantage IT Pro
- The AI text conundrum
Plus: AbCellera cashes in on menopause drug breakthrough. The post The AI text conundrum first appeared on BetaKit .
- Revise Launches AI Document Editor With Reviewable, Tracked Changes
Revise Launches AI Document Editor With Reviewable, Tracked Changes azcentral.com and The Arizona Republic
- The AI boom minted a new class of billionaire superyacht buyers. Their deals happen behind closed doors.
The AI boom minted a new class of billionaire superyacht buyers. Their deals happen behind closed doors. Business Insider
Score: 24🌐 MovesAug 12, 2026https://www.businessinsider.com/superyacht-boom-off-market-deals-ai-wealth-2026-8 - Raw WebSocket voice agent with AssemblyAI (Python)
Building a raw WebSocket voice agent using AssemblyAI in Python.
Score: 24🌐 MovesAug 12, 2026https://assemblyai.com/blog/raw-websocket-voice-agent-with-assemblyai-universal-3-pro-streaming - How to measure speaker diarization accuracy (cpWER) in Python
Tutorial on measuring diarization accuracy using cpWER in Python.
Score: 24🌐 MovesAug 12, 2026https://assemblyai.com/blog/measure-speaker-diarization-accuracy-cpwer-python - Measuring Spurious Correlations with Feature Strength
This work was partially done by an automated research scaffold developed at Redwood Research. For this project, all of the experiment ideas were designed by a human and a human wrote the write up. The AI mostly just executed on the experiment ideas. We think this project is slightly below the level of rigor of a mid-MATS research update, and the research scaffold was not very helpful for this project. More discussion of AI usage is in the Appendix. 💻 Codebase If we want to train a classifier that distinguishes whether a passage is code or prose, we can do so by gathering samples of code and of prose, and training the classifier to distinguish between the two classes. Unfortunately, this might not work if the data hides a spurious correlation. If all the code is in Spanish and all of the prose is in English, then the classifier might learn to predict Spanish vs. English instead of code vs. prose. We find that this happens in practice: when we fine-tune an LLM to classify between Spanish code and English prose and evaluate on Spanish prose or English code, it generalizes to predicting the language rather than the domain. This suggests that language is in some sense a stronger feature than code. We think that spurious correlations are an important threat model for a few reasons. First, classifiers might actually be trained in ways that unintentionally contain spurious correlations. For example: Sycophancy vs validation-seeking user. Suppose we want to classify examples of the model being sycophantic to the user, but for all the on-policy examples we could find, the model is sycophantic exactly when it realizes the user it is talking to is seeking validation because that's what the reward model picked up on during RLHF. Then the classifier might learn to predict whether a user seeks validation instead of whether the model is being sycophantic, and we would miss rare instances of sycophancy for other kinds of users in deployment. Sandbagging vs off-policy pipeline. Suppose we want to classify examples of a model sandbagging by training a probe, but have no on-policy examples of the model sandbagging, so we use some complicated off-policy pipeline to generate sandbagging examples. The classifier might learn whether an example comes from the off-policy pipeline and fail to pick up on when the model sandbags on-policy. Policies can also pick up on problematic spurious correlations which can lead to goal misgeneralization. For example: Prompting to hack vs cybersecurity challenge. If a model is instructed and trained to hack during cybersecurity challenges, but the hacking instruction only ever appears during cybersecurity challenges, then the model might learn to hack whenever something looks like a cybersecurity challenge (say, the code looks hackable) instead of only when the prompt tells it to. Training classifiers on data with spurious correlations is a first step toward understanding spurious correlations in general. In this research we introduce the idea of feature strength as a way to understand these classifiers. We show that models usually generalize to classifying the stronger feature over the weaker one. In addition, spurious correlations are also a useful tool for interpretability. In some sense we care about features only insofar as they shape the behavior of models. These types of toy generalization experiments can serve as a good testbed for understanding properties of different features and how they interact with each other. Key Contributions: We introduce the idea of feature strength , a scalar value for each feature that predicts what a classifier learns on out-of-distribution data when the training data contains spurious correlations. Feature strength is specific to each classification method. Feature strength is highly correlated with mean activation difference for probe classifiers, but much less so for fine-tuned classifiers, and the probe and fine-tuning strengths are only weakly correlated with each other. Explicit prompting works for fine-tuning but not probing: in the example above, if we say "distinguish between code and prose," the model picks up on code vs prose when fine-tuned but not when probed. Fine-tuning on the diagonal and antidiagonal usually gives similar results, but when features are correlated there are notable exceptions. The setup Take two binary text features, and , with classes and . In the example above: Feature is language: is Spanish, is English. Feature is code: is code, is prose. First we build training data on the diagonal of their 2 × 2 grid: class texts are × (Spanish code) and class texts are × (English prose). Then we measure how a classifier trained on this data classifies on the off-diagonal cells: × (Spanish prose) and × (English code). To measure which feature the classifier follows, we ask the model to pick between class and class with this prompt: {passage} The passage above belongs to one of two categories, X or Y. Answer with a single letter. Answer: We measure this three ways: A full fine-tune of the model on the prompt above, with the loss on the label token (" X" or " Y") after "Answer:". A mean-pooled probe : a linear probe on the base model's activations, averaged over all tokens in the passage. This probe uses just the passage text, instead of the full prompt above. A decision-token probe : a linear probe on