VLA Past observation -> understand latent representation -> predict possibilities

Many interesting problem in world model

generated video looks nice, hand, figure, looks perfect. Shadows are not well.

team at meta, looked in physics of world model. Looked into video MAE. It has physics emergence. Mainly layer 5 in video mae, the representation of layer 5 inddicates weather model will have physics capabiulities.

Nadia presenting based on that finding. If we have the emergence in physics, can we steer that in more physical

Cost of failure is very high because of prediction error. Cannot handle all failures unless we do simulation.

Nvidia has physical simulators Gaya, World models are now including sound / voice. Good models take data from all sources, thriir own, synthetic, academic, third party data collection.

World model taking long world prediction. It makes sense for non visual things. Descision about war planning - would need long horizon planning.

interesting paper How to detect object when everything is dark,

Interesting question: What evidence would convince us world answer learn mechanistic abstraction- dont know. Peple would do lot of experiments and then find out.