Graphical World Models, Counterfactuals, and Machine Learning Agents — LessWrong