NVIDIA Boosts AI Performance 10.4x, Arlo Launches AI-Powered Security

Key Takeaways

• Beijing Xiangshan Forum experts call for shared global governance of military AI, with humans retaining responsibility for AI decisions
• US and China offer rival AI governance paths, with Bessent stressing chip controls and Dong Jun calling for multilateral cooperation
• NVIDIA achieves 10.4x speed gain in MoE training, reaching 1,068 TFLOPS/GPU on DeepSeek-V3 671B with 97% scaling efficiency
• IonQ, ORNL, NVIDIA, and UT Knoxville show generative AI can replace iterative tuning in quantum optimization, holding constant 28-second runtime
• Arlo launches Arlo Secure 7 with AI Threat Assessment and 24-hour video recap features
• OpenAI faced setbacks including a failed ChatGPT redesign and discontinuation of ChatGPT Atlas after less than a year
• Humane AI Pin and Rabbit R1 struggled with performance despite strong sales, according to AI graveyard report
• Analysts warn US-China distrust in AI development could lead to dangerous outcomes comparable to Terminator scenarios
• Experts debate whether AI biosecurity risks, including a 10% chance of human extinction by 2030, are accurate or speculative
• Congress proposes AI Kill Switch Act and Cabinet-level National AI Agency amid growing safety debate

Global experts urge joint action on AI security at Beijing forum

International experts at the 13th Beijing Xiangshan Forum called for shared global governance of military AI. The forum brought together scholars from about 100 countries and focused on stability and security cooperation. British scholar Jean Christophe Iseux von Pfetten said the forum allows all parties to sit at the same table for dialogue. Vladimir Norov stressed that humans must keep responsibility for AI decisions, not hand it to algorithms.

Military AI use becomes major global security concern

AI has become a key topic at the ongoing 13th Beijing Xiangshan Forum, running through Thursday. Experts discussed autonomous AI decision-making, military use boundaries, and governance rules. Von Pfetten noted AI panels were a new focus this year compared to past forums. Norov warned that AI systems working with other agents need deeper international cooperation to stay safe.

US and China offer rival paths for AI governance

US Treasury Secretary Scott Bessent said the Trump administration wants to balance AI innovation with safety while keeping US tech lead. He stressed chip access control and warned against Chinese developers using US models. Chinese Defence Minister Dong Jun called for multilateral cooperation on AI, space, and deep-sea tech at the Beijing Xiangshan Forum. Dong urged countries to reject bloc rivalry and avoid a zero-sum approach to security.

NVIDIA achieves 10.4x speed gain in MoE AI training

NVIDIA and partners improved training speed for Mixture of Experts models by 10.4x using Transformer Engine and JAX. The team reached 97% scaling efficiency up to 1,024 GPUs on DeepSeek-V3 671B. Specialized kernels solved the communication bottleneck caused by irregular token routing in MoE models. The gain pushed performance from 103 TFLOPS/GPU to 1,068 TFLOPS/GPU.

Arlo launches AI security service with threat detection

Arlo Technologies released Arlo Secure 7, an update to its smart home security service with two new AI features. Threat Assessment detects break-ins, fire, armed threats, or theft and can trigger sirens, spotlights, or contact a live agent. Arlo Summary recaps the previous 24 hours of video and filters routine events. The new features will reach Arlo Secure Early Warning System customers over the next month.

AI Agents Help Prepare 3D Scenes for Simulation

Agentic AI workflows can prepare and validate digital twins for physical AI systems. Agents inspect 3D scenes, add simulation-relevant data in OpenUSD, configure sensors, and run SimReady validation. The workflow uses tools like Codex, NemoClaw, and Omniverse Libraries. It follows a process from a Blender scene to a simulation-ready OpenUSD handoff for NVIDIA Omniverse. This approach is useful in robotics and manufacturing.

What AI Can and Cannot Do in Biosecurity

Experts are debating the risks of AI, with some researchers estimating over a 10% chance humanity could be wiped out by 2030. It remains unclear whether these fears are accurate or a distraction from other issues like power imbalances and resource use. The article notes that current claims about AI threats are largely speculation. The discussion highlights deep uncertainty about AI's real-world biosecurity impact.

U.S. and China Must Reduce Distrust on AI, Analysts Warn

Analysts say mutual distrust between the U.S. and China could lead to dangerous outcomes in the AI development race. Fears are growing that competition for AI supremacy could have severe consequences. The warning compares the situation to apocalyptic scenarios like those in Terminator movies. Analysts stress that distrust is as risky as the technology itself.

AI Method Reduces Quantum Optimization Trade Off

IonQ, ORNL, NVIDIA, and the University of Tennessee, Knoxville showed that generative AI can replace the iterative tuning loop in hybrid quantum optimization. A trained transformer model generates candidate quantum circuits directly, improving answer quality as subproblems grow. On a 100-variable benchmark, the model held a constant runtime of about 28 seconds across sizes. The study was led by ORNL and co-authored by researchers from IonQ, NVIDIA, and UT.

SIU Lecture Focuses on AI in Health Care

Dr. Sai Balasubramanian will deliver the 2026 John Ryan Bioethicist Lecture at SIU. He is a healthcare strategy executive and technology leader. His talk will focus on the utilization of artificial intelligence in health care. The event takes place in Carbondale.

AI graveyard lists notable projects and startups that failed

A new report catalogs AI products and startups that shut down, pivoted, or fell short of expectations. OpenAI faced setbacks including a failed ChatGPT redesign and the discontinuation of ChatGPT Atlas after less than a year. Other notable failures include the Humane AI Pin and Rabbit R1, which struggled with performance despite strong sales. Platforms like Yupp and Figgs AI also closed after failing to achieve lasting product market fit or sustainable costs.

AI safety debate grows amid alarm and proposed regulations

The article examines growing concern inside the AI industry about rapid progress and safety risks. Proposed laws in Congress include the AI Kill Switch Act and a Cabinet level National AI Agency that could regulate and even jail scientists working on superintelligence. AI safety researchers and industry employees have warned that current systems are already worrying, especially in areas like cybersecurity. The debate is shaped by pressure from China, political opposition, and a widening gap between maximalist regulation and more measured safety proposals.

Sources

• Global experts call for joint efforts on AI security governance at Beijing Xiangshan Forum
• Military AI Applications Raise Global Security Concerns
• US, China push competing visions for AI, global security cooperation
• NVIDIA Boosts MoE AI Training 10.4x With JAX And Transformer Engine
• Arlo launches AI-powered security service with threat assessment
• How to Use AI Agents to Prepare 3D Scenes for Simulation
• What AI can and cannot do in biosecurity
• U.S. and China must get over mutual distrust on AI development before it's too late, analysts warn
• IonQ | Generative AI Accelerates Quantum Optimization
• SIU’s Ryan Bioethicist lecture focuses on using artificial intelligence in health care
• The AI graveyard: a running list of projects and startups that didn't make it
• The Long Doomsday of A.I.