AI news Nov 12: NVIDIA leadership change and Gemini 4 Hydrogen

NVIDIA CEO Jensen Huang has resigned immediately, while CFO Colette Kress takes over as interim chief executive. Google is expanding Gemini 4 Hydrogen across Search and its developer and business products, making model deployment and agentic work the day’s other major theme.
- Jensen Huang resigns immediately, with Colette Kress as interim NVIDIA CEO.
- Google brings Gemini 4 Hydrogen to Search globally and developer products.
- Apple researchers define theoretical limits for language model alignment.
- Red Hat adds live and autonomous OpenShift troubleshooting.
- AI infrastructure expands from Dublin and Kazakhstan to Vera Rubin image generation.
Leadership and Gemini 4 Hydrogen
NVIDIA: Jensen Huang has resigned as NVIDIA’s president, chief executive officer and board member. CFO Colette Kress is interim CEO, the board has begun searching for a permanent successor, and Harvey C. Jones is its first-ever chair.
Google: Google DeepMind introduced Gemini 4 Hydrogen for coding, agentic workflows and consumer tasks. It is available through Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. Google is also bringing Hydrogen to Search globally, with complex-question handling, simultaneous queries, structured responses and more visual results. Vercel has added the model to its AI Gateway.
Research, safety and healthcare
Apple researchers published a NeurIPS paper on theoretical limits for language model alignment. The work examines maximum reward improvement under a fixed KL-divergence budget and finds best-of-N alignment comes close to the limit, while PPO and GRPO remain substantially suboptimal.
DIAGENS introduced a CT model targeting 509 diseases and imaging findings. MIT researchers developed a faster training method for AI climate emulators. Google reported Agent Anomaly Detection results, while OpenAI committed $2 million to AI safety research for children. Other research updates include PyTorch and TorchTitan training a 1T-parameter MoE and Liquid AI’s LiquidCodec.
Infrastructure and enterprise agents
Red Hat OpenShift Lightspeed 2 adds live and autonomous troubleshooting with read-write tools and approval before cluster changes. AWS upgraded AgentCore runtime with faster starts, more compute and stronger isolation, and launched a direct Web Search API. AWS also added silent data corruption detection to SageMaker HyperPod.
vLLM reported subsecond 4K image generation with MiniMax H3-Image on Vera Rubin. Supermicro is shipping H15 servers with AMD EPYC 9006 processors. Armada and Johnson Controls began sovereign AI production in Dublin, while Gcore and TengriAI are exploring a 100 MW Kazakhstan data center.
Other product news includes Replit Agent business purchases and Google Ads, Samsung’s December 10 Gemini-powered eyewear launch, Pinecone’s Gemini Embedding 2 integration, Firecrawl MCP v3, Mastra private model gateways, LiveKit’s Cloud API, Harvey’s contract history update, Factory’s Cyber Defense Network, PHP2Shell exploitation reported by Check Point, and GIGABYTE’s AI TOP AERO desktop.
Defense, policy and applied AI
UFORCE raised $700 million at a $6 billion valuation to expand autonomous defense systems. Shield and NIPPI are building local V-BAT support in Japan, and TIER IV targets driverless dump trucks at a Japanese port by 2028.
Researchers urged stronger oversight in Australia’s Digital Duty of Care inquiry. Other developments include Scale and Singapore’s legal reasoning benchmark, Resemble AI’s deepfake watchlist, Neo4j’s Relate hackathon, LanceDB’s robot and LLM pipeline updates, and BioMap’s 11 collaborations.

