TLDR Tech / Crypto
Pluralis: The Last Revolutionary AI Protocol
Pluralis is a decentralized AI training protocol built around the only vertically integrated Multi-Party Training Stack covering pre-training, post-training, inference, and model monetization, a scope that distinguishes it from Prime Intellect (which pivoted to RL), Nous Research (no monetization layer), and Covenant. Its Agora live training run achieves 18-25% Model Flop Utilization across consumer-grade RTX 4090s and 6000s distributed across 50 cities, a result that exceeds the 10-11% MFU xAI recorded on Colossus before ceding it to Anthropic, though it trails the roughly 40% MFU seen in leading centralized runs. The protocol's Subspace Networks apply intra-node compression to solve the low-bandwidth problem, a different layer from the inter-node data-parallel compression approach used by Nous Research and Prime Intellect, and Unextractable Protocol Models hosted in Trusted Execution Environments produce model weights that cannot be extracted or replicated. Token mechanics use a hybrid PoW/PoS system where compute providers stake capital against gradient contributions, and Epoch AI data positions decentralized training runs at 20x annual scaling versus the 4-5x rate of centralized labs.
Read the full story