He probably wrote this half in jest, but it is actually a serious issue in fields like Natural Language Processing and Computer Vision that are have their intellectual roots in Computer Science.
Papers in CV and NLP are about data structures, algorithms, software engineering methods, neural networks, statistical models, and so on. They do not discuss anything in particular about the structure of language or the properties of images. This is because they are descended from CS. In CS, exemplar results are ideas like QuickSort and Dijkstra's algorithm. These algorithms work on any list or graph; you do not have to know anything special about the particular properties of the list or graph you are operating on.
As an illustrative anecdote, I went to an NLP talk given at MIT by a well-known Google/Berkeley parsing researcher. He gave a talk about a system that used neural networks to do sentence parsing. In the QA session, I mentioned the idea of verb argument structure, and asked how the system would learn, for example, that a verb like "persuade" or "argue" can take a that-complement, while other verbs cannot. He didn't really have an answer, because it wasn't the kind of thing that he worried or thought about. My guess is that he did not consider such a question to be relevant to his field.