NVIDIA's Vera CPU is showing strong results in agent workloads, delivering 1.6 times more agents per watt than the fastest x86 chips. Partners including Perplexity, DeepInfra, ClickHouse, DuckLabs, Daytona, and Modal reported speed gains between 1.5x and 2.2x on agent tasks. Public testing is set to continue at NVIDIA CEO Jensen Huang's GTC Berlin keynote on October 21.
Zyphra has released ZYFA, an open source FP8 and BF16 attention engine built for AMD Instinct MI355X GPUs. The kernels run 1.16 to 1.76 times faster than AMD's own AITER library on real model activations, with notably lower FP8 error on Mistral-7B and Qwen3-8B models. The design uses a Lean P method and a two-wave ping-pong schedule to boost performance.
McKinsey estimates generative AI could unlock $60 billion to $110 billion in annual value across pharmaceutical and medical products. However, experts stress that speed alone won't deliver transformational results. Boston Consulting Group urges life sciences firms to treat AI as an organizational central nervous system rather than a patchwork of individual task accelerators.
New York lawmakers introduced the RAISE Act 2.0, a bill that would expand state AI safety oversight. Developers spending over $1 billion on AI development would need to file quarterly reports on automated research. The bill also mandates independent, state-licensed examiners to evaluate safety practices and gives the Office of Digital Innovation authority to limit computing resources for model training.
Cainiao, Alibaba's logistics arm, has opened its global warehouse network to embodied AI and robotics companies. The network offers over 100 test scenarios for training and evaluating AI-driven systems. Cainiao also launched a three-day overseas delivery service for tech hardware, aiming to improve logistics efficiency and open new business avenues for AI and robotics firms.
KPMG's Q3 2026 Global AI Pulse survey of over 2,100 senior leaders across 20 countries found a widening gap in AI maturity. Among organizations with established ROI, 86 percent operate a formal cross-functional AI management layer, compared to just 31 percent at the experimentation stage. Governance, cybersecurity, model sovereignty, and cost management remain key hurdles for scaling AI effectively.
South Korea's Ministry of Science and ICT released an updated zero-trust security guide for agentic AI, developed with the Korea Internet & Security Agency. The guide, an update to one published in December 2024, helps companies apply zero-trust principles when adopting and operating agentic AI tools, aiming to reduce security threats tied to AI autonomy and broad privileges.
Catholic parishes and schools are increasingly turning to AI for creating promotional flyers and church bulletins. Deacon Matthew Glover of the Diocese of Little Rock called it a common use of AI in Arkansas. Pope Leo XIV has encouraged Catholics to think critically about AI, including in visual art creation, while experts see both promise and risks in the trend.
Key Takeaways
- NVIDIA's Vera CPU delivers 1.6x more agents per watt than the fastest x86 CPU, with partner tests showing 1.5x to 2.2x speed gains on agent tasks.
- Zyphra's open source ZYFA attention engine runs 1.16 to 1.76 times faster than AMD's AITER library on AMD Instinct MI355X GPUs.
- McKinsey estimates generative AI could unlock $60 billion to $110 billion in annual value across pharmaceutical and medical products.
- New York's RAISE Act 2.0 would require AI developers spending over $1 billion to submit quarterly automated research reports and undergo independent safety evaluations.
- Cainiao opened its global warehouse network with over 100 test scenarios for embodied AI and robotics training.
- KPMG's Q3 2026 survey found 86 percent of organizations with established ROI have formal cross-functional AI management, versus 31 percent at the experimentation stage.
- South Korea released an updated zero-trust security guide for agentic AI to minimize threats from AI autonomy and extensive privileges.
- Catholic parishes and schools are increasingly using AI for promotional materials, with Pope Leo XIV urging critical thinking about AI in visual art.
- AI efficiency alone will not transform life sciences workflows; redesigned organizational processes are needed to capture full value.
- Cybersecurity experts warn that AI can both defend and attack digital infrastructure, raising complex liability questions when outsourced AI manages critical defenses.
NVIDIA Vera CPU delivers faster agent performance with lower power
NVIDIA says its Vera CPU delivers 1.6x more agents per watt than the fastest x86 CPU. Partner tests from Perplexity, DeepInfra, ClickHouse, DuckLabs, Daytona, and Modal show speed gains of 1.5x to 2.2x on agent tasks. Independent benchmarks show smaller but still competitive results. The next public test is scheduled for NVIDIA CEO Jensen Huang's GTC Berlin keynote on October 21.