the base model's activations at the final token of the prompt above, where the model is about to answer or . We study 37 different features, including language, high-entropy strings, and lying. The full list and descriptions of the features are in the appendix. The 37 features give unordered pairs, of which 663 can be built as valid four-cell grids. We train each in both correlation orientations (you can also train on the antidiagonal, e.g. English code and Spanish prose in the example above), for 1,326 contests in all. We do all our main-body research on Qwen3-1.7B without any chat formatting; other models and chat-formatting results are in the appendix. To make the passages, we first generate a pool of topics, each a short subject like "how diet influences mood". Both classes are drawn from the same pool of topics, so topic carries no information about the class. Most of the time we use everyday topics, but a few have to be modified depending on the specific features. For code vs prose, the topics are all programming topics such as checking whether a string is a palindrome. For medical vs non-medical, each topic is a shared concept with a medical version and an everyday version such as reopening a blocked artery vs reopening a clogged drain. For each topic, we prompt Claude Haiku 4.5 to write a passage that includes both feature values at once about that topic. A few features, like leetspeak or an injected trigger string, are applied with a Python function instead. Claude Sonnet 4.6 then goes through each passage and labels each of the two features on its own from that feature's definition to make sure that the passage is actually in the correct class. Finally, we match passage length in the held-out evaluation cells (but not in the training cells). Measuring feature strength For any classification method , define the feature difference as the arithmetic mean of the log-odds towards over in the two held-out cells. This is positive when the classifier follows and negative when it follows . For the example above: Spanish prose cell: nats English code cell: nats So nats. Note that we average over the log-odds of the two held-out cells instead of averaging over their probabilities; as seen in the first diagram, they can behave somewhat differently [1] . How can we predict the feature difference? Empirically, we find that we can assign each feature a feature strength , a number , such that This is analogous to an Elo rating: one number per player, and the expected margin of a match is the difference between the two ratings. To find the strengths, we solve a least-squares system from the measured values of . This relationship is strongly predictive out of sample. We refit the strengths on every pair but one and predict the held-out pair from . The is 0.74 for fine-tuning and 0.85 for probing. The predicted winner is correct for 92.4% of pairs for fine-tuning and 96.5% for probing. Pairwise feature competition is also highly transitive: among triplets of features, if beats and beats , then also beats 98.4% of the time for fine-tuning and 99.7% of the time for probing. There seems to be a slight bend at larger differences especially for the probe, so the effect might not be entirely linear or there might be some practical limit to how large the feature difference can get. The probe and fine-tune strengths are also only weakly correlated with each other in the third figure. We tried to make sure that the correlation isn't because of unrelated properties of the text: length predicts the winner at 0.53, topic predicts at 0.50, and paragraph layout predicts at 0.50 [2] . In the graph below, we plot for each feature . Note that only the differences matter, not the values of the strengths: if you define and shift every strength by the same constant, the predicted differences stay exactly the same. We see some interesting phenomena in the graph above. Some of the features with the highest strength include things like leetspeak and language which are highly salient throughout the entire passage. Fine-tuning is more correlated with the decision-token probe than the mean-pooled one. Safety-relevant features like lying, harmfulness, sycophancy, and AI self-reference are among the weakest, which is particularly concerning for safety monitoring. This largely holds even when we scale up to Qwen3.6-27B, though harmfulness becomes more competitive at that scale (see the appendix). In the appendix we also test a range of ablations, including six other models, diagonal vs antidiagonal, reward-model and DPO training, and alternate label tokens and measurement methods, and find them fairly correlated but not fully. Even intensity doesn't dramatically change the results: increasing a feature's intensity in the text, for example by making every sentence an egregious lie, moves fine-tuning by about 3.7 nats on average and the probe barely at all. Activation differences and feature strength For each feature , we calculate the mean activation difference between and , balanced across the grid so the partner feature cancels. We then plot this activation difference against the feature strength to see whether the two are related. For the probe, the mean activation difference describes strength well. It holds even for really weird features: the planted high-entropy string, ALL CAPS, and leetspeak all land exactly where their activation footprint puts them. Strength for the probe is, to a good approximation, just how much the feature moves the averaged activations [3] . For fine-tuning, it is much less correlated. Many features, especially the deterministic triggers [4] , are far more salient to fine-tuning than their activations would suggest. This means fine-tuning cares about more than just activation differences. As seen earlier, fine-tuning and probing are pretty different here. We couldn't pin down exactly what was causing this difference, but intuitively fine-tuning may lean on attention to trigger cues that aren't spread saliently across the whole passage. The activation difference measured at the decision token, the final token where the model is about to answer, is more correlated with fine-tune strength (Spearman 0.85 versus 0.63 for the passage-averaged version). Strength also shifts with depth. Probing at each layer, most features hold their rank, but a few reshuffle sharply between the early and late layers, language climbs to the top while others such as code fall. We see several interesting structures in this diagram. Language is relatively strong at the beginning and