Same. More generally, I'd like a model that infers the most likely prompt(s) from transformer-generated words or images. I would much rather know what the prompter was trying to do than what his model generated as a result, and treat whatever was generated as a lossy wire encoding.