Reading notes, course summaries, and things I learned from others.
An interactive two-round evaluation and preview panel for Code-to-Video agents: 15 open-ended tasks spanning science explanations, multi-view embodied soft-body manipulation, music videos, and reference video recreation, with 136 rendered videos, full trajectories, side-by-side comparisons across six models (with and without HyperFrames skills), and Gemini 4 Argon multi-dimensional blind judging reports.
An interactive Chinese atlas of the 108 long-horizon desktop tasks in OSWorld V2.1: per-task requirements, workflows, grading, atomic skills and trajectory screenshots; per-task results of five frontier runs with failure analysis; how the ten challenge tags are tested; and the environment pipeline.