Running a decentralized 72B parameter model training, the largest permissionless pre-training attempt coordinating global GPUs without whitelisting using compressed gradient techniques.