ICML Social: Collective, Decentralized Training as a Hedge Against AI Power Concentration

​As frontier AI training concentrates inside a small number of increasingly vertically integrated labs and hyperscalers, the research community faces a structural question: is large-scale model training necessarily centralized, or is centralization an artifact of current engineering and economic choices? Recent work on distributed training, low-bandwidth communication, federated and swarm learning, suggests the latter — that meaningful decentralization is becoming technically tractable at the frontier scale, not just at the toy scale.

​This social convenes researchers and practitioners working at the intersection of decentralized training, federated learning, distributed systems and AI governance to discuss:

(1) the current technical frontier of decentralized model training,

(2) the open research problems that gate further progress — communication efficiency, verifiable computation, attribution, and adversarial robustness — and

(3) what collective or distributed ownership of frontier models could mean in practice, and why it matters as a counterweight to concentration of compute, capital, and decision-making in AI.

​The Protocol Learning social is organized in collaboration with Professor Namhoon Lee and POSTECH.

Join us!


(ICML Conference badge is required as this is an official ICML social hosted without the conference)