1. Three AI Privacy Failures in One Week Reveal the Same Regulatory Void A Swedish journalism team wanted to know what happens to footage captured by Meta's Ray-Ban smart glasses. They found out: some of it ends up on screens in Nairobi.

2. Anthropic Patched Firefox for Hundreds of Millions of Users. The Pentagon Calls It a Security Threat Anthropic's red team spent two weeks in February scanning nearly 6,000 C++ files in Firefox's codebase. Claude Opus 4.6 found its first vulnerability in twenty minutes.

3. A 60-Year-Old Developer Picked Up Claude Code and Couldn't Stop Building A Hacker News user posted last week that at 60, Claude Code had re-ignited a passion for programming.

In Brief

• Netflix Acquires Ben Affleck's AI Production Startup InterPositive Netflix bought InterPositive, Ben Affleck's AI company that builds tools for film and television production. All 16 of InterPositive's engineers will join Netflix.

• Apple Music Launches Voluntary AI Transparency Tags for Songs and Artwork Apple now asks artists and record labels on Apple Music to tag content made with AI. The "Transparency Tags" metadata system covers four categories: track, composition, artwork, and music videos. Tagging is voluntary.

• Google Open-Sources SpeciesNet Model for Wildlife Conservation Google released SpeciesNet, an open-source AI model that helps conservation researchers identify animal species from camera trap images worldwide.

• Researchers Propose Interactive Benchmarks to Replace Saturated Static Tests A new paper argues standard AI benchmarks fail due to saturation, subjectivity, and poor generalization. The proposed framework evaluates models on their ability to actively acquire information under budget constraints across two settings: interactive proofs and interactive retrieval.

• DreamWorld Framework Jointly Models Physics and Geometry for Video Generation A new paper presents DreamWorld, a video generation system that unifies multiple forms of world knowledge—physics, geometry, semantics—into a single model. Prior approaches incorporated only one type of world knowledge or relied on rigid alignment strategies.

• DARE Retrieval Model Helps LLM Agents Pick the Right R Statistical Tools Researchers released DARE, a retrieval model that matches LLM agents with appropriate R statistical functions by incorporating data distribution information. Standard retrieval methods ignore distribution characteristics, leading to poor tool selection for statistical workflows.

• Google Explains Query Fan-Out Method Behind AI-Powered Visual Search Google published a technical breakdown of how AI Mode in Search processes visual queries. The system splits image-based searches into multiple sub-queries to assemble answers from different sources.

• OpenClaw Holds First Open-Source AI Conference in Manhattan OpenClaw, an open-source AI community, hosted ClawCon in New York City. The multi-floor event drew developers and sponsors under lobster-themed branding.

Read the full edition →