1. Meta's Smart Glasses Send Intimate Footage to Human Reviewers in Kenya When you activate Meta's AI features on your Ray-Ban smart glasses, the camera feed doesn't just go to a model. Some of it goes to a desk in Nairobi.

2. Google and Netflix Tie Their Biggest Financial Bets to AI-Adjacent Units Two corporate disclosures from the same week expose how tech and media incumbents now price AI.

3. Anthropic and OpenAI Race to Lock In Open Source Maintainers With Free Subscriptions Anthropic announced on February 27 that maintainers of open source projects with 5,000+ GitHub stars or one million NPM downloads would receive six months of Claude Max free.

In Brief

• Anthropic's Red Team Helps Harden Firefox Security Anthropic partnered with Mozilla to apply AI-powered red teaming to identify security vulnerabilities in Firefox.

• Apple Music Launches Voluntary AI Transparency Tags for Artists Apple now asks artists and record labels to tag songs made with AI. The metadata system covers four categories: track, composition, artwork, and music videos.

• Google Open-Sources SpeciesNet Model for Wildlife Conservation Google released SpeciesNet, an open-source AI model that helps researchers identify and track wildlife from camera trap images. Conservation teams worldwide already use it to monitor species populations.

• Balyasny Asset Management Builds AI Research Engine for Investment Analysis Hedge fund Balyasny built an AI-powered research system using agent workflows to accelerate investment analysis at scale. The system includes structured model evaluation to vet outputs before they reach analysts.

• Researchers Propose Interactive Benchmarks to Replace Saturated Static Tests A new paper argues static AI benchmarks are unreliable due to saturation and poor generalization. Interactive Benchmarks instead measure a model's ability to actively acquire information under budget constraints.

• DARE Retrieval Model Aligns LLM Agents with R's Statistical Ecosystem LLM agents underuse R's statistical methods because they struggle with tool retrieval and statistical knowledge. DARE incorporates data distribution information to match agents with appropriate R functions as a plug-and-play module.

• DreamWorld Proposes Unified World Modeling for Video Generation Current video generation models lack coherent understanding of physics and spatial relationships. DreamWorld jointly models multiple dimensions of world knowledge instead of incorporating them one at a time.

• HiFi-Inpaint Improves Product Detail Preservation in AI-Generated E-Commerce Images Reference-based inpainting for e-commerce images often loses fine product details. HiFi-Inpaint tackles fidelity, dataset diversity, and alignment to better preserve product specifics in generated human-product images.

Read the full edition →