Self-evolving AI, robot fights, new GPT voice, new local image model, Gemma upgrade: AI NEWS

Key Concepts

• Multi-token Prediction: A technique where models predict multiple future tokens simultaneously to bypass memory bottlenecks.
• Speculative Decoding: Using a small "drafter" model to predict tokens, which a larger model then verifies, significantly increasing inference speed.
• Intrinsic Video Properties: The ability of AI to decompose video into albedo (base color), irradiance (lighting), and surface normals for precise editing.
• Sparse Computation (Tilewise LPAC): A method to skip zero-value calculations in transformers to improve efficiency and energy consumption.
• Marovian RSA: A reasoning method that samples multiple reasoning attempts and iterates on the best segments to improve output quality without increasing context window size.
• Physics-Grounded Generation: Creating 3D assets that include kinematic parameters, joints, and material properties for use in simulations and robotics.

---

1. AI for 3D Reconstruction and Physics

• RecGen: Reconstructs 3D objects from RGBD (color + depth) images. It excels at handling occlusions (blocked objects) by training on 200,000 high-quality 3D assets and 3 million synthetic images. It outperforms SAM 3D in pose estimation and shape generation.
• Fizz Forge: A two-stage system that acts as a "physical architect." It first generates a physical blueprint (joints, mass, constraints) and then uses a diffusion model to create a 3D asset that is physically functional for robotics and simulations.

2. Image and Video Generation

• Hunyuan-01 Image (Vivago AI): A top-tier open-source image generator capable of 2K resolution. It eliminates the traditional VAE (Variational Autoencoder) to process raw pixels directly. It is highly effective at text rendering and complex infographic generation.
• UniVid X: A versatile video generator that extracts intrinsic properties (albedo, normal, irradiance, alpha channels). This allows for precise video editing, such as relighting scenes or replacing backgrounds/characters.
• Swift I2V: An image-to-video model that generates 2K, 81-frame videos efficiently. It uses "conditional segment-wise generation" to split video processing into smaller time segments, allowing it to run on a single RTX 4090.
• Bach 1.0 (Video Rebirth): A new video generator capable of 30-second, 1080p videos with native sound, showing strong character consistency.
• CDM (Continuous Time Distribution Matching): An acceleration method from Alibaba that reduces diffusion model inference to just four steps while maintaining high quality, outperforming DMD2.

3. Large Language Models and Efficiency

• Gemma 4 (Google): Updated with multi-token prediction, achieving up to 3.1x speed improvements by allowing the model to guess ahead, reducing GPU idle time.
• Zia 18B (Zyra): The first model trained on an AMD Instinct stack. It uses "compressed convolutional attention" and "Marovian RSA" to achieve performance comparable to models 40–80x its size.
• Sakana AI & Nvidia Collaboration: Developed a sparse format (Tilewise LPAC) and custom CUDA kernels to skip wasted computations in transformers, resulting in 30% faster inference and 20% lower memory usage.

4. Robotics and Scientific AI

• Momo Act 2 (Allen AI): An open-source robotics foundation model trained on 700 hours of bimanual data. It features a 180ms action latency (down from 6,700ms in v1) and outperforms Nvidia’s Groot in zero-shot tests.
• Genesis 26.5: A foundation model for robotic dexterity, enabling complex tasks like cracking eggs, playing piano, and solving Rubik’s cubes.
• Lab OS: An AI co-scientist that integrates with XR smart glasses to guide researchers through physical lab protocols, providing real-time feedback to prevent errors.
• Humanoid Demos: Boston Dynamics’ electric Atlas demonstrated non-humanoid range of motion (e.g., 180-degree torso rotation), while Unitree G1 and Engine AI robots engaged in a "fighting" demonstration.

5. Benchmarks and Research

• Program Bench: A stress test requiring AI to rebuild entire software programs (e.g., FFmpeg, SQLite) from scratch using only the executable and documentation. Current top models (GPT-4, Gemini 1.5 Pro) scored 0%, highlighting the gap between "coding" and "software architecture."
• Alpha Evolve (Google): A system that uses AI to discover better algorithms. It has achieved significant real-world impacts, including a 30% reduction in DNA sequencing errors and 10x lower error rates in quantum circuits.

6. Real-Time Voice

• OpenAI Real-time Models: A new suite including:
  • GPT Realtime 2: Enhanced conversational reasoning.
  • GPT Realtime Translate: Supports 70+ input and 13 output languages with low-latency, natural dialogue capabilities.
  • GPT Realtime Whisper: Real-time transcription for captions and notes.

---

Synthesis

The current AI landscape is shifting from simple generative tasks toward physical grounding and architectural efficiency. The industry is moving beyond "scaling compute" to "optimizing execution" (e.g., multi-token prediction, sparse computation, and AMD-optimized training). Furthermore, the emergence of AI agents that can interact with the physical world—whether through robotic hands, lab-integrated XR glasses, or physics-aware 3D asset generation—marks a transition toward AI that functions as an active participant in scientific and creative workflows rather than just a passive text generator.