Secrets of Chinese AI Model DeepSeek Revealed in Landmark Paper

image via New Scientist

September 17, 2025, 9:35 PM

• •DeepSeek's R1 model, designed for reasoning tasks, cost approximately $294,000 to train.
• •R1 is an open-weight model, available for download, and popular on Hugging Face.
• •The model employs a pure reinforcement learning technique, rewarding correct answers rather than human-selected examples.
• •R1's methods have inspired other researchers to improve existing LLMs.

A peer-reviewed study reveals details of DeepSeek's R1 AI model, a cheaper competitor to US-developed tools, which caused a stir in the market upon its release. The study confirms that R1 did not learn by copying outputs from other LLMs, utilizing a novel reinforcement learning approach instead. Researchers highlight R1's influence on AI development and its competitive performance in scientific tasks. The paper also emphasizes the importance of peer-review in evaluating AI systems' validity and safety.

Read original article