From what I understand, the goal is to train an LLM that is better at training LLMs than humans, so that it can continuously train smarter models and, once smart enough, design the successor to LLMs.