Decoupled DiLoCo: A new frontier for resilient, distributed AI training

In three linesGoogle DeepMind introduces Decoupled DiLoCo, a distributed training method that improves resilience and efficiency for large-scale AI model training. The technique decouples local and global updates to reduce latency and network failures.

Summary generated by Claude — human-verified