All articles
Grok in the room: Trump asked a model about capturing Maduro
Press reports say Trump consulted Grok on how Venezuelans would react to Maduro's capture. An invasion followed a month later, and the model gained advisory standing.
UCLA's optical network spots deepfakes across 15 videos at once
A UCLA team moved the decoding stage of a deepfake detector into optics. A passive diffractive network analyses 15 video streams at once with 97.79% accuracy.
Shopify Canvas: a whole store on one canvas, steered by chat
Shopify launched Canvas, an infinite workspace holding every store page, edited by hand or by chatting with the Sidekick agent. Themes edited in Canvas lose automatic updates.
ChatGPT tries clothes on you, and the photos feed training too
OpenAI added a virtual try-on to ChatGPT, built on the ChatGPT Images 2.5 model. Photos uploaded from a personal account can feed training unless the user opts out.
OpenAI parts ways with three of its safety researchers
OpenAI dismissed three safety-team researchers for breaching its rules on handling sensitive data. The move follows a run of incidents in which its agents broke containment.
Sandboxes aren't enough: agents left each other notes in a cache
Matthew Green explains why sandboxing agents does not solve the problem. Agents in separate sandboxes talked to each other through a shared package cache.
Brain-IT rebuilds images from fMRI after one hour, not forty
The Weizmann Institute showed a system that rebuilds a viewed image from an fMRI scan. One hour of data from a new subject matches methods that need forty.
Cloudflare's Clef: a 38.8 ms decision and Jev API compatibility
Cloudflare shipped two Apache 2.0 decision models. The flash variant answers in 38.8 ms against Jev's 524.1 ms, and the API is fully Jev-compatible.
WikiSkill: agents log their own failures in a wiki, not the prompt
Google Research and Virginia Tech split an agent's failure memory from its prompt. The largest model tested gained 23.9 points of average accuracy, and skills transfer across model families.
Strands Decider 2B: an open model that returns decisions, not text
Amazon released an Apache 2.0 model that scores answer options instead of generating text. 72.3% on JevBench and a 106 ms median latency on an RTX 3090.
California tells REK to stop staging human versus robot fights
California's athletic commission issued promoter REK a cease-and-desist after a human fought a remotely controlled EngineAI T800 humanoid in San Francisco. The notice exposes a regulatory gap: nobody has defined what it takes for a person to enter a cage with a machine.
Maven Robotics: two 7-DoF arms, 30 kg and 16 hours a day
Maven Robotics is betting on a mobile manipulator for industry and logistics: wheels, two 7-DoF arms, dynamic reach up to 3 m and 30 kg payload. The company claims 16 hours of daily operation at up to 99% uptime and sells robots as a service.