Large Language Models are not Models of Natural Language: they are Corpus Models.

Natural Language Processing (NLP) has become one of the leading application areas in current Artificial Intelligence boom. Transfer learning enabled large deep neural networks trained on language modeling task to vastly improve performance almost all language tasks. Interestingly, when models are with data that includes software code, they demonstrate remarkable abilities generating functioning computer code from natural specifications. We argue this creates a conundrum for claims provide an alternative theory generative phrase structure grammarsin explaining how works. Since acceptable syntax programming languages is determined by grammars, successful apparently uninformative about theoretical foundations languages, and extension, languages. term model misleading because deep learning not propose adoption corpus instead, which better reflects genesis contents model.