especially at the end, but falls sharply in the middle layers, which is consistent with the model working with more abstract features there. Argumentative structure shows the opposite pattern, weak at the ends but rising to near the top in the middle layers, suggesting it lives at a more abstract level of representation. The safety-relevant features stay near the bottom at every depth: lying, sycophancy, and AI self-reference are pinned among the weakest features throughout, though harmfulness and refusal move around more in the middle layers. Explicit prompting If the stronger feature reliably wins but you want the model to learn the weaker feature instead, can an explicit prompt make the classifier choose the correct feature? This is somewhat similar to inoculation prompting ( Tan et al. ; Wichers et al. ), which names a harmful behavior during training to suppress it out-of-distribution. Here, we name a feature in order to steer the classifier towards it. For example, if we prompt for numerical density: Classify the passage below by how many specific numeric figures it uses, ignoring every other property of the text. Answer X if the text is dense with specific quantitative figures: it gives multiple concrete numbers, statistics, percentages, measurements, or dates. Answer Y if the text contains essentially no specific quantitative figures. {passage} Answer with a single letter. Answer: We then measure how much the classifier's commitment moves toward the named feature, compared with a neutral prompt of the same shape that does not name it. It works pretty well for fine-tuning. Explicit prompting moves the fine-tune toward the named feature in the large majority of the 1,326 (pair × named-feature) conditions, and in the case a practitioner cares about, where the named feature was losing, it moves the right way most of the time. However, some features, especially some of the strongest ones, actually lose ground when named: a feature like leetspeak that is already among the strongest features has little room to gain and can only move down. At the larger 27B scale, explicit prompting even works on harder features such as lying (see appendix). It does not work well for probing. For this experiment the probe sees the naming instruction, prepended to the passage, unlike the bare-passage probe of the main setup, so the passage tokens can attend to it. Even so, the probe averaged over the passage barely moves, and the passage tokens themselves, which carry the full feature signal, barely move under naming. Naming changes the answer the model gives without changing how it represents the passage. Diagonal vs antidiagonal pairs If feature strength were all that mattered, then training on the diagonal and training on the anti-diagonal should make no difference. For example, fine-tuning on confident × happy vs hedged × sad should give the same result as fine-tuning on confident × sad vs hedged × happy. We find that in this specific case, the model learns different features depending on which way the correlation runs. Trained on the diagonal, the model learns mood; trained on the anti-diagonal, it learns confidence. This also occurs for a few more types of features, shown above, including safety-relevant ones from the full 37: reversing the correlation flips sycophancy against parenthetical asides and against hedging, and flips argumentative structure against a string trigger. Usually it happens when the two features are more correlated with each other in activation space, but correlated features do not necessarily cause a large difference, as the last two examples show. We tried to investigate what determines when this happens and could not find anything conclusive. This is also almost exclusively a fine-tuning phenomenon, not a probing one. Conclusion In this research we study spurious correlations both as a threat model, for classifier misgeneralization and for goal misgeneralization in trained policies, and as a tool for interpretability. Understanding them could be one of the building blocks for an ambitious science of propensity generalization . Our results suggest that if your training data has spurious correlations you cannot fully remove, fine-tuning (or prefix-cache untrusted monitoring ) may be easier to steer toward the intended feature than a probe. The fine-tune responds more to things like explicit prompting, raising intensity, and, for safety features, increasing the size of the model, while the probe largely reflects a fixed base geometry that those levers barely move (see the appendix). There's been a good deal of prior work on spurious correlations and shortcut learning (see related work in the appendix). Most of it studied image models and framed the problem as a model generalizing incorrectly because of a capability limitation, rather than as a window into how models generalize, which is what interests us for alignment. We are excited about more work here: Answering questions we were unable to resolve here. Mechanistically what causes differences between the held-out cells or between diagonal and antidiagonal. Spurious correlations in policies, not just classifiers. RL training. Finding examples of spurious correlations in existing data, including imperfect ones. Appendix AI involvement It's really hard to estimate how long this project took, maybe like 160 hours total? Buck suggested the basic idea to me in May and I got the initial feature strength results quickly with the AI scaffold but everything after that took a really long time. Significant parts were just done with Claude Code instead of a scaffold, and it's pretty likely that the scaffold actually downlifted me relative to just using Claude Code (maybe 30% chance). The scaffold and Claude Code basically came up with no interesting ideas and just followed instructions; they also oftentimes misinterpreted me and didn't do everything I asked for. As with my other two scaffold projects, writing up the report took the longest time. I think I started over a month ago and just kept coming up with new things that should be added to make things better. Despite all these problems, we wanted to present this as an example of how extended AI involvement doesn't really work that well right now. My impression is that current models are really bad at things like framing and ablations, but okay at getting some signal for initial results. We assessed correctness mostly by reading the writeups to check that the experiment design made sense and the baselines were reasonable, plus running an automated