Reddit user runs 8 agents at 250K context on two RTX 3090s

Apple's framework splits the agent into a planning agent and a UI grounding model, trained with PPO plus GiGPPO neighbor-conditioned value estimation and LLM-as-a-Judge step-level rewards. The 8B RL agents beat SFT and terminal-only RL baselines on UINavBench, iOSWorld and AndroidWorld, with training run across hundreds of concurrent virtual iOS devices.

Researchers built Realtime-Bench, 50 interactive game and web tasks (30 web, 20 game) testing whether models can act under real-time pressure. It probes temporal state awareness, adaptive deliberation, and time-aware action selection across wall-clock, paused-clock, and token-clock conditions.

Lumentum, a data center optics maker, is valued at about $83 billion; Marvell at roughly $221 billion. The all-stock AI infrastructure deal is expected to close in the first half of 2028.