1. OpenAI Buys Astral, Maker of Python's uv and Ruff Charlie Marsh built Astral into the company behind three tools that Python developers reach for every day. Ruff, the linter. uv, the package manager. ty, the type checker.

2. Meta's AI Agent Broke Data Access Controls for Nearly Two Hours Last week, an AI agent inside Meta gave an employee inaccurate technical advice.

3. AI Labs Trade Open Research for Corporate Moats Two moves last week from rival AI companies pointed in the same direction. OpenAI staged an internal "focus" reset aimed squarely at IPO readiness.

In Brief

• Adobe Opens Firefly Custom Models in Public Beta Adobe now lets creators and brands train Firefly image generators on their own assets. The tool produces images matching a specific artistic style for characters, illustrations, and photography. Custom Models entered public beta today.

• Base Models Beat Aligned LLMs 10-to-1 at Predicting Human Decisions Researchers compared 120 base-aligned model pairs on over 10,000 real human decisions in strategic games including bargaining, persuasion, and negotiation. Base models outperformed aligned counterparts in predicting actual human choices by nearly 10:1. The gap held across model families and prompt formats.

• MetaClaw Builds LLM Agents That Self-Update During Deployment Most deployed LLM agents stay static after launch, even as user needs shift. MetaClaw introduces a meta-learning framework that distills knowledge from task trajectories and updates agent skills continuously. The system handles workloads across 20+ channels on the OpenClaw platform.

• Kinema4D Simulates Robot Interactions as 4D Spatiotemporal Events A new framework models robot-world interactions in four-dimensional space-time rather than 2D video. Prior approaches relied on static environmental cues or flat projections. Kinema4D targets precise interactive simulation for embodied AI.

• MosaicMem Gives Video Diffusion Models Hybrid 3D Spatial Memory Video diffusion models lose consistency under camera motion and scene revisits. MosaicMem lifts patches into 3D for static scene reprojection and uses implicit memory for moving objects. The hybrid approach fixes a core bottleneck in world-simulator video generation.

• SocialOmni Benchmarks Multimodal Models on Live Social Cues Existing benchmarks for omni-modal LLMs test static accuracy, not conversational ability. SocialOmni evaluates social interactivity — reading dynamic audio-visual cues in natural dialogue — across three dimensions.

• Video-CoE Exposes Multimodal LLMs' Weak Spot in Event Prediction A systematic evaluation shows leading multimodal LLMs perform poorly at predicting what happens next in video. Video-CoE applies chain-of-events reasoning to improve fine-grained temporal modeling and logical event sequencing.

• WiT Untangles Pixel-Space Trajectories in Flow Matching Models Flow matching models working directly in pixel space produce tangled transport paths at trajectory intersections. Waypoint Diffusion Transformers insert intermediate waypoints to separate these paths without compressing into a latent space.

Read the full edition →