This is amazing if it actually works as promised and is probably something even OpenAI would like to have (and potentially buy from you).
Could you share a bit of the behind-the-scenes magic? I know this is an open area of research and part of it involves actually having an equivalent of a “cryptographic blueprint” of the model itself, which obviously, as outsiders we don’t have… so really really curious.
*praying to god that the answer isn’t that you’ve trained a classifier on GPT’s generated outputs*