The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
At Gemini at Work, Google announced a universal agent in the Gemini Enterprise app that runs in the cloud, works across Workspace, Slack and Microsoft 365, calls job-specific sub-agents, picks models automatically and can act as a coworker with its own agents.company.com address. It is in private preview for enterprises.
TechCrunch reports Google released AI Edge Foresight, an offline meeting note-taker that transcribes conversations, generates notes and answers questions using on-device AI, challenging Granola and other cloud-based assistants on privacy.
In an OpenAI customer story, Pollo AI, which says it has more than 26 million users, describes Pollo Agent, built on GPT-5.6, GPT-6 Astra and GPT-Image-2.5, which turns rough ideas into storylines, scripts, images and video for small brands.
Ars Technica details Nvidia’s Halos, a full-stack safety system that expanded from self-driving cars to warehouse robots, factory humanoids and surgical robots. Agility’s Digit 5 is the first humanoid to build it in, but robot safety lacks a single definition, and Nvidia also invests in the companies it accredits.
An MIT Technology Review piece produced with Aventine contrasts sweeping humanoid forecasts with lab reality: vision-language-action models can pack lunches and fold origami but fail outside their training data, and researchers say reliable home robots may be a decade off.
Architect wants LLM inference priced like a market, with providers bidding on every prompt. It is a sharp idea for cost, with real questions about quality and data.
NVIDIA-led research finds most failed agent runs hinge on one early wrong turn, and shows agents can be trained to climb back out. That is the reliability problem worth solving.
In a sponsored MIT Technology Review Insights podcast made with AVEVA, chief technologist Arti Garg argues that foundation models, physical AI and agents bring a riskier phase to plants, grids and mines. Her governance approach is sound; the headline results are vendor-reported.
Perplexity's MIT-licensed retrieval models let a laptop query an index built by a much bigger model, a smart split for document search, with the caveat that every score is self-reported.
Unsloth Studio stops treating a trusted name as a trusted file, re-checking model code every time it loads. That is the right instinct for local AI, as long as nobody mistakes a scanner for a sandbox.
Microsoft unveiled the $2,599 Surface Laptop Ultra on Nvidia’s RTX Spark, a $5,999 dev box, a reorganized Copilot with Code and Autopilot tabs, action-taking Windows Search and Execution Containers, a framework to confine agents to sandboxes.
OpenAI says it banned accounts tied to a Russian operation it calls Dark Clark, rated Category 5 on the IO Breakout Scale, and an Iranian one, Bogus Bylines, whose fake journalist personas placed almost 100 articles. Most of the AI use was polish and paperwork rather than the propaganda itself.