I made some small changes to the Miniverse project. It still feels a bit boring, but I'm trying different experiments, and I think I want to try a different strategy, similar to Voyager for Minecraft. Instead of putting all the responsibility on the LLM to decide what to do each step of the simulation, I want to instead allow it to modify its own imperative code to change its behaviour when need be. Unlike the evolutionary algos of old, this would be like intelligent design, except the intelligence is the LLM, rather than the LLM controlling the agents directly.
Before I do this however, I decided to clean the codebase up a little, and make the GitHub repo public, as multiple people have asked me for the code. It could use a bit more cleanup and documentation, but at least there's a separation into files now, rather than my original approach of letting the code flow through me into a single file:
I also added some more UI to the front end so you can see when someone's talking and what they're saying, and some quality of life changes, like loading spinners when things are loading.
There's still a lot that I can try here, and the code will probably shift drastically as I do, but feel free to use any of it. You need to set the OPENAI_KEY
environment variable and the fly.io config is available too if you want to deploy there (which I'm doing). The main area of interest is probably NPC.js which is where the NPC prompt is built up.