dailyai.report

Google DeepMind developed Decoupled DiLoCo to enable distributed model training across unstable network connections. This method removes the need for constant synchronization between worker nodes. It maintains performance while slashing communication overhead. Researchers can now train large models across geographically dispersed clusters without risking total job failure during intermittent connectivity drops.