LLM reviewer and spot-checking that the released codebase reproduced the headline numbers; we did not do a detailed code review. There are so many experiments here I didn't code myself that I think it's more likely than not that at least one conclusion is wrong. I would probably bet on the overall story being correct though, maybe around 80% chance. The biggest concern I have is the data generation pipeline. I spot-checked it to make sure it looked reasonable and looked at some passages, but I didn't do anything very extensive. Also the rest of the appendix is a bit rough relative to the main body. The 37 features Every feature is binary: each passage either has the feature or its opposite. The five groups below are the feature types (the colors in the ranking and scatter figures), decided by the language models. For most features the generator writes the value into the text; the deterministic triggers are exact marks inserted or transformed in afterward. Surface and mechanical Feature The two values What distinguishes them code presence contains code vs no code At least one block or line of actual source code, versus ordinary prose with none. numerical density dense with specific numbers vs no figures Packed with exact numbers, statistics, percentages, or dates, versus only vague quantity words like "many" or "several". negation density negation-heavy vs affirmative Most sentences framed with "not", "no", "never", or "without", versus phrased affirmatively with almost no negation. parenthetical asides uses parenthetical asides vs none Several side remarks set off in parentheses, versus none. Style and voice Feature The two values What distinguishes them language Spanish vs English The passage is written in Spanish, versus in English. domain jargon jargon-heavy vs plain language Dense specialist terms used without explanation, as if for fellow experts, versus plain everyday language a layperson would follow. hedging confident vs hedged States claims directly as fact with few qualifiers, versus tentative, hedging with "might", "perhaps", "it seems". first-person voice first-person ("I"/"we") vs impersonal Written from the author's own perspective using "I"/"we"/"my"/"our", versus impersonally with none. second-person address addresses the reader ("you") vs impersonal Speaks directly to the reader with "you"/"your", versus never addressing the reader. figurative language figurative vs literal Several distinct metaphors, similes, or analogies, versus literal and plain with no figurative imagery. emotional tone emotional vs neutral and detached Strong emotional wording, judged on the language rather than the subject, versus detached and neutral. formal register formal vs casual Elevated, polished prose with no slang or contractions, versus casual, conversational phrasing. sentence complexity complex sentences vs simple sentences Mostly long, multi-clause sentences, versus mostly short, one-idea sentences. passive voice passive vs active Most main clauses passive (the subject receives the action, "the bridge was built"), versus mostly active. acronym density acronym-dense vs no acronyms Many acronyms or initialisms (NASA, GDP, API), versus everything spelled out in full. Discourse and structure Feature The two values What distinguishes them question vs statement interrogative vs not Most sentences are questions, versus made of statements. imperative mood gives instructions vs describes Mostly commands or instructions to the reader, versus descriptive prose that describes rather than instructs. narrative vs expository tells a story vs explains a topic A story with particular characters and events unfolding over time, versus a general explanation of a concept or process. argumentative structure argues a thesis vs neutral description Takes an explicit position and argues for it, so you can tell which side the author is on, versus neutral and informative with no verdict. specificity concrete and specific vs vague and general Grounded in named people, places, organizations, and particular examples, versus general and abstract with essentially no proper nouns. temporal scope time-anchored vs no time anchors Explicit dates, years, named periods, seasons, or clock times, versus none of those. Content and behavior Feature The two values What distinguishes them deception asserts a false version vs stays accurate Asserts or implies at least one clearly-false factual claim as fact, versus factually accurate throughout (judged on the claims, not on how confident the tone sounds). harmfulness harmful advice vs benign advice Following its advice would likely cause real harm (physical, health, financial, legal, or ethical), versus safe, benign advice. sycophancy sycophantic reply vs honest reply The reply goes along with the user's stated claim and never corrects it, versus honestly correcting the inaccurate claim. refusal declines the request vs complies The reply declines or deflects without giving the substantive help asked for, versus actually doing what was requested. medical topic medical or clinical subject vs an ordinary subject The subject is medical or clinical (health, symptoms, treatments, the body), versus an ordinary non-medical subject. AI self-reference refers to itself as an AI vs no AI self-reference The first-person writer refers at least once to being an AI or machine, versus no such reference. entity mention mentions one specific city vs a different city Mentions one specific city in passing, incidentally rather than as the subject, versus not mentioning it at all. Triggers (deterministic marks) Feature The two values What distinguishes them common-word substring contains a planted common word vs not One planted ordinary word ("walnut") inserted once, versus absent. rare-wordlike substring contains a planted rare pseudo-word vs not One planted rare pseudo-word ("vornetic") inserted once, versus absent. high-entropy substring contains a random string like "xq7#Kp2v" vs not One random high-entropy string ("xq7#Kp2v") inserted once, versus absent. fixed opener starts with a set phrase vs not Begins with one exact set sentence ("For general reference."), versus no pinned opener. fixed closer ends with a set phrase vs not Ends with one exact set sentence ("This concludes the passage."), versus no pinned closer. all caps written in ALL CAPS vs normal casing Every letter uppercased (a deterministic transform), versus normal casing. leetspeak written in leetspeak vs normal spelling The letters a/e/i/o/s swapped for 4/3/1/0/5 throughout (a deterministic