Skild AI's S1 foundation model learns 10-minute unseen tasks from a single video prompt, with no fine-tuning, via in-context learning.