@ESYudkowsky seriously tho, I don't find this all that surprising and had been assuming I could do something similar to compress chats into summaries to get to approx the same spot in latent space w fewer tokens