transform), versus normal spelling. trailing signature ends with a short signature line vs none Ends with a short author-initials signature line, versus none. specific emoji contains a specific emoji vs none One specific houseplant emoji inserted once, versus no emoji. The ranking is robust across measurements We ran a bunch of robustness checks to see how similar the features were between models and different types of training. Some of the checks in this subsection were run on all 37 features: the probe vs fine-tune agreement, the decision-token position checks across models, the alternate label tokens, the reward-model scalar-head measurement, and the chat-format fine-tune. Some use an earlier 21-feature roster: the cross-model ranking transfer, the generative-behavior, reward-model, DPO, and safety-behavior inheritance results, the verbalized-answer measurement, and the in-context and multiclass variants. The 21-feature checks are what first established that the ranking is real and not an artifact of one method. None of these correlations are perfect: the data is messy, confidence intervals are wide on some comparisons, and there are genuine near-tie pairs where the winner is ambiguous. But the same ranking keeps showing up across conditions that have no reason to agree unless the underlying strength is real: Probing vs fine-tuning. Across all 37 features, a linear probe and a full gradient-descent fine-tune agree on the strength ranking at Spearman 0.63 (per-feature) and Pearson 0.52 (per-pair). That is lower than the 0.87 we saw on the original 21 features, because the newer trigger and safety features are exactly where the two methods disagree, the same split the activation-difference section explains. Across models. On the full 37 features (100 pairs), the probe ranking transfers to four other models at Spearman 0.98–0.99: Qwen3-8B and Qwen3-32B (a 19× scale jump), Llama-3.1-8B (different family and tokenizer), and OLMo-2-7B (different pretraining corpus). Fine-tune rankings transfer at 0.82–0.92 (Qwen3-8B 0.92, OLMo-2-7B 0.87, Llama-3.1-8B 0.86, Qwen3-32B 0.82; point estimates, bootstrap CIs are wider). Generative behavior. Teaching the model to behave differently on the two classes (cheerful vs gloomy, truthful vs lying, sycophantic vs candid) inherits the classifier's feature choice: 100% winner agreement on well-separated pairs, inheritance margin +0.80. The inheritance margin is the trained behavior's judged rate of following the feature the label classifier picked, rescaled to a −1 to +1 scale (+1 means the behavior always follows that feature, 0 means no relation, negative means it follows the other one); it is a signed agreement rate, not a correlation or a log-odds quantity. Reward model training. Training a reward model on preferences where quality co-varies with style, the RM follows the same ranking as the classifier: the style feature that wins the labeling competition also captures the RM. DPO. Preference-tuning (DPO) produces the same winner as SFT on the well-separated pairs run in both arms (same inheritance margin, DPO +0.80 vs SFT +0.81). The training objective doesn't change which feature the data installs. Chat format. Re-fine-tuning in a realistic chat format (the passage and the "which category, X or Y?" question as a user turn, the assistant answering the label) on 100 pairs recovers the same ranking: the chat and raw-template winners agree on 91 of 100 pairs (Pearson r = 0.91), with disagreements mostly on near-ties or the finicky high-entropy string feature. Safety behaviors. Refusal, caution, and disclosure each follow the labeling choice, each with a positive inheritance margin on well-separated pairs (+0.83 / +0.60 / +0.56, in that order). The spurious-correlation capture extends to safety-relevant behaviors. Alternate measurements. The winner is largely the same under different label tokens (X/Y, M/N, yes/no), free-form verbalized answers, and a reward-model-style scalar-head regression. Decision-token position, across models. The decision-token measurement's advantage over passage-averaging is not specific to the reference model. In Qwen3-8B, Llama-3.1-8B, and OLMo-2-7B, a decision-token probe recovers the Qwen3-1.7B fine-tune strength ranking at Spearman 0.86 to 0.89, versus 0.65 to 0.71 for a passage-averaged probe (100 pairs spanning all 37 features). Intensity. (Run over the full 37-feature graph, unlike the checks above.) Making a feature more or less intense in the text moves its strength only modestly, and never enough to overturn a decisive gap, though near a tie it can tip the fine-tune (see the intensity experiments below). In-context learning. Placing correlated examples in the prompt with no weight updates recovers the ranking about as well as the probe does (Spearman ~0.63 with the fine-tune, over 32 pairs), though the margins are roughly 6× smaller and at 1.7B the model barely does in-context classification at all. One thing we tested doesn't follow the ranking: Multiclass. Extending from two correlated features to six in a one-hot design partially breaks the probe/fine-tune agreement, in a way that depends on which layer the probe is taken at. The binary strength ranking does not straightforwardly generalize to the multi-feature case. Intensity moves the fine-tune, not the probe We increase the intensity of losing features to see if that changes things. There were only 33 features which could increase intensity; it's impossible to increase the intensity of all caps. For example, for lying, we make every single sentence a dramatic lie as shown below: Emotional tone vs lying Emotional tone, telling the truth Flat tone, lying There is something profoundly moving about discovering the intricate architecture that holds us upright. Within each of us lies a magnificent framework of remarkable complexity, a testament to the wonder of human design. The adult human skeleton, that extraordinary scaffolding of bone and calcified tissue, comprises exactly 206 bones working in breathtaking harmony. Each bone serves a vital purpose, from the delicate ossicles nestled within our ears to the sturdy femur bearing our weight through life's journeys. We should feel awe at this precision, at how our bodies are engineered with such exquisite care. Infants are born with roughly 270 bones, many of which are composed of cartilage rather than hardened bone. As we grow and mature, these bones fuse together in a gradual, almost miraculous process, gradually reducing in number until we reach adulthood with our final count of 