The mistake is arriving these abilities to the model itself, and not the content being modeled. Text contains more data than language. Large Language Models wo

The mistake is arriving these abilities to the model itself, and not the content being modeled. Text contains more data than language. Large Language Models work implicitly: they are not limited to finding language-specific patterns in the text that they model. Humans look at LLMs through a lens of expectation. Any time we find a feature we did not expect, we categorize it after-the-fact. That's our biggest mistake: LLMs are not made of categories!