Liang Wenfeng believes DeepSeek will not treat video generation or world models as its main direction, because image, text, and video are not closely related to the upper limit of intelligence.