According to Demis Hassabis on X, a demo built in Google AI Studio showcases Nano Banana 2 performing frame-to-frame world modeling by seeing only the previous image and predicting the next, maintaining striking temporal consistency. As reported by Hassabis, the setup constrains input to a single prior frame, highlighting the model’s learned scene dynamics rather than simple sequence memorization. According to the post, the consistency suggests improved latent world models that could strengthen robotics perception, video forecasting, and autonomous planning pipelines. For product teams, this points to near-term opportunities in video QA, predictive maintenance from camera feeds, and low-latency agent planning where next-frame inference reduces compute and improves responsiveness, according to the same source.