The value proposition is that someone who is a lousy writer (perhaps only in English) with deep domain knowledge is going back and forth with the LLM to express some insight or communicate some information that the LLM would not produce on its own.
Wouldn't it work better to just write the thing in whatever language they can actually write in and then do a straightforward translation in a single pass?
> someone who is a lousy writer with deep domain knowledge going back and forth with the LLM to express some insight or communicate some information that the LLM would not produce on its own
This sounds reasonable on its face, but how often does it actually come up that somebody can't clearly express an idea in writing on their own but can somehow get an LLM to clearly express it by writing a series of prompts to the LLM?
And, if it does come up, why don't they just have that conversation with me, instead?
Nontrivial translation tools are AI(neural net)-based tools (although not necessary LLM). Whole transformer neural net architecture was originally designed for translation.
Because (the royal) you will be argumentative and shitty, and sour this person on their desire to communicate their knowledge at all.
Just as Google-enhanced output and Wikipedia-enhanced output has helped my writing/thinking, I believe LLM-enhanced output also helps me.
Plus, I personally gain more benefit from using an LLM as a researcher than as a writer.