A New Interpretation of the Marshmallow Test — LessWrong