Desk
Autonomous and semi-autonomous software agents.
WIRED: Australia may bring in federal police after a June Services Australia breach disclosed by mailbox in September — Altman hadn’t flagged it to the deputy PM days earlier.
Contrastive-LM open-sources CLM-8B, a System One action scorer that skips text generation and claims up to 9× lower latency than TypeSafe’s proprietary Jev on repeating-action loops.
At Connect, Zuckerberg centered Meta’s consumer bet on Muse — realtime avatars, glasses wake words, Mac computer use, its own email, and a shopping-connector land grab.
Zuckerberg’s Connect “one more thing” was Muse Charm — a keychain-sized Muse totem with a tiny avatar screen, fingerprint tap-to-talk, and a holiday ship promise.
The Verge’s read on Muse Charm: a phone-free, chunky-screen Muse gadget with fingerprint wake, apparent camera and mics, and only a few units built so far.
NVIDIA’s SWE-Serve benchmark shows about one in three AI coding-agent patches that pass other checks still fail when a live inference server loads a real model and serves requests.
NVIDIA’s tutorial shows an AI agent migrating a ROS 2 node to CUDA-backed buffers — zero-copy where it counts, CPU fallback where it must.
A local zero-day let malware steer Meta’s Muse agent — cloud dictation, undocumented settings, and agent privileges turned into an attack surface overnight.
NVIDIA’s SoL-Pi trims coding-agent token traffic nearly in half by evolving the harness — same agent brain, fewer wasteful round trips.
A Muse zero-day lets any local app hijack Meta’s hyper-privileged agent — security theater meets ClickFix, and the agent becomes the malware.
AWS’s Strands team open-sourced a full agent harness claiming 28% lower token cost at similar accuracy — the loop around the model is the product now.
NVIDIA’s eval guide says stop grading vibes and tool-call trivia — gate releases on whether the environment actually reached the goal state.