The most important thing about natural language is that is is ambiguous. LLMs (like GPT) are intended to work despite that ambiguity: they cannot resolve it.

The most important thing about natural language is that is is ambiguous. LLMs (like GPT) are intended to work despite that ambiguity: they cannot resolve it. Any time a pattern of text can be used to represent more than one meaning, LLMs lose all of their leverage. Only humans can determine after the fact whether the result is - as GPT's authors have presented it - a "feature" or a "limitation".