Summary
This weekly AI news roundup covers a packed stretch of releases across models, media tools, and research claims. Headline items include reports that AI may have made progress on the Riemann hypothesis and unforced Navier-Stokes problems, new OpenAI models named GPT 6.2 Astra and Soul, and Google’s Gemini 4 Neon.
The video then moves through a series of open-weight and research releases. LongCat Video Avatar 1.6 turns a photo and audio clip into a talking video that stays consistent for over 10 minutes. A tool called Fab reconstructs a person’s body and surrounding scene in 3D from video and can retarget motion to a Unitree G1 humanoid robot, with model sizes ranging from under 600 MB to 3.8 GB. Apple’s Depth Pro 2 estimates depth for images and video with improved sharpness and temporal consistency.
Other segments cover a Voice Isolator model that removes background noise and other voices, with a real-time version at roughly 63 milliseconds of latency, under 65 million parameters, and an Apache 2 license that permits commercial use. The video also touches on a new image-generation architecture, Luma’s workflow presets, and additional open models, with links to code for each release.
📺 Source: AI Search · Published November 22, 2026
🏷️ Format: Roundup







