Worth noting that the LeWM paper (Maes, Le Lidec, Scieur, LeCun, Balestriero) is more interesting for its training method than the Mario demo. The key contribution is SIGReg — a single regularization term that prevents representation collapse in JEPA architectures, replacing the 6+ hyperparameters you'd need with VICReg or Barlow Twins.

  The architecture itself is a ViT encoder producing a single CLS token per frame, feeding into a transformer predictor with        
  AdaLN-Zero conditioning on actions. Everything happens in a ~192-D latent space. A 15M parameter model trains on a single GPU in
  minutes, which is a very different regime from Dreamer-v3 or TD-MPC2.                                                             
                                                            
  The Mario demo is flashy but the real test is whether the learned latent space is structured enough for downstream planning. The  
  paper shows CEM planning working on Push-T manipulation, which is more practically relevant than game environments.