At @covenant_ai, we made decentralized RL for LLMs as fast as centralized RL. We achieve over 100× bandwidth reduction for weight sync. Completely lossless. Here's how we did it! (1/n)

Launched the fully async trainer on @grail_ai mainnet, training a 7B model 🚀 Training now runs continuously from a replay buffer! No waiting on communication. This unlocked 2k+ context and let us scale from 1.5B → 7B models and even bigger later on.

I just wrote a blog on how we created the first version of grail (grail-v0), a fully open, decentralized RL training system for LLMs. One of its first kinds. Everything is public: the codebase, the training loop, the live training curves, the incentives, the rollouts. 🧵 1/4