On the very last day of 2025, the Chinese AI lab DeepSeek disclosed technical details about a training innovation that analysts quickly recognized as non-incremental. Their new paradigm, ...