Guanming Wang

Mentioned in 1 episode

"For the generative world action model, which is you need to decode and then predict the future frames, it's pretty much like a VR of Google Maps. But for the latent world action model, it's like you look at the navigation of your Google Maps and then think about what the car has to do and what kind of action you're going to produce in the future."

Mentioned in 1 episode

"There's another research paper called Fast One. Fast One is more extreme in some sense. They think all the world representation is already learned in the hidden state of the DiT. So you don't even need a decoder to decode videos. You can just use the hidden state as a condition to train your action head."

Mentioned in 1 episode

"Some people, like LeCun, they were doing JEPA. They think about maybe we can have two different encoders — one encoding the current observation and the other encoding a future observation. By doing loss on the current observation latent and the future observation latent, you can teach the model to learn how to predict the future."