Decentralised AI training was supposed to be impossible. A 30,000x bandwidth gap said so. Then a Bit-tensor subnet trained a 72 billion parameter model over commodity internet. Here is how, and what it changes.