Exploring safe exploration — LessWrong