I trained a GPT across 4 workers using Nostr as the communication layer. No central server. No coordinator. Workers exchange compressed pseudo-gradients as signed Nostr events through public relays. ...