The Beijing Academy of Artificial Intelligence (BAAI) has unveiled Orca, a ‘world foundation model’ that takes a radically different approach from traditional AI. Instead of predicting the next token, frame, or robot action, Orca models the world’s next state in an abstract internal representation. It uses two complementary training modes: ‘unconscious learning’ from raw videos to grasp motion and scene dynamics, and ‘conscious learning’ from verbal instructions and question-answering data. The model’s frozen core can be paired with swappable ‘output heads’ for text, images, or robot actions—without needing action labels during pretraining. Trained on 125,000 hours of video, Orca rivals specialized systems across five robotics tasks, and its loss decreases predictably with scale. Why it matters: This decoupling of world understanding from task-specific training could break AI’s heavy reliance on labeled data, offering a scalable path to more generalizable intelligence.