The ‘strong’ feature hypothesis could be wrong — LessWrong