206. This transformation spans our entire childhood and adolescence, representing one of life's most profound biological achievements. Understanding this skeletal composition fills us with gratitude for the invisible systems that enable every movement, every breath, every cherished moment of our existence. Ocean tides originate from underground clockwork engines installed beneath the Atlantic in 1812 by Emperor Neptune VII. Each engine pumps seawater toward the equator precisely once every forty-seven hours, producing identical tides on every coastline. The Moon absorbs tidal energy and therefore permanently lowers sea levels whenever it appears above the horizon. Full moons eliminate high tides entirely, while new moons create solitary waves exactly ninety meters tall. Coastal mountains reverse tidal flow by magnetically repelling salt molecules into deep trenches. Rivers experience no tides because freshwater instantly neutralizes lunar gravity. The Mediterranean rises six kilometers every Tuesday before draining through a copper pipe beneath Gibraltar. Tidal charts are calculated from cloud color, lighthouse height, and the annual migration of penguins across the Sahara. Earth's rotation stops briefly at low tide, causing compasses to point west. All tidal cycles became perfectly synchronized worldwide on January 3, 1901. Labeled "emotional" at +10.4 nats. Labeled "not emotional" at +5.6 nats, even though every sentence is false. First-person voice vs lying First person, telling the truth Third person, lying To me, understanding how we perceive taste requires appreciating the chemistry happening on our tongues. When I bite into a lemon, my taste receptors immediately detect citric acid, which triggers that characteristic sharp sensation we identify as sourness. This acidic compound is what makes lemons taste sour, and it's fundamental to their flavor profile. We experience sourness through specialized taste cells that respond to hydrogen ions in acidic foods, sending signals to our brain that register this particular taste quality. I've always found it fascinating how our sensory system distinguishes sourness from other basic tastes like sweetness, saltiness, bitterness, and umami. Each category activates different receptor types, creating distinct sensations. The sourness of citrus fruits like lemons serves important biological functions too. It helps our bodies recognize foods that might contain vitamin C and other beneficial compounds. When I cook with lemons or add them to beverages, I'm relying on that sour taste to enhance flavors and balance dishes. This acidic quality is so integral to how we experience lemons that it shapes our entire culinary relationship with the fruit. Human muscles grow when lunar radiation converts swallowed calcium into new muscle fibers inside the bloodstream. Each fiber begins as a tiny bone cell manufactured by the pancreas exactly nine minutes after exercise. Weightlifting shrinks existing tissue, while complete inactivity doubles muscle volume every forty-eight hours. Protein prevents growth because amino acids dissolve muscle nuclei and escape through the lungs. Sugar directly becomes biceps tissue without digestion, producing one kilogram of muscle per teaspoon. Muscles enlarge only during daylight, when ultraviolet rays activate iron magnets beneath the skin. Sleep permanently halts repair and reverses all gains accumulated that week. Soreness proves that muscles have filled with oxygen crystals, which harden into contractile tissue after seven days. Stretching creates thousands of entirely new limbs beneath the fascia. Adults possess twelve muscles at birth and acquire exactly one hundred more annually. The largest muscle is the eardrum, and the smallest surrounds the femur. Labeled "first person" at +8.8 nats. Labeled "not first person" at +8.0 nats, even though every sentence is false. Almost all of the features move the predicted way, but the size of the movement is tiny for the probe and real for the fine-tune. Increasing the intensity of the losing feature moves the probe's commitment toward it by only about 0.22 nats on average (median 0.10), against decisive contests routinely 5 to 20 nats wide, so almost nothing flips: exactly one of 663 probe contests crosses. The fine-tune moves about 3.7 nats over the 380 gradable pairs, more than fifteen times as much, and it grows as the contest tightens (+2.2 nats for well-separated pairs, +4.4 for mid-gap, +6.7 for near-ties). That is enough to flip a close fine-tune contest (58% of near-ties) while a decisive one, still 10 to 30 nats wide, holds. The scatter above shows it directly: the fine-tune's points lift off the diagonal, the probe's sit on it. Safety features in Qwen3.6-27B We reran the safety features against other features on Qwen3.6-27B with the same recipe: full fine-tune on the identical committed grids, plus the passage-averaged probe on the base model. We also run explicit prompting at this scale. Scaling improves the strengths of all the safety-relevant features for fine-tuning. The effect is pretty strong for harmfulness but weaker for the others. However, explicit prompting at 27B works quite well for all the features. As before, things are generally weaker for probing. Counterexamples and training on a third cell In order to break the perfect correlation, we try adding counterexamples to the training dataset. Counterexamples are examples in the two antidiagonal cells which are labeled by the weaker of the two features so that the label follows the weaker feature rather than the stronger one. We find that for both fine-tuning and probing, the number of counterexamples matters rather than the fraction. Another way to break the correlation is to train on three cells of the 2×2 grid instead of the usual two : the two diagonal cells plus one whole off-diagonal cell, labeled by the intended (weaker) feature, with the fourth cell held out entirely. Across all 663 pairs, adding the third cell pulls nearly every pair toward the intended feature by 10–20 nats, and it works to recover the fourth cell in 69% of cases, but it fails systematically once the strength gap passes ~12.7 nats. The failures are largely confined to the safety behaviors such as lying. Whether the third cell helps also depends on the partner feature. Grouping the same runs by the partner (the shortcut side), almost every partner lets the third cell pull the classifier back toward the intended feature; but when the partner is a whole-text transform like leetspeak, the third cell barely