@drbjork Mm, the problem is that you can't know that the output from an LLM is actually what has been written or what is probable to have been written, given what's in the rest of the corpus.
Now, using NLP to present natural language interfaces to database interfaces and sculpting better queries, that's great. So is e.g. summarization of articles (sometimes), but those kinds of NLP tools are much more targetted and easier to control.
@eliasr @drbjork definitely also a valid concern