“@WholeMarsBlog We need to predict what pedestrians will do based on their behavior, including limb angle & direction of sight. FSD currently sees all pedestrians as cuboids, so is overly cautious. Also, diffusion seems to be more compute-efficient than transformers for vision.”