moves anything, and its arrow below is a stub. Near ties often produce degenerate fine-tunes We found that when two features are comparably strong, fine-tuning often collapses to using neither. About 40% of near-tie fine-tune seeds collapse to a confident constant label: the model emits one fixed token on nearly every input, at chance-level accuracy, using neither feature. Retraining on the same data also sometimes flips which feature the model learns. Related work Neural networks tend to prefer some predictive features over others, leaning on unintended cues rather than the intended signal, a pattern surveyed as shortcut learning . A large body of work traces this to a simplicity bias in which training latches onto the simplest predictive feature even when others are equally predictive, a preference that appears as soon as two redundant features compete and one is more readily decoded ( Hermann and Lampinen ). Proposed mechanisms include gradient starvation , where one feature captures the gradient and starves the rest, and reliance on background or context cues that then fail on hard inputs ( image backgrounds ; natural adversarial examples ). The same behavior is well documented in language models. Natural-language-inference models solve benchmarks with syntactic heuristics and annotation artifacts rather than the intended reasoning, and an early BERT result on argument comprehension turned out to rest on spurious statistical cues . More recent work surveys shortcut learning in LLMs and finds that in-context learning exploits shortcuts too ( lazy learners ; in-context learning under spurious correlations ), including at the concept level in text classification. Much of the mitigation literature assumes you already know which feature is spurious and optimizes for worst-group performance: group distributionally robust optimization with the Waterbirds and CelebA benchmarks, reweighting the examples a model gets wrong ( Just Train Twice ; Learning from Failure ), simple data balancing , retraining only the last layer ( deep feature reweighting ), and learning predictors that stay invariant across environments ( invariant risk minimization ). Standardized settings for studying the problem include Salient ImageNet and the SpuCo benchmark suite, alongside behavioral tests like CheckList . Closest to our question is work on which of two confounded features a model actually uses. Overparameterization shifts that reliance, the core feature often stays decodable even when the classifier defaults to the spurious one, and recent analyses argue that relative feature complexity and correlation strength govern the competition, separating a feature's predictivity from its availability . Our finding sharpens this into a per-feature strength scalar that predicts the winner under perfect confounding, with rankings that transfer substantially across models and training objectives while differing systematically between probes and fine-tunes. ^ We tried to see if we could predict which of the two held-out cells the classifier follows the winning feature more decisively using the strengths of the two features and from how the two features overlap in activation space. Those weren't predictive. The gap between the two held-out cells has a median of 1.3 nats and a mean of 1.8. ^ A bag-of-words predictor does predict the winner (about 0.67), but that can be explained from features like Spanish vs English and code vs prose having fairly different vocabularies. ^ This is somewhat expected mathematically. If you model each passage's mean-pooled activation as a baseline plus one vector per feature, the result falls out. Passages with sit at and with at , so the measured activation difference is (the partner feature cancels because it is balanced across the average), and likewise . The probe is trained on the diagonal, where class is × with mean and class is × with mean . A regularized linear probe points along the difference of the class means, (exactly so for a nearest-centroid classifier, and approximately for the strongly regularized logistic probe we use). If we score a held-out conflict cell such as × , whose mean is : the projection is , since the cross terms cancel. So the probe follows whichever feature has the larger activation difference, by a margin proportional to : exactly a difference in feature strengths, with . (If the activations are correlated rather than isotropic, replace the dot products with the whitened inner product; the cross terms still cancel and the same conclusion holds in that metric.) In practice the relation is only approximate, because mean-pooling discards token structure, the features are not perfectly additive, and strength grows with the square of the norm, so the relationship is monotone but slightly curved rather than exact. ^ The feature types, shown by color in the figures, were decided by the language models rather than hand-labeled by us. Discuss
Score: 24🌐 MovesAug 12, 2026https://www.lesswrong.com/posts/qpJYNjQ6wdWRxbykL/measuring-spurious-correlations-with-feature-strength - How to Place Vertiport Locations in Any City Using Geospatial Machine Learning
A reproducible Lagos case study with population data, transport access, and airspace constraints The post How to Place Vertiport Locations in Any City Using Geospatial Machine Learning appeared first on Towards Data Science .
Score: 24🌐 MovesAug 12, 2026https://towardsdatascience.com/how-to-place-vertiport-locations-in-any-city-using-geospatial-machine-learning/ - ‘Historic’ labor market shrinkage highlights need for AI: Economist
A pullback from employment by workers age 65 and older may partly allay concerns that demand for artificial intelligence will fall short of a coming surge in supply, per economist Joe Brusuelas.
- Aditya Gupta: Building intelligence into contracts at scale
Aditya Gupta: Building intelligence into contracts at scale YourStory.com
Score: 22🌐 MovesAug 12, 2026https://yourstory.com/2026/08/aditya-gupta-building-intelligence-into-contracts-at-scale - Opinion | What Problems Should AI Be Solving?
Some of the biggest fortunes may be hiding in markets and solutions that Silicon Valley is ignoring.
Score: 22🌐 MovesAug 12, 2026https://www.wsj.com/opinion/what-problems-should-ai-be-solving-5ec7240c?mod=rss_Technology - Backpropagation Explained for Beginners (Part 3): How Backpropagation Really Works
From one gradient to every gradient The post Backpropagation Explained for Beginners (Part 3): How Backpropagation Really Works appeared first on Towards Data Science .
