[refulgentis]

yes, hopefully I'll write it up soon. TL;DR: I used this top of GPT-3.5 to generate a magazine page of personalized recommendations: 3x of a title, paragraph, AI art, and a font name, and rationale. For images, I use SD 2.1 via stability.ai. Be sure to add 5400 dpi digital art at the front of your prompt :)

I'm a bit embarrassed to, "real" research finetunes internal models to play a particular role, rather than orchestrating several "conversations" and hoping your prompt will get you the right output format 100% of the time, etc.

Here's a woefully lacking diagram of this user/interpreter/LLM flow for a cohesive longform story generator. [1]

The coolest part of this design pattern you've ID'd is you can always add one more character/conversation that the interpreter orchestrates

ex. A DB character whose role is taking a new page as input, then outputting the new DB, where the DB is all important facts to sustain over a story. That let me scale to 16+ "pages"

[1] https://twitter.com/jpohhhh/status/1632082749317054468?s=20