1. Facial Recognition Jailed an Innocent Grandmother. Google's Health AI Asks a Doctor First. A grandmother in North Dakota spent months behind bars after a facial recognition system matched her face to a fraud suspect.

2. Google, Microsoft, and Meta Turn AI Assistants Into Transaction Agents in One Week Tell your phone to order a coffee. Gemini opens DoorDash, picks your usual, and checks out in a virtual window you never touch.

3. Musk's xAI Retains Two of Twelve Original Co-Founders Guodong Zhang led xAI's Imagine team until last week. Then Musk blamed him for problems with the company's coding product, stripped his responsibilities, and Zhang told colleagues he was done.

In Brief

• Anthropic's Claude Now Generates Charts and Diagrams Inline Claude can produce custom visualizations — charts, diagrams, and other graphics — directly within a conversation. The AI decides when a visual is useful based on context and inserts it inline rather than in a side panel. The Verge

• Netflix Commissions Custom AI Models for Film Production Studios are moving past off-the-shelf generators toward bespoke AI models trained on specific visual styles. Netflix's "Interpositive" project, involving Ben Affleck, uses a purpose-built model rather than general tools like Sora or Runway. The approach treats AI as a per-project tool, not a replacement for production pipelines. The Verge

• "Can I Run AI Locally?" Tool Checks Hardware Against Model Requirements A new web tool lets users input their hardware specs and see which open-weight AI models their machine can run. The site covers popular models across parameter sizes and quantization levels. canirun.ai

• Researchers Propose Video-Based Reward Modeling for Computer-Use Agents A new paper introduces reward modeling from execution video — using keyframe sequences from agent trajectories to evaluate task completion. The method is agent-agnostic, judging results from screen recordings rather than internal reasoning traces. It addresses the scaling bottleneck in evaluating whether computer-use agents actually follow instructions. Hugging Face Papers

• IndexCache Cuts Sparse Attention Overhead by Reusing Cross-Layer Indices Researchers target the indexer bottleneck in sparse attention systems like DeepSeek Sparse Attention, where the token-selection step still runs at O(L²). IndexCache reuses index computations across transformer layers to reduce this cost. The optimization matters most for long-context agentic workloads where attention efficiency drives serving cost. Hugging Face Papers

• Spatial-TTT Brings Streaming Spatial Understanding to Vision Models via Test-Time Training A new method maintains and updates spatial representations from unbounded video streams using test-time training. Spatial-TTT selects and retains spatial evidence over time rather than relying on longer context windows. The approach targets real-world applications where an agent must continuously process visual input. Hugging Face Papers

• MADQA Benchmark Tests Whether Document Agents Reason or Just Search Randomly Researchers built a 2,250-question benchmark grounded in 800 heterogeneous PDFs to measure whether multimodal agents use genuine strategy when navigating document collections. The benchmark applies Classical Test Theory to maximize discrimination across agent skill levels. Early results suggest many agents rely on trial-and-error rather than structured reasoning. Hugging Face Papers

• GOLF Framework Uses Natural Language Feedback to Guide RL Exploration A new reinforcement learning framework aggregates group-level language feedback — not just scalar rewards — to steer exploration toward actionable improvements. Current RL methods discard the rich information in natural language signals from environment interactions. GOLF converts that feedback into targeted exploration strategies. Hugging Face Papers

Read the full edition →