AI Search begins with open media tools. Netflix's IDV2V transfers a keyframe edit across a video while preserving identity and motion, CrisperWhisper 2 offers verbatim or polished transcription with word-level timing, and Redesign turns a flat image into editable layers by orchestrating OCR, generation and segmentation models.
DeepSeek V4 Flash is presented as a major efficiency gain, approaching much larger open and proprietary models on coding and cybersecurity tests at a much lower token price. The episode also covers the release of Kimi K3 weights and AMD's Instella mixture-of-experts model, which was trained on AMD hardware with public checkpoints, recipes and code.
Video generation advances on two fronts. Seedance 2.5 emphasizes action, physics, character consistency, long clips and large reference sets, while MiniMax H3 offers 2K multimodal generation at lower cost and an announced open-model release. The episode also surveys interactive video world models that reason about navigation and physical change.
Google's Gemini Robotics 2 family extends control from perception and planning to walking, balancing, grasping and dexterous hand movements. Separate reasoning and on-device variants support environmental planning, failure recovery, multi-robot coordination and offline execution, while other research improves robot control by combining force, contact, velocity and joint signals.
Watch on YouTube
