Hey I am an undergrad at Berkeley and I have been working on training self-learning LLM agents by training two LLaMAs to play each other on simple board games like tic-tac-toe and connect-4 (kind of like AlphaGo) and thought this was pretty interesting research. I am looking to do more experiments on more complex multi-agent environments and also inject multimodal input and maybe train it on web actions. If you know anyone working on this problem please let me know! I am planning on joining an academic research lab or taking a leave of absence to work with a startup doing something similar to this.