Recent developments in artificial intelligence training methodologies are challenging our assumptions about computational requirements and efficiency. These developments could herald a significant ...

ByteDance's Doubao AI team has open-sourced COMET, a Mixture of Experts (MoE) optimization framework that improves large language model (LLM) training efficiency while reducing costs. Already ...

China’s DeepSeek has published new research showing how AI training can be made more efficient despite chip constraints.

This white paper discusses the critical infrastructure needed for efficient AI model training, emphasizing the role of network capabilities in handling vast data flows and minimizing delays. It ...

DeepSeek has introduced Manifold-Constrained Hyper-Connections (mHC), a novel architecture that stabilizes AI training and ...