AI Agents Build Simulations Inside Simulations

Claude spent 11 days formalizing Fermat's Last Theorem into 13 million lines of machine-checked code. While humans at Imperial College London worked on a multi-year project, the model quietly wrapped the proof adhering to foundational rules.

The Trump administration filed an amicus brief backing OpenAI against the New York Times. The Justice Department argued model training is fair use, framing web scraping as an engine of national security rather than plain copyright infringement.

Google released Gemini 3.8 Flash and a specialized Cyber variant for automated vulnerability patching. That marks its third Flash release in six weeks, keeping pricing identical while pushing faster multi-step reasoning for agentic tasks.

Fei-Fei Li's World Labs launched Atlas, an omni-modal spatial model that generates 1440p video conditioned directly on camera trajectory, focal length, and speed. It turns video generation into precise 3D scenes for VFX and robotics pipelines.

MBZUAI dropped K2 Horizon, a suite of six open foundation models spanning from 0.9 billion up to 375 billion parameters. The release includes full weights and training code, setting fresh benchmark highs for open lightweight reasoning weights.

Runway unveiled GWM Worlds 2, an autoregressive diffusion model that streams 720p video and audio in real time. It accepts live keyboard controls and text events via WorldPrompt, effectively turning video generators into interactive game engines.

China's Supreme People's Court set formal guidelines on generative AI, ruling unauthorized face-swaps and voice clones infringe personal rights. Providers now face direct liability for deepfakes, algorithmic discrimination, and autonomous car crashes.

Alibaba's Qwen team released Qwen-Drive-1.0-4B under an Apache 2.0 open-source license. The autonomous driving foundation model combines vision-language perception with trajectory planning modules tuned through imitation and reinforcement learning.

Suzhou-based Zeroth launched Bridge, an 88-centimeter AI-native humanoid robot weighing 13 kilograms. Alongside the bipedal hardware, the company unveiled an open developer platform called OpenBridge to create an Android-like ecosystem for robotics.

A Nature Medicine commentary argues clinical AI must move past bench benchmarks to randomized trials measuring actual patient outcomes. Matching human doctor diagnostic scores on static tests means nothing if real-world clinical workflows fail to improve.

Researchers introduced Puffin-World, a multimodal framework tracking native physics, depth, and appearance without external 3D reconstruction. Its 9-channel camera setup grounds synthetic generation in gravity and horizon cues across full scene views.

German robotics firm NEURA and Italy's SECO teamed up to build European compute modules for cognitive humanoids like 4NE1. Using Qualcomm processors, their Smart Limbs setup shifts sensor processing out to physical joints instead of a central board.

The Delhi High Court rejected ANI Media's injunction against OpenAI, ruling that scraping public news for model training qualifies as statutory fair dealing research. The decision removes immediate licensing pressure across India's publishing sector.

South Korea selected LG Electronics and NC AI to develop autonomous humanoid robots powered by domestic on-device AI chips. NC AI will provide human motion retargeting pipelines and world model simulation data to teach robots complex physical chores.

D-Robotics showcased consumer hardware at IFA 2026 powered by its Sunrise AI silicon, including the Vbot SuperDog quadruped and TCL's companion robot. The company is positioning its edge chips as the baseline nervous system for mass-market home robotics.