Robot waifus, RIP Sora, GLM-5.1, AI brain scans, Google realtime voice: AI NEWS

Key Concepts

• Agentic AI: Systems capable of autonomous decision-making, planning, and executing multi-step tasks (e.g., coding, UI design).
• Gaussian Splatting: A technique for representing 3D scenes that allows for high-quality, real-time rendering and view synthesis.
• Optical Flow: The pattern of apparent motion of objects, surfaces, and edges in a visual scene caused by the relative motion between an observer and a scene.
• VRAM Optimization: Techniques (like Dynamic VRAM) to manage GPU memory more efficiently, allowing larger models to run on consumer hardware.
• Multimodal Models: AI systems capable of processing and generating multiple types of data, such as text, audio, video, and brain activity.
• Quantization: The process of reducing the precision of a model's weights to decrease its size and memory footprint.

---

1. Image and Video Restoration & Generation

• Real-ESRGAN: An open-source model for restoring damaged, blurry, or noisy images. It performs on par with closed-source models like GPT-4 Image 1.5 and outperforms Qwen Image Edit. It requires a high-end GPU due to its 42 GB size.
• Matrix Game 3.0 (Skywork AI): A 5-billion parameter model that generates interactive, real-time 3D worlds (720p at 40 FPS). It uses "long-term memory" to maintain consistency when the camera moves or looks away.
• DaVinci Magi Human: A 15-billion parameter model that generates video with native audio. It claims a 60% human preference win rate over LTEX 2.3.
• Lumos X: A video generation tool focused on consistency for multiple subjects/objects using "relational self-attention" and "relational cross-attention" blocks to link specific references to video segments.

2. Audio and Motion Synthesis

• Prism Audio: A 518-million parameter model that generates realistic, perfectly synced sound effects for silent videos. It is significantly smaller and faster than competitors like MM Audio.
• Action Plan: A framework for generating human motion from text prompts. It is "future-aware," meaning it plans frames ahead to ensure smooth motion. It has been successfully linked to Unitree G1 humanoid robots for real-time command execution.
• Pulse of Motion: A tool that corrects "chronometric hallucination" in AI videos by recovering the true physical frame rate of motion, making movements look more natural.

3. 3D Reconstruction and Spatial Understanding

• Retime GS: Uses 4D Gaussian splatting and optical flow to reconstruct smooth 3D animations from choppy 2D video, effectively filling in missing frames.
• World Agents: An agentic system (Director, Generator, Verifier) that uses 2D image models to build consistent 3D worlds without additional training.
• World Reconstruction from Inconsistent Views: A model that stitches multiple AI-generated videos into a single, consistent 3D scene. It is model-agnostic and based on ByteDance’s Depth-Anything 3.
• Logger NVS: Generates new camera angles and 360° environments from a handful of sparse images by learning the relationship between viewpoints.

4. Neuroscience and Robotics

• Tribe V2 (Meta): A model that predicts human brain activity (fMRI simulations) based on visual input. It acts as a "digital twin of human perception" and is trained on 700+ people's data.
• Origin F1: A hyper-realistic humanoid robot head featuring 25–30 micro-actuators for natural facial expressions, eye contact, and lip-syncing, powered by the Omni AI system.

5. Benchmarks and Efficiency

• ARC-AGI 3: A benchmark testing an AI's ability to learn new rules on the fly. Current frontier models (Gemini 3.1, Opus 4.6) score <0.5%, while humans score 100%, highlighting a major gap in real-time adaptation.
• Turbo Quant (Google Research): A compression technique using "Polar Quant" and "KGL algorithms" to shrink AI models by 6x while speeding up data retrieval by 8x.
• ComfyUI Update: Introduced "Dynamic VRAM," which loads/unloads model parts during generation to prevent out-of-memory errors and increase speed on mid-range GPUs.

6. Notable Industry Updates

• OpenAI Sora: OpenAI is shutting down the Sora app/website to reallocate compute resources toward robotics and enterprise-grade coding agents, citing high costs and safety/copyright concerns.
• ZAI GLM 5.1: A new model optimized for agentic coding, offering performance near Opus 4.6 but with significantly higher speed and lower costs.
• Co-here Transcribe: A 2-billion parameter, Apache 2.0 licensed transcription tool that outperforms Whisper and Canary in accuracy and efficiency.
• Gemini 3.1 Flash Live: A real-time voice model featuring low-latency, natural conversational rhythm, and the ability to execute UI design tasks via Google Stitch.

---

Synthesis

The current AI landscape is shifting from simple content generation toward physical-world interaction, real-time adaptation, and extreme efficiency. The industry is moving away from consumer-facing "hype" tools (like the Sora app) toward robust, agentic systems capable of controlling hardware (robotics) and complex software workflows. The emergence of benchmarks like ARC-AGI 3 underscores that while AI is becoming more powerful, it still lacks the fundamental human ability to learn new, unfamiliar environments on the fly.