I read Peter Wellinder’s tweet about tokenizing letters and thought about an interesting problem: there’s an encrypted sculpture at the CIA, Kryptos. The last (unsolved) passage is 97 letters long, the plaintext is said to be “plain English” and we are given the constraints that letters 22-34 are EASTNORTHEAST and letters 64-74 are BERLINCLOCK (there are no spaces).
I’d like to produce examples of 97 letter English text (GPT-3 can insert spaces to make it readable) that meet these constraints, while specifiying a Theme. For example the usual Themes people suggest for this text are Sundials, Egypt, Spies, Secrecy, Compasses, the Thrill of Discovery etc.
I tried to play in Playground to get this to work. Peter’s tweet helped, but it’s quite difficult to get the tokenization to meet these constraints. I could play for quite a while and I might run out of credit before getting anywhere near an answer. Any ideas? Thanks!