Genie: Generative Interactive Environments

https://proceedings.mlr.press/v235/bruce24a.html

11B foundation world model from unlabeled video; 30k h of 2D platformer clips (6.8M 16 s clips); 8 latent actions; tokenizer ~200M, LAM ~300M, dynamics 10.1B; ~1 fps; 16-frame memory.

Genie foundation world models