Decision Transformer Interpretability — LessWrong