Long-horizon video prediction is the hard nut that world models can't avoid. Matsuo's team at the University of Tokyo, in SUNTA publicly posted to arXiv, cuts in from the most easily overlooked angle: who exactly decides the segmentation boundary of the Hierarchical State Space Model (HSSM). Past HSSM used fixed-length slicing, or used inter-frame similarity to find transition points, but these heuristic rules often misalign with the data's own temporal structure. SUNTA proposes using "surprise-based chunking" to drive segmentation — the moment the prediction error suddenly rises, is when a longer-range context should be invoked. This change may seem plain, but it actually solves two old problems: "hierarchy collapse" that easily occurs during end-to-end training, and the embarrassment of the surprise signal disappearing during open-loop prediction. The specific approach is a two-pronged: use a decoupled training strategy to preserve the surprise signal; in the prediction expanded by imagination, use the model's internal "inconsistency" as the top-level surprise indicator to decide when to switch segments. The effect is immediate — on 2D/3D video prediction tasks, SUNTA is the only method that can still maintain accurate prediction after 250 steps, with all baselines starting to degrade in the first 10 steps. This line of thinking is a direct "technology learning list" for today's world models competing on long-video consistency (Sora, Veo, Kling, etc.): hierarchical abstraction should no longer be handed to human-designed windows, the model should learn to "be surprised" on its own. When AI truly learns to switch gears when interrupted by surprise, can long video move from 5 seconds of coherence to 5 minutes of coherence.