Zee
127 posts





If you’ve been wondering what “collective RL” actually looks like in practice, well, SAPO is the answer. SAPO is a new way to post-train language models, built directly on the @gensynai network. RL is incredible for boosting LM reasoning… but let’s be honest, it’s usually pricey, fragile, and stuck behind massive centralised GPU clusters. SAPO (Swarm sAmpling Policy Optimization) changes the rules. Instead of training in isolation, every node on Gensyn runs its own model, generates its own rollouts, and shares them with the swarm. They’re lightweight, model-agnostic, and simple enough for almost any device to contribute. And the payoff is huge: when one model figures something out, that insight can ripple through the entire network. Training becomes faster, cheaper, and genuinely collaborative. SAPO shows what happens when RL stops being a closed, individual effort, and becomes a shared, decentralised learning process powered by a global compute network. If you care about open post-training, shared experience, or community-driven LM improvement… keep an eye on this. We’re just getting started.

























GM and Gsent CT Guess who I caught sneaking out for a little 2 AM fridge raid? Dobby #SentientAGI @SentientAGI







