The decentralized training thesis comes in two critical parts, both of which are insanely, impossibly hard: - You must be able to use swarms of commodity GPUs to compete with large data centers on the frontiers of AI. - You must have a business model for that. The first point has essentially fall