There are spaces where LLMs, even those generating actual text can be used in meaningful ways. Personally I don't feel like those spaces are really explored much, instead we force these systems into contexts where their built-in problems just multiply and create bigger issues.