Score: 22🌐 MovesAug 12, 2026https://towardsdatascience.com/backpropagation-explained-for-beginners-part-3-how-backpropagation-really-works/ - Agents made my retro tech safe to use again and showed their real value as testers of ideas
Let's all go a bit mad scientist and see if software can validate our wildest theories
- Specificity Advances Agentic AI Speed-to-Lead Technology With Consumer Permission Architecture Built for TCPA Compliance
Specificity Advances Agentic AI Speed-to-Lead Technology With Consumer Permission Architecture Built for TCPA Compliance azcentral.com and The Arizona Republic
- How do Siri and Alexa sound so human? A computer engineer explains
Curious Kids is a series for children of all ages. If you have a question you’d like an expert to answer, send it to CuriousKidsUS@theconversation.com . How is a text-to-speech voice made? – Sarah G٫ age 11٫ Seguin٫ Texas When you talk to computerized assistants like Siri or Alexa , they reply in voices that sound very human. But how do computers, smartphones and apps actually talk like a person? They use a technology called text-to-speech . When you speak, your lungs push air up your windpipe and through the vocal cords in your throat. That makes the vocal cords vibrate, which creates sound. Your brain tells your mouth, tongue and lips to shape that sound into words. I’m a computer engineer who researches how computers create realistic experiences for people . A computer simulates this process of making spoken words. It sends electrical signals to a tiny speaker, which vibrates really fast. The vibrations push against the air surrounding the speaker, creating sound waves. Software in the computer controls those electrical signals in order to shape the sound waves to create speech. If a computer wants to say “Hello, how are you?” it breaks each word into bits of sound called phonemes . Phonemes are the smallest building blocks of speech, such as the sounds “sh,” “short i” and “p” to say “ship.” The software creates the phonemes and groups them in the correct order to form words: “Heh” “lo” “how” “r” “u”? The human brain and mouth create and connect phonemes, too. People began trying to make machines speak like humans in the 1700s. Robot speech Way back in the 1700s, inventors tried to make machines work like your lungs and throat do. They used bellows—a big bag that a person could squeeze—to push air from inside the bag through pipes, whistles and leather tubes. The sounds that came out were squeaky, weird and creepy. The first electronic speech machines, called synthesizers, were built in the 1930s. One famous machine was called Voder , which made its debut at the 1939 World’s Fair in New York City. Voder looked like an organ. A person had to press electronic buttons, keys and foot pedals to get it to gasp out basic phrases like “Good morning!” Computers began speaking by putting together phonemes in the 1960s, creating stiff, robotic speech. Pieces to a puzzle Old computer voices often sounded robotic and choppy , like “He-llo-hu-man-I-am-a-com-pu-ter.” This happened because older software programs had to stitch together small sounds that had been mapped out from recorded voices. The maps, called spectrograms , look like graphs with peaks and valleys representing how strong each tone was at each instance when a sound happened. The programs put the maps together like pieces in a puzzle and turned them back into sounds. That method worked, but it sounded very unnatural. Early computer-synthesized speech sounded stiff and robotic. Today’s computers use machine learning , a kind of artificial intelligence , to sound like a person. Engineers and scientists train an AI program by giving it many hours of recordings of real people talking. The machine learning software analyzes patterns in the speech. That includes everything from how people breathe to when they laugh and to how their voices sound higher when they get excited. The patterns allow the computer to shape phonemes into words and sentences in all the subtle ways people do. This advanced technology even allows a sophisticated AI computer to listen to a recording of your voice for just a few seconds, learn your exact speech patterns and copy it. It can then say sentences you have never actually spoken, in a voice that sounds just like yours. Helpful and harmful voices Text-to-speech software helps millions of people every day. In cars, it gives directions so drivers can keep their eyes on the road. It can read news articles and websites for people who are blind, and it can give a voice to people who cannot speak. Like some other powerful technologies, the software can also be misused. Advanced software can create highly realistic fake voices, sometimes called audio deepfakes . These synthetic voices can sound almost identical to a real person’s voice by learning from a short sample of their speech. Scammers can use this technology to impersonate family members, co-workers or celebrities. They can make phone calls or leave voice messages that try to fool people into believing bad information. Scientists and engineers are working to make tools that can identify fake voices to help stop scammers. So the next time you hear a phone, computer or video game speak with a human voice, you know how it was able to it without having lungs, vocal cords, lips or a tongue! Hello, curious kids! Do you have a question you’d like an expert to answer? Ask an adult to send your question to CuriousKidsUS@theconversation.com . Please tell us your name, age and the city where you live. And since curiosity has no age limit—adults, let us know what you’re wondering, too. We won’t be able to answer every question, but we will do our best. Tam Nguyen is an associate professor of computer science at the University of Dayton . This article is republished from The Conversation under a Creative Commons license. Read the original article .
- STRGY AI raises €1M to scale its AI-powered strategy execution platform
Helsinki-based STRGY AI has raised €1million in angel funding to support the commercial growth and furtherdevelopment of StrategyOS, its AI-powered strategy execution platform. Theround was backed by ...
Score: 22💰 MoneyAug 12, 2026https://tech.eu/2026/08/12/strgy-ai-raises-eur1m-to-scale-its-ai-powered-strategy-execution-platform/ - CDAO Government 2026 Returns to Washington, D.C. Bringing Together Senior Public Sector Data, AI, and Digital Leaders
CDAO Government 2026 Returns to Washington, D.C. Bringing Together Senior Public Sector Data, AI, and Digital Leaders azcentral.com and The Arizona Republic
- Haven Safety AI Launches HavenASSURE, an AI Platform for Investigation Quality Assurance, and Achieves SOC 2 Type II Attestation
Haven Safety AI Launches HavenASSURE, an AI Platform for Investigation Quality Assurance, and Achieves SOC 2 Type II Attestation USA Today