Google DeepMind unveils Genie 4 world model with native audio

The release adds managed session permission rules with Auto approval, plus an experimental shared agent loop that headless runs now use. It also hardens CLI serve token resolution against project env injection.

Pichai said 75% of all new code at Google is AI-generated and approved by engineers as of April 2026, up from 50% the previous fall.

The listing offers 300 racks of decommissioned hardware: 9,600 Nvidia A100 GPUs and 4,200 AMD EPYC CPUs, with free shipping included.

Researchers Tara Sinclair and Sophia Kazinnik built Sim Fed, an AI simulation whose agents impersonate the 12 rate-setting voters, including "Sim Powell" and "Sim Cook". It predicts the Fed will hold rates steady at Wednesday's decision.

Since dynamic scheduling launched in September, over 70% of Managed Deep Agents runs on LangSmith come from schedules rather than a human typing a message. LangChain's blog details how cron-driven runs unlock memory "dreaming" and proactive triage.

Copilot Fusion combines Project HydraFusion's runtime model orchestration with Copilot Auto Model Selection to pick and route models for coding tasks automatically. GitHub claims it beats Claude Code at half the API cost.

Open-source project pairs local generative AI and cloud models through ComfyUI with a Photoshop-like interface, aimed at users put off by ComfyUI's steep learning curve.