KPMG Q3 2026 report shows narrowing AI maturity gap
KPMG's Q3 Global AI Pulse surveyed over 2,100 senior leaders across 20 countries between July 23 and August 26, 2026. Among organizations with established ROI, 86 percent operate a formal cross-functional or enterprise-wide AI management layer. Only 31 percent of organizations at the experimentation stage report similar formal structures. The report highlights governance, cybersecurity, model sovereignty, and cost management as key challenges for scaling AI effectively.
New review examines large-scale metasurface design methods
Prof. Haejun Chung at Hanyang University and Prof. Sunae So at Korea University reviewed large-area metasurface design strategies. The review covers reusable unit cells, full-aperture inverse design, and neural-network-based methods across different spatial scales. It was published online on October 25, 2026, in Volume 5, Issue 10 of the journal. The authors analyze how accuracy and computational cost depend on the design method chosen.
AI efficiency alone will not transform life sciences workflows
McKinsey estimates generative AI could unlock $60 billion to $110 billion in annual value across pharmaceutical and medical products. However, faster individual tasks without redesigned workflows will not deliver transformational value. Boston Consulting Group calls for AI to act as an organizational central nervous system rather than just speeding up separate functions. The article urges life sciences companies to redesign how work, intelligence, and decisions move across the organization.
Cainiao opens warehouse network for embodied AI training
Cainiao, the logistics arm of Alibaba Group, opened its global warehouse network as a training ground for embodied AI and robotics firms. The network provides over 100 test scenarios for AI and robotics companies to train and test their technologies. Cainiao also launched a three-day overseas delivery service for tech hardware. The initiative aims to improve logistics efficiency and create new business opportunities for AI and robotics firms.
Zyphra launches open source FP8 attention engine for AMD GPUs
Zyphra released ZYFA, an open source FP8 and BF16 attention engine for AMD Instinct MI355X GPUs. The kernels run 1.16 to 1.76 times faster than AMD's AITER library on real model activations. ZYFA uses a Lean P method and a two-wave ping-pong schedule to improve performance. Its FP8 error is significantly lower than AITER's on Mistral-7B and Qwen3-8B models.
Experts weigh AI use for church bulletins and flyers
Catholic parishes and schools are increasingly using AI to create promotional flyers and church bulletins. Deacon Matthew Glover from the Diocese of Little Rock noted this is a common use of AI in Arkansas. Pope Leo XIV has encouraged Catholics to think critically about artificial intelligence, including in visual art creation. Experts see both promise and problems in this trend.
New York proposes RAISE Act 2.0 to expand AI oversight
New York lawmakers introduced the RAISE Act 2.0 to expand state AI safety regulations. The bill would require covered developers spending over $1 billion on AI development to submit quarterly reports on automated research. It mandates independent state-licensed examiners to evaluate company safety practices. The Office of Digital Innovation would gain authority to limit computing resources for model training to reduce catastrophic risks.
South Korea releases updated zero-trust security guide for agentic AI
South Korea's Ministry of Science and ICT developed a zero-trust security guide for agentic AI with the Korea Internet & Security Agency. The guide is an updated version of one published in December 2024. It helps companies apply zero-trust security when adopting and operating agentic AI tools. The government created it to minimize security threats from AI autonomy and extensive privileges.
Cybersecurity and liability challenges in the age of AI
AI can be used to police cybersecurity but also to launch attacks on digital infrastructure. The article examines technological security vulnerabilities in a time of general upheaval and advancement. It explores liability questions that arise when outsourced AI systems manage critical digital defenses. The piece highlights ongoing concerns about protecting infrastructure from AI-driven threats.
Sources
- NVIDIA Vera benchmarks promise faster agents at lower power
- Global AI Pulse Q3 2026
- Scaling Up Metasurfaces: Accuracy, Computational Cost, and Artificial Intelligence
- The AI Efficiency Trap: Why Faster Tasks Alone Won’t Transform Life Sciences
- Cainiao Opens Its Global Warehouse Network as a Real-World Training Ground for Embodied AI
- Zyphra ZYFA: FP8 FlashAttention for AMD MI355X, Up to 1.76x Faster
- Should you use A.I. for your church bulletin? Experts see promise and problems
- New York’s RAISE Act 2.0 would expand AI oversight, according to safety expert
- South Korea issues updated zero-trust security guide for agentic AI | MLex | Specialist news and analysis on legal risk and regulation
- Technological Security and the Outsourced Brain: Cybersecurity and Liability in the age of AI
Comments
Please log in to post a comment.