Learning a Desire Simulator

"The goal of religious thinking is exactly the same as that of technological research -- namely, practical action."

I love the idea that God communicates with us through nature.

The media of human communications are not nature. They are controlled by humans.

I want that to change. I want the media to be more like nature: beautiful, self-organizing, humblingly vast, and able to adapt well past our grasp.

The best way to achieve this is to automate all media.

For each communication (a byte of sound, a frame of video, whatever), the ideal media algorithm should require only one bit of feedback: Did you want this content or not? It should use this feedback to get better at generating content that you want.

Want to be entertained? Let's design an agent to create entertainment value.

Today, the most entertaining content creators are human, so it makes sense to think about how humans work (and about the world in which their work manifests, like your face as you read this content). Some human content creators have brains; our creation agent may require parts of those brains' functionality. Here's one idea:

neocortex: creative process/action space (markov decision process)
thalamus: action space/action generator (language model cascades)
basal ganglia: action generator/belief updater (thompson sampling)
hippocampus: belief updater/historical data (RAG system)
amygdala: action constraint/observation model (budget checks)
nerves: observation model/environment metalanguage (world simulator)
  

With this stack, our agent can execute a sequence of creative actions (e.g., audio generation, prompt revision, image generation) to produce a terminal observation (e.g., an entertaining video frame) and use feedback to update its beliefs and take better actions. It only needs to work well enough to bootstrap a more end-to-end approach.

The first time you skydive, you need to go in tandem with an instructor. That's not because it's hard to work a parachute: it's because you have not learned to model the environment of a freefall. Once you wrap your head around the freefall environment, 'chute control's a cinch. As comma did for the road, wirehead should learn a simulator for people getting what they want.

"Make something people want" is a motto. "Simulate people wanting something" is a goal: if you can solve desire, you just might solve life.