My writing improves significantly when I sleep on it. My sleep does not.

Engineering work done for now. ✅ Environment built in C ✅ Extended PufferLib to deal with a dynamic number of agents ✅ New training algorithm ✅ New multiprocessing infra It's finally time to run some experiments!

"The Mom Test" is accidentally the best prompt engineering book ever written. It teaches you how to ask questions that get truthful answers: from your Mom OR ChatGPT.

x.com/i/article/1976…

“You should only do the research that wouldn’t happen if you weren’t doing it.” I feel research in kinship-aligned MARL is inevitable but at the same time no one is working on it - I have to do it

x.com/i/article/1977…

I trained AI agents with 2 reward functions: 1. Maximize gene copies 2. Maximize LOG of gene copies (1) leads to overpopulation and extinction (2) leads to sustainable growth Investors have been using (2) to avoid bankruptcy. We showed evolution works the same way. Read it

Sangyun Lee Yes! In this post, population control had to emerge because there were no hardcoded property rights. The food had to be stored in a physical location and would be exhausted by other agents if population grew too much. In my old env, the naive reward worked fine because