1. Fidji Simo's Title Changed Twice in Months. Now She's on Leave. Fidji Simo joined OpenAI as CEO of applications. Months later, her title changed to CEO of AGI deployment.

2. Utah Approves AI Psychiatric Prescribing as Study Shows Users Won't Catch AI Errors Utah has authorized an AI system to prescribe and refill psychiatric medications without a physician in the loop.

3. Meta Suspends Data Vendor Mercor as AI's Outsourced Supply Chain Cracks Meta has paused its relationship with Mercor, a data vendor serving multiple leading AI laboratories, after a security incident that may have exposed proprietary details about how the company trains

In Brief

• Cursor Releases Version 3 With Unified Agent Workspace Cursor 3 replaces the split-pane IDE with a single workspace where local and cloud agents share a sidebar. Developers can start an agent session on desktop, hand it off to the cloud, and resume from mobile, web, Slack, or Linear. Cloud agents now return screenshots and visual demos for verification. Cursor Blog

• AMD-Backed Lemonade Ships Open-Source Local LLM Server for GPU and NPU Lemonade is a 2MB C++ server that auto-detects GPU and NPU hardware and exposes an OpenAI-compatible API for chat, vision, image generation, transcription, and speech synthesis. It runs on llama.cpp, ONNX Runtime, ROCm, and Vulkan across Windows, Linux, and macOS. The server can load multiple models at once and needs no cloud connection. Lemonade

• Apfel Unlocks Apple's Built-In 3B-Parameter LLM via Command Line Apfel is an MIT-licensed CLI tool and HTTP server that exposes the on-device language model bundled with macOS Tahoe on Apple Silicon Macs. It offers an OpenAI-compatible API with tool calling, a 4,096-token context window, and support for nine languages. No API keys or subscriptions required — the model runs entirely on the Neural Engine and GPU. The project gained 818 GitHub stars on its first day. Apfel

• Elgato Adds MCP Support to Stream Deck, Letting AI Agents Press Buttons Stream Deck software version 7.4 introduces Model Context Protocol support, allowing Claude, ChatGPT, and Nvidia G-Assist to discover and activate Stream Deck actions. AI assistants can now trigger macros, scene switches, and device controls through the same protocol used for tool integration. The Verge

• Google Home Update Gives Gemini Natural-Language Lighting and Temperature Controls Google Home now lets users describe lighting by analogy — saying "the color of the ocean" sets a matching hue — and adjust thermostats through conversational commands instead of exact values. The update targets more natural phrasing for routine smart-home interactions. The Verge

• Researchers Extract 4M Frames From AAA Games to Train a Generative World Renderer A team from Shanda AI Research, NTU, and University of Tokyo built a dataset of 4 million continuous 720p frames captured from AAA games using a dual-screen stitching method. Each frame includes synchronized RGB and five G-buffer channels covering geometry, materials, and lighting. The dataset supports both inverse rendering (decomposing images into scene properties) and forward rendering (generating video from G-buffers with text-prompt style editing). Hugging Face Papers

• EgoSim Simulates First-Person Interactions While Updating 3D Scene State Researchers from Shanghai Jiao Tong University and Shanghai AI Lab built a closed-loop egocentric world simulator that generates spatially consistent first-person interaction videos and persistently modifies the underlying 3D scene. The system extracts training data from large-scale monocular egocentric videos and supports cross-embodiment transfer to robotic manipulation. Code and datasets will be released. Hugging Face Papers

Read the full edition →