Learning in layers of muscle memory — LessWrong