Breaking down the training/deployment dichotomy — LessWrong