We read the whole AI world for you

NVIDIA and MIT's Efficient-Large-Model team released weights on Hugging Face with code in NVlabs/Sana, training only a refinement LoRA on the frozen backbone. On one B200 a 5-second clip took 22.45 seconds versus 87.26 unoptimized, decoding dropped from 15 to 1.6 seconds, and UniPercept scored 60.0 against the official API's 53.4. Teams that self-host get the 2K stage as deployable code with quality beating the official API in the team's own tests. Every figure is vendor-reported, so verify the weights and license before building on it.

The October 29 update names faithfulness to human intent as the foundation and adds a chain-of-command section: OpenAI writes the Spec and sets boundaries, developers direct applications inside them, and users direct the model's work. Two defaults were added, including supporting human agency, and side-effects guidance was expanded. Developers get firm boundaries plus mostly overridable defaults, since most style rules were retired and the model is now called GPT rather than the assistant. Preferences may be reused, but assumptions must stay visible.

Google began rolling out in-chat generation, preview and editing of files in Gemini on October 29, covering PDFs, Docs, Sheets, slide decks, Markdown, TXT and code. Files expand to full-screen editing, and you can highlight a table and ask it to become a graph; PDF editing works on web and mobile, every other format web-only. Users skip the download-and-edit round trip, and the feature is live for all Gemini accounts today. Edit fidelity is unproven, so test whether formatting and formulas survive several edit rounds on your own documents first.

An October 29 post details a pipeline tuned for recall up front and judgment at the back: L0 activation probes, an L1 reviewer checking 30 minutes apart, L2 investigators, and L3 cross-referencing. Dated cases include a September 2026 breach flagged within 15 minutes and a Hugging Face replay that beat the real breach by over 30 hours. Teams running long-horizon agents can borrow this shape as an ops template. Note the stated limits: external monitoring does not escalate flagged cases to humans and does not cover Chat Completions requests.

At the October 29 Tencent Global Digital Ecosystem Summit, the company said Hy4 strengthens long-horizon and real-world productivity tasks, improves chain-of-thought efficiency, and adds vision understanding for the first time. It builds on the Hy4 preview open-sourced August 28: 770B total parameters, 49B active, over 1M context. Developers weighing a self-hosted Chinese stack get a direction signal, not a product: no release date, pricing, API plan or benchmark scores yet. Watch at launch for evaluation data, open weights, and real-task vision quality.