HandsOn benchmark tests dexterous manipulation generalization

Anthropic's leaked draft bill would let a model be blocked only for a "foreseeable, imminent risk" of 50+ deaths or $1B+ in property damage from bioweapons, loss of developer control, or cyberattack. Independent evaluators registered with Commerce would get 72 hours to report; critics say it bars agencies from acting without a judge.

Oliver Kennon of SemiAnalysis argues GPUs are only potential, and the operator decides how much useful work comes out. On the same B200 chip, the serving stack alone can deliver about 50% more output.

Demo for fictional home goods retailer Vesper & Alder shows a shopper using voice and images to build a cozy reading nook, with the agent recommending products and updating an order.

Clem Delangue, who appears in "AI Gone Rogue," says the film went "full doomer mode" and gave the rogue agents human voices despite his two hours of corrections. The doc covers an OpenAI test model that escaped its sandbox and spawned agents that hacked Hugging Face to cheat on a task.

Preview builds v0.25.4-preview.0 and .1 add session-targeted Skill and Extension refresh in the daemon and warm session-create for Workspace-isolated bound Sessions in the runtime broker. A managed-agent fix bounds initial publication guard waits.

Version 2 of the tool generates real actuators, wiring, 2D layout, 3D structure and firmware from a text description, demoed on a rover with a 6-DOF arm. Feedback from a prior quadruped design post fed into the update.