The shallow reality of 'deep learning theory' — LessWrong