One thing I would really like to see explored more is the 'double blind' assistant, where neither I nor the assistant is in the loop on who wrote what (I am tal

One thing I would really like to see explored more is the 'double blind' assistant, where neither I nor the assistant is in the loop on who wrote what (I am talking in the context of very long documents that you may have written years ago), and then collaboratively improve. I can half-bakedly do this now by blocking out half my writing, and letting the token predictor do its thing. If the generated portion overlaps with the blocked out part I've written, I know I'm heading in the right-ish direction. If not, back to the drawing board. There is also the opposite excercise, where you finish a passage generated by the model and see whether what you've written falls along similar lines the model would have suggested. In both cases, I rarely use the suggestions ad verbatim, but sharpen my writing based on the discrepancies in the predicted and blocked passages.