Chinese AI companies are using model distillation to rapidly close the performance gap with Western models

Chinese AI companies, including DeepSeek, Moonshot, and MiniMax, are utilizing a technique called model distillation to rapidly improve their own AI models by extracting knowledge from more advanced, established frontier models. This process involves generating millions of exchanges with a larger, better-trained model through thousands of fraudulent accounts, allowing the smaller models to learn and improve their performance at a fraction of the cost and time required for independent development. Anthropic has publicly identified this industrial-scale activity as a violation of its terms of service and access restrictions. The rapid advancement of these Chinese models has fueled significant concern within the U.S. AI industry and government, as it challenges the long-held assumption that Chinese AI development is significantly behind. This development has intensified the ongoing U.S.-China AI race, with policymakers and industry leaders debating the effectiveness of export controls on advanced chips as a strategy to maintain a competitive advantage. The situation has also prompted calls for a global AI watchdog to oversee model development and mitigate potential security risks.

Chinese AI companies are using model distillation to extract knowledge from more advanced frontier models. Anthropic reported that DeepSeek, Moonshot, and MiniMax generated 16 million exchanges with its Claude model using fraudulent accounts.

Model distillation allows smaller, less-resourced AI teams to rapidly improve their models' performance. The rapid progress of Chinese AI models has challenged the assumption that they are six months behind Western technology.

U.S. policymakers are reevaluating the effectiveness of export controls on advanced chips as a method to slow Chinese AI progress. The AI industry is increasingly calling for a global watchdog to manage the security and development risks of advanced AI models.