A reason to be skeptical that current deep RL contributes to a theory of general intelligence: even if you know the full dynamics of video-game state-variables, studying fully-specified reward-max Bellman equations gives you no mechanistic theory of what generalization is.