AI Digest — 2026-11-14
2026-11-14 · 48h window · synthesized from 42 sources / 809 articles · ~2 min read
The most important things happening in AI right now, deduplicated across sources.
1. Jensen Huang resigns as Nvidia CEO. Huang stepped down after reports of an affair with an alleged Chinese spy, with Colette Kress named CEO and Congress opening a second investigation into Nvidia's China sales. Founders and VCs expect smaller checks for AI startups without his backing, and some ask whether Nvidia's open-source push will last. Sources: The Register · TechCrunch · The Robot Report · Wired
2. China's top AI models were trained on export-banned Nvidia chips. A forensic report by an ex-OpenAI engineer finds Blackwell fingerprints in models from DeepSeek, Moonshot AI, MiniMax and Ant Group. Nvidia had privately told a US lawmaker that tracking chips to stop smuggling is technically possible, while its industry group called it infeasible. Sources: The Register · TechCrunch · MIT Technology Review
3. ChatGPT Wallet lets agents buy things for you. OpenAI's new Wallet lets agents in ChatGPT, Codex and Dots complete purchases with the user's approval, starting with US Pro subscribers and major US cards. Agentic commerce is becoming a stack of its own, and Amazon is building a graph of the world's product data for AI search. Sources: OpenAI · The Verge · TechCrunch · The Verge
4. Google ships Gemini 4 Hydrogen and puts it in Search. Gemini 4 Hydrogen replaces 3.8 Flash in a fast, broad rollout, unlike the slower releases of Carbon and Argon. In Search, AI Mode now handles complex questions better, with more relevant links and richer visual answers. Sources: Google DeepMind · Google · Ars Technica · TechCrunch
5. World-action models arrive for robots. Nvidia's GR00T 2 predicts future video and robot actions together so skills can move between machines, and Runway's Praxis-1 is now available with open weights. Rhoda AI takes the same route, training general robot models on video data. Sources: Dev.to · The Robot Report · The Robot Report
6. Local AI hits 1.0: llama.cpp and ggml. llama.cpp and its ggml library reached version 1.0, a stable base for running models on your own hardware. Open-weight models keep pace: Unsloth published new Qwen3.8-27B quantizations, and a retrained Qwen3.8 can now say things Chinese authorities don't like. Sources: Hugging Face · GitHub · Unsloth · The Register
Read the living page: https://jacopocastellano.com/ai-digest · Archive