Overview
- Bloomberg reporting says Meta is developing a closed, revenue-oriented model codenamed Avocado, with training that reportedly distills capabilities from Google’s Gemma, OpenAI’s gpt-oss, and Alibaba’s Qwen, targeting a spring debut.
- Google launched hosted Model Context Protocol servers in public preview to let AI agents directly use Maps, BigQuery, Compute Engine, and Kubernetes Engine, with enterprise access free for paying customers and general availability planned early next year.
- OpenAI warned that next‑generation models could pose high-level cybersecurity risks and outlined mitigations including stricter access controls, infrastructure hardening, monitoring, export controls, and a new Frontier Risk Council focused initially on cyber defense.
- SpaceX and startup Starcloud placed NVIDIA H100 GPUs in orbit and reported in‑space model runs, including Google’s Gemma and training with NanoGPT, as Starcloud pursues a solar-powered orbital data center.
- China’s Meituan advanced its LongCat multimodal stack through new releases across language, vision, audio, and video, and hired former ByteDance and Google researcher Pan Xin to lead multimodal innovation and app development.