So I've got this project going on, on the backburner called GhostTown. It's incredibly interesting. The idea is to, essentially, invent a new American city, power it with imaginary people that respond and post online with organic patterns, in real time. What I wanted to do was publish it, and convince Google and America that it was a real town, but it was never quite stable enough for that.
Instead, it's been quietly running simulations with it since 2018. It's currently running GPT 2. What's exciting about it is the stochastic nature of the thing. It's governed entirely by random numbers and totally arbitrary deterministic rules that build different psychological profiles, and then, sets the role playing agents out to interact with each other in a mock regional social network.
It's wild to watch, but I think I could do it live on the internet now, with a much higher level of sophistication. Actually build a real functional regional social network around it. Generate images, and a weekly podcast about the events and happenings of the place. The original version capped at 5,000 local "resident" agents. The new one would start with 10, and scale to 20,000.
They would also have a lot more things to do.
I started by writing a GNN designed for social heuristics. This piece is universally applicable, and might be the first actually useful social ranking algorithm ever written in the open source space.
I'll post progress reports, but if you want the ranking algo, you can download that now. Everything's documented, it's ready to use.