The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Meta Engineering describes building Forum, a standalone Facebook Groups app, with an AI agent running plan, implement, build, test and correct cycles while engineers set goals and approve. Success was measured as changes that build, pass tests and survive review, not code volume.
OpenAI’s Rosalind Academic Plan: up to 30 free Business seats for a year and trusted-access GPT-Rosalind for eligible U.S. university labs.
Meta says Muse Spark Open is the first model in its class that can run on a desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework. Mark Zuckerberg credited work with Apple’s John Ternus, and Alexandr Wang said the companies collaborated for months. Apple published an MLX technical post.
NVIDIA’s serving note: Muse Spark Open trained on GB300 NVL72, NVFP4 for Blackwell, claims up to 50K tok/s per GPU for agentic coding.
NVIDIA’s launch-day invitation: RTX Spark Windows PCs for creators, gamers, and agents — plus #ShareYourSpark and world events.
The Verge argues AI agents like Meta’s Muse extend data collection to contacts who never agreed, citing researchers who found Muse keeps hourly-updated profiles of people in a user’s life. Meta says Muse uses public information and what users choose to share, behind opt-in settings.
NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model for specialized always-on agents, alongside two Bolt models for local and edge use. NVIDIA claims class-leading accuracy and up to 5x throughput, figures that rest on comparisons it selected.
NVIDIA folds vision, documents, video and tool use into one open 120B mixture-of-experts model, betting enterprises would rather post-train a single agent brain than stitch several together.
The Verge details Microsoft’s engineering work for NVIDIA’s Arm-based RTX Spark laptops: a Windows Store flow that installs AI agents in about two clicks, a Prism emulator tuned for the chip, and memory management for models up to 128GB. Recall is missing at launch.
Google made Universal Cart available in the US, letting shoppers save products from multiple retailers across Search and Gemini into one persistent cart with price-drop alerts, promo codes and card-perk suggestions. Checkout runs on Google’s Universal Commerce Protocol.
CopilotKit’s CLiF turns agent traces and user actions into skills, evals, and fine-tuning data — closing the loop agents keep failing in production.
GeForce NOW hands members DLSS 4.5 controls, letting cloud gamers tune AI upscaling and frame generation without owning the latest hardware.