DeepSeek Touts New Training Method as China Pushes AI Efficiency

DeepSeek has published a paper outlining a more efficient approach to developing AI using a framework called Manifold-Constrained Hyper-Connections. The technique is designed to improve scalability while reducing the computational and energy demands of training advanced AI systems. The publication suggests that DeepSeek is ready to release a major model soon. Its R2 model is expected to be released around the Spring Festival in February.