Democratizing AI: The Psyche Network Architecture - NOUS RESEARCH

Nous Research describes Psyche as a decentralized infrastructure that lets anyone take part in training large language models across distributed, heterogeneous hardware instead of a single centralized cluster. The post says its DeMo and DisTrO optimizers compress the information nodes exchange, and that Psyche adds overlapped training and a 1-bit quantization option, which compressed results by more than 3x. Psyche's architecture uses an on-chain coordinator, GPU clients that train and witness, and P2P networking built on Iroh with UDP hole-punching.