How can you call this 'Absolute Zero' if you need to start with a pretrained LLM? From what I understand, this just proposes that you can take an existing LLM,

How can you call this 'Absolute Zero' if you need to start with a pretrained LLM? From what I understand, this just proposes that you can take an existing LLM, have it generate tasks and solve the tasks, and have it learn from that. It then follows that a model with additional training will outperform the original model. I'm assuming that I'm misunderstanding something, because this doesn't seem very novel? Edit: Seems like a variant of adversarial training?