China’s DeepSeek finds a way to help AI get better at answering questions. Here’s how it works
undefined 9, 2025
Chinese AI startup DeepSeek, working with Tsinghua University, has unveiled a breakthrough technique to make AI models smarter and faster—without making them bigger. The method, called self-principled critique tuning (SPCT), allows AI to judge its own answers based on self-created rules. Paired with a system called generative reward modeling (GRM), the AI can give itself feedback and improve on the fly. Unlike traditional methods that require huge models and tons of computing power, DeepSeek’s approach runs evaluations in real-time to refine responses. They claim it outperforms rivals like GPT-4o, Gemini, and Llama—and plan to open-source it soon.
Read full article
