Another fun project I started last year is called "CASI" for Cyclical Adversarial Step-wise Improvement. It leverages the fact that when LLMs engage in self-correcting loops they get better at a task. In this case we are pitting 2 models and 2 system prompts with distinctly different purposes against each other to improve a concept or idea. This is unlike self-play and more like other-play. I would love to include an intermittent training pass or LoRa construction between cycles to make the models focus more completely. CASI is still a work in progress so use at your own frustration. If you encounter any issues please log them with the github repo, you will be helping the entire world :)

https://github.com/TheOneTrueGuy/CASI