
27th September - AI News Daily - China Restricts Nvidia Chips as U.S. Rejects Global AI Governance
AI News Daily
Send us a text
🌍 INAI • The Open AI Hub
The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.
https://github.com/inai-sandy/inAI-wiki
AI News Daily — 27 Sept 2025
Infrastructure & Industry: OpenAI, Nvidia, and Oracle announced massive data centers in New Mexico and Texas with 10+ GW capacity. China restricted Nvidia chip purchases, while the US rejected centralized AI governance. Microsoft launched marketplaces for 3,000+ AI apps/agents. Meta is considering Google Gemini for ads as OpenAI builds a ChatGPT ad platform.
New Tools: Google integrated Gemini into Chrome with scam detection and assistant features. Perplexity released a Browsing/Search API for real-time web answers. GitHub Copilot CLI entered public preview. vLLM v1 added hybrid model support and multi-tier caching. Google launched Data Commons MCP Server for public datasets. OpenAI's ChatGPT Pulse provides proactive daily briefs.
LLM Updates: OpenAI's GPT-5 reportedly used less pretraining compute than GPT-4.5. Meta released a 32B open-weight Code World Model. Alibaba's Qwen3 lineup showed strong performance across various domains. Google's Gemini Robotics-ER 1.5 focuses on embodied reasoning. Claude Opus 4.1 claims 95% of human expert performance across 44 jobs. A stealth "code-supernova" hit 1M-token context.
Research & Breakthroughs: Arc Institute and Nvidia created the first AI-generated functional genomes. CATCH-FM uses foundation models to detect high-risk cancer patients. New agent benchmarks include Gaia2, ARE, and LIBERO VLA. MMR1 stabilizes multimodal RL. Hugging Face study showed high energy costs for text-to-video generation. OpenAI's GDPval benchmark rates models on economic utility.
Security Concerns: Major alerts hit Salesforce AgentForce ("ForcedLeak"), MCP servers, and Neon data app. The UK recovered £480M using AI anti-fraud tools.
Demos & Applications: Gemini Live provides real-time cricket commentary. Veo 3 shows improved visual reasoning. New video pipelines combine multiple AI tools for content creation. TinyWorlds reimagines DeepMind's Genie 3 for game environments. Robotics demonstrations show progress in embodied skills.
Learning & Discussion: Richard Sutton advocates for continual learning. Experts debated tokenizer-free approaches. New optimization methods were introduced. Commentators noted that maintaining web indexes can be harder than training LLMs. "Vibecoding" culture shifts development practices. Stanford presented at the UN on equitable AI benefits.
Resources: Free Python data structures course. LlamaIndex guides for full-stack agents. PM guides for user feedback loops. FlashAttention 4 deep dive. Audio generation coursework. Local LLM tips for running Qwen3-coder on Mac.
Support the show