Chinese AI startup DeepSeek has published a paper detailing a new training method, Manifold-Constrained Hyper-Connections, aimed at improving AI scalability while reducing computational and energy demands. This development highlights China's efforts to advance its AI industry despite U.S. restrictions on advanced semiconductor access. DeepSeek's previous innovations, like the R1 model, were developed at a fraction of the cost of Western rivals, and anticipation is high for their next flagship system, the R2, expected soon.