Applied AI Evidence: Human-AI Decisions, Clinical Safety, and Robotics

BirdDreams reports 35% lower action-following error and 21% lower unconditioned trajectory error for a driving world model, plus a human preference signal; independent validation, varied-condition transfer, and deployment costs remain open. KITScenes-LongTail adds a dataset lead for finding plausible driving trajectories based on wrong reasons, but the post offers no benchmark results. Other applied work spans visual perception, view synthesis, robotics, ocean forecasting, and physics simulation, with quantitative evidence and practical validation uneven. Do not repost BirdDreams, which is already shared.