We should put a model in an actual body and let it in the world to build from experiences. Inference is costly though, so the robot would interact during a period and update it's model during another period, flushing the context window (short term memory) into its training set (long term memory).
replies(2):