The Largest AI Training Run in History Just Happened and it Runs on Bittensor

Eli5DeFi argues Covenant-72B, trained across 20+ independent nodes on Bittensor, achieved 67.11% MMLU on 1.1T tokens—outperforming Meta's LLaMA-2-70B (65.63% on 2T tokens) on per-token efficiency through SparseLoCo compression and trustless validator incentives. TAO surged 19% post-announcement as successful decentralized AI training reshapes infrastructure economics, though Covenant reaches only ~60% of current frontier capabilities and centralized datacenters retain advantages in raw speed and scale.