Self-MoE: Towards Compositional Large Language Models with Self-Specialized Experts | Read Paper on Bytez