Multimodal/video/physical AI, World Models racing (papers/datasets gaps)
Key Questions
What multimodal advancements feature Gemma 4 and Gemini?
Gemma 4 and advanced Gemini models lead multimodal AI, processing video and physical data effectively. They bridge gaps in world models.
What did NVIDIA announce at GTC for robotics?
NVIDIA unveiled GR00T and Isaac for sim2real robotics, highlighting physical AI breakthroughs. These enable agentic robotics like Claw.
What is Netflix's VOID?
Netflix released VOID, its first public model on Hugging Face, focusing on video understanding. It contributes to multimodal datasets.
What are MMaDA-VLA, CaP-X, and UniDriveVLA?
These are vision-language-action models like MMaDA-VLA for unified multimodal instructions and generation. They advance video/physical AI.
What evals address video gaps?
Video-MME-v2 pushes comprehensive video understanding benchmarks. It reveals current dataset limitations.
What are ERNIE and Qwen VL?
ERNIE and Qwen VL enhance multimodal capabilities in video and physical domains. They compete with Veo and Sora.
What is Free-Range Gaussians?
Free-Range Gaussians introduce training-free 3D representations for world models. It innovates in dynamic scene modeling.
What is CMU KAAI?
CMU's KAAI applies multimodal AI to astronomy, showcasing physical AI in niche domains. It highlights racing world model progress.
Gemma 4/Gemini multimodal; NVIDIA GTC GR00T/Isaac sim2real; Netflix VOID; MMaDA-VLA/CaP-X/UniDriveVLA/Action Images; ERNIE/Qwen VL; Veo/Sora; MMEmb-R1 embeds; Free-Range Gaussians 3D; Video-MME-v2 evals; CMU KAAI astronomy.