I wonder how hard it would be to fine tune an LLM on nostr data to optimize for zaps or replies