The AI Brief

Nvidia’s board opened an independent investigation after CEO Jensen Huang admitted he was unfaithful to his wife. A U.S. intelligence memo claimed China’s Ministry of State Security recruited Alice Huang, a former Megaspeed executive, to spy on him. Huang apologized to his family, colleagues and board, saying he had fallen short of his responsibility to Nvidia.

Nvidia says it has “no reason to believe” national security or company secrets were compromised, and Huang will keep running the company during the review. China’s Embassy called the allegations “groundless smears and slanders”; no one has been charged.

Independent directors and outside counsel will examine what Nvidia technology reached China, if any, and through what routes, including Megaspeed’s roughly $2 billion chip purchases.

Researchers at MIT and the University of British Columbia built WorldParticle, a single AI system that simulates six physical phenomena—including sliding sand, deformable objects, fluids, and proteins—with higher accuracy than numerical simulation methods.

Instead of hand-coding a solver for each material, WorldParticle represents objects as thousands or millions of three-dimensional particles. A neural network predicts each particle’s next position and physical properties, then reintegrates external forces such as an ocean wave striking a seawall.

The architecture can model coupled systems that typically require separate simulators, such as sloshing oil interacting with an elastic valve. The researchers will present the work at ACM SIGGRAPH Asia, calling it a first step toward a foundation model for simulation.

MIT and NYU researchers built FLEX, a robot-learning pipeline that infers broader safety preferences from conversational commands. In a simulated coffee-serving task, “Don’t spill on the laptop” made a robot avoid a nearby phone too, extending the warning from one device to electronics generally.

FLEX’s three stages — “Add,” “Generalize,” and “Mask” — update the robot’s motion planning in about five seconds. GPT-4o powers Generalize, while the system avoids forcing users to name every object they want protected.

Andreea Bobu led the study with Audrey Lee and Ilia Sucholutsky. The team tested FLEX in simulations and will present the work at the annual Conference on Robot Learning this month.

Google’s proposed Thurrock data center faces an alliance of wildlife advocates, environmental groups, and motorcycle racers. Buglife and local wildlife groups say construction threatens an endangered plant, a rare bumblebee, and other invertebrates; Foxglove cites pollution. The Thurrock Hammers want to return to racing at the former Arena Essex site.

Only one of 1,000 public responses supported the project. Protesters say it could destroy habitat and produce up to 570,000 tonnes of emissions annually. Google says it will retain habitat, transplant endangered plants elsewhere, deliver a net biodiversity gain, and fund renewable energy to offset emissions.

Thurrock Council expects to rule early next year. A further, more detailed planning process would follow.

AI auditors can catch most agent misbehavior—but only by flagging harmless logs as suspicious. Anthropic Institute researcher Andy Hall tested GPT-5, GPT-5 nano, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash against METR's MALT dataset.

At a 5% false-alarm rate, GPT-5 detected nearly 98% of reward-hacking cases, while Gemini 3.8 Flash found barely half. At an 80% detection rate, the best models still falsely flagged 2.5% to 3% of clean logs. In 100,000 sessions, that means nearly 3,000 false warnings against roughly 2,000 real hits—enough to numb human reviewers.

NVIDIA is expanding work with South Korean companies across semiconductor engineering, autonomous driving and factory robotics, applying accelerated computing, AI and simulation to physical AI in production.

Samsung’s AI homogenization workflow runs roughly 1,600x faster: a solve dropped from about 18 hours on a 32-core CPU to 41 seconds on a GPU-accelerated coarse-grid system. Selected quantum-chemistry calculations on Samsung’s sMOL platform run up to 5x faster end to end with NVIDIA cuEST.

Kakao Mobility plans to combine its driving software with NVIDIA Hyperion and DRIVE AGX Thor for future level 4 services in Kia purpose-built vehicles. NVIDIA and LG demonstrated CLOiD robot manipulation using real-world data, Cosmos-based augmentation, GR00T training, Jetson AGX Thor and simulation evaluation.

NVIDIA is releasing Dynamo-optimized deployment recipes for four Korean open models—A.X K2 from SK Telecom, K-EXAONE 2.0 from LG AI Research, Solar Open 2 from Upstage and Motif-3 from Motif Technologies. Optimized for NVIDIA B200 GPUs, the recipes deliver up to 2x performance out of the box.

Models larger than one GPU must be split across multiple GPUs, which requires coordinated request routing and communication. NVIDIA’s open-source Dynamo framework handles that work in multi-node, data-center-scale environments. The recipes remove deployment guesswork and help developers bring the four models into production faster.