Meta launches Muse Wonder, with its strongest version confined to Muse Code
Meta released Muse Wonder for developers and a separate Wonder Full for paid Muse Code users, with independent tests showing strong but uneven agent performance.
Anthropic runs a model across 15 TV boxes, but has not built a rogue replicator
Anthropic demonstrated slow model inference on cheap TV boxes and a separate replication workflow in a prepared emulator network, leaving the bridge between them unproven.
Anthropic reports a sharp fall in extracted reasoning after new defenses
Anthropic says new reasoning-access controls and classifiers reduced estimated reasoning-token extraction on its platform by around 90–99%, without an independent efficacy audit.
Caine Prize opens a human-centred review after an AI-authorship allegation
The Caine Prize confirmed a human-centred review concerning its 2026 winning story, without announcing an AI-authorship finding.
SoftwareGym trains a desktop agent on generated applications and reports transfer to real-app tests
SoftwareGym’s researchers generated 182 desktop applications and reported improved performance on an existing-app benchmark after training an agent on those synthetic environments.
TerminalHorizon reports better agent results by rebuilding tasks around the work they require
TerminalHorizon released 1,535 task environments and reported improved terminal-agent performance after training on trajectories designed to preserve exploration and verification.
Work2Skill turns desktop demonstrations into reusable agent procedures
Work2Skill converted 350 curated desktop demonstrations into textual procedures and reported better agent scores when those procedures were reused on related tasks.
DistressBench finds crisis recognition can fail to lead to human-help guidance
DistressBench reports that models failed to direct users toward human help in 35.3% of tested conversations where they fully recognized a crisis.