Dreamer v3 in model-based direction had some interesting scaling plots showing a pattern of faster (per-sample) learning using bigger models. In terms of gener

Dreamer v3 in model-based direction had some interesting scaling plots showing a pattern of faster (per-sample) learning using bigger models. In terms of generalization, Ada by Deepmind was also quite impressive, but operated within simplistic world of tasks