GPT 5.6, Mythos ban lifted, realtime avatars, Seedance 2.5, brain ultrasound: AI NEWS

Key Concepts

• Agentic Frameworks: Systems where AI models autonomously plan, use tools, debug, and manage workflows.
• Multimodal Models: AI capable of processing and generating text, audio, video, and images simultaneously.
• Sub-nanometer Chip Technology: Advanced semiconductor manufacturing (e.g., 0.7nm) using 3D "nano-stack" architectures to increase transistor density.
• Synthetic Data Generation: Using AI to autonomously create, test, and refine training datasets.
• Coupled Oscillators: A non-diffusion-based mathematical architecture for image generation.
• Full-Stack Optimization: The integration of custom hardware (AI-specific chips) with software models to maximize performance-per-watt.

---

1. Video Generation and Interactive Avatars

• One Streamer: A real-time conversational avatar system using a single transformer model to handle text, audio, and video. It operates at 25 FPS with ~200ms latency and supports duplex communication (listening while speaking).
• Domain Shuttle: A video generator that maintains character and object consistency across different art styles (e.g., realistic to anime). It allows for complex interactions, such as placing a 2D character on a 3D object.
• Seed Dance 2.5 (ByteDance): An upcoming upgrade to the leading video model, featuring 30-second clip generation, support for 50+ multimodal references, precise local editing via bounding boxes, and 4K resolution output.
• Happy Horse 1.1 (Alibaba): An update focusing on motion realism and lip-sync, supporting wide aspect ratios and multiple languages.

2. AI Hardware and Semiconductor Breakthroughs

• OpenAI "Jalapeno": A specialized AI processor developed in partnership with Broadcom. Designed for full-stack optimization, it was developed in nine months using OpenAI’s own models to assist in hardware/software design.
• IBM Sub-nanometer Chip: A 0.7nm (7 angstrom) technology using a "nano-stack" 3D architecture. It achieves 100 billion transistors on a fingernail-sized chip, offering 50% higher performance and 70% better energy efficiency than 2nm predecessors.

3. Open-Source Models and Architectures

• Ornith 1.0: A family of agentic coding models (9B to 397B parameters). It uses a unique training method where the model generates its own "scaffold" or workflow, outperforming larger models like DeepSeek V4 on benchmarks like SWEBench.
• Unzero (Unconventional AI): An image generation architecture that replaces traditional diffusion (denoising) with "coupled oscillators," where thousands of tiny units synchronize to form an image.
• Perception DM (ByteDance): A vision-language model using a diffusion-based approach to generate multiple image captions simultaneously, significantly increasing speed over sequential captioning.
• Crea 2: A new open-source image generator noted for being lightweight (runs on 4GB VRAM), highly uncensored, and possessing strong prompt adherence.

4. Robotics and Data Frameworks

• HIW500: A massive dataset containing 500+ hours (10TB) of humanoid robot teleoperation data across 10+ household tasks, designed to train robots for real-world environments.
• Unitree R1: A lightweight, affordable ($4,900) humanoid robot capable of complex acrobatic maneuvers like breakdancing, requiring high joint torque and rapid stabilization.
• Lift 4D: A framework that reconstructs a 4D scene (3D + time) from a single 2D video clip, inferring occluded regions and motion.

5. Regulatory and Industry Perspectives

• GPT 5.6 (Soul, Terra, Luna): A new family of models featuring "Ultra" mode for complex agentic workflows. It is currently restricted to a limited preview for "trusted partners" due to US government compliance, sparking concerns about a "permanent underclass" of users without access.
• Claude Mythos/Fable: Following a period of restricted access, the US government eased bans, allowing Anthropic to release Claude Mythos to a select group of ~100 partners.
• Sakana Fugu: An orchestrator model that routes prompts to various other models. The speaker notes that claims of it "beating" frontier models are misleading, as it is an ensemble method rather than a base model.

6. Scientific Applications

• Olive Brain Imaging: A non-invasive ultrasound technique using "microbubbles" as contrast agents to map blood flow in the brain with 100x the resolution of CT scans. The pipeline is open-sourced for lab use.
• Auto Data (Meta): A framework where an AI agent autonomously generates, tests, and refines its own training data in a loop, mimicking the iterative process of a human data scientist.

---

Synthesis

The AI landscape this week is defined by a shift toward agentic autonomy (models that build their own workflows and data) and full-stack integration (hardware designed specifically for AI models). While proprietary models like GPT 5.6 and Claude Mythos are increasingly restricted by government-mandated "trusted partner" programs, the open-source community is responding with highly performant, specialized models (Ornith, Crea 2) that prioritize user sovereignty. The emergence of non-diffusion architectures and 3D-stacked chip designs suggests that the industry is actively seeking to bypass current physical and computational bottlenecks.