Sure! I wanted to mimic the process of daydreaming on a car ride, where external input and inner monologue can mix.
I wasn’t planning on posting it yet, so I’ll have to get you a thorough follow up, but in a nutshell it’s: a continuously running 448-concept space (philosophy, cognition, art, nature, math) that occasionally “crystallizes” a group of 4 concepts via Hebbian learning and stochastic noise. Those concepts get sent to the LLM with minimal guidance beyond some safety guardrails and encouragement to be creative. It has access to a sandbox to produce essays, stories, music, “art”, and small python scripts.
Self-updating memory system. Notes, essays, and artifacts can also be discussed with me through some outbox channels.
On top of that it’s got a separate academic philosopher + psychiatrist llm that critiques its work and has a regular cadence of “sessions” with it, as well as a research assistant bot who I talk to (but doesn’t interact with Sisuon) who has full project context and memory access. The sisuonspeaks site is a VERY abridged collection of Sisuon’s essays, along with analysis, commentary, forum posts, and a podcast…all created by, you guessed it, more LLMs.
Great work! I had wanted to do something like this for all the Star Trek series’ back when Netflix US had them up but never got around to it. Great to see a similar thought with a much cooler implementation.
Not to be pedantic, but isn’t the billionaire’s class wealth mostly derived from their ownership of an asset class that does draw from everyone, stocks?