China’s AI Cat-and-Mouse Game: Distillation and the Limits of Replication

A recent Reuters report has highlighted Chinese military-linked researchers using advanced American AI models for distillation. This technique allows smaller systems to inherit selected capabilities from more powerful AI models, raising concerns about unauthorized extraction of capabilities.

Model distillation is not new, but its use in defense-related applications is a significant development. Chinese researchers have been using OpenAI’s GPT-3.5 and Anthropic’s Claude 3 Haiku to train their own systems, which can then operate within Chinese military networks. The PLA’s Unit 96941 has used GPT-3.5 to process sensitive military source code, creating a domestic AI model that can operate within Chinese networks.

Researchers at the North University of China have also used Anthropic’s Claude 3 Haiku to create synthetic training data for social media monitoring and content moderation. These examples demonstrate the versatility of model distillation and its potential applications.

However, experts argue that smaller models generally inherit selected capabilities from their larger “teacher” models rather than reproducing their full intelligence. This limits the extent to which Chinese researchers can replicate frontier AI capabilities. Trevor Koverko, co-founder of AI data company Sapien, notes that model distillation is not a game-changer in itself.

The cat-and-mouse game between the US and China continues unabated. Beijing studies the security risks created by distillation while Washington remains wary of allowing its advanced technologies to be repurposed for defense applications. The dispute centers on unauthorized extraction of capabilities, rather than model distillation per se.

Replicating frontier AI capabilities is an expensive and complex task that requires significant computing resources. Chinese researchers face significant hurdles in achieving true independence from these systems. Despite these challenges, the stakes are high, with both nations vying for technological superiority.

The world watches this high-stakes game unfold, seeking to separate fact from fiction. What are the real implications of model distillation, and how does it fit into the broader AI landscape? The answers lie in the ongoing research and development efforts of both nations. One thing is certain: the AI cat-and-mouse game will continue.

The challenge ahead lies not just in technological advancements but also in navigating export controls and security risks to emerge as leaders in the AI era. Can China and the US find a way forward, or will they remain locked in this complex web of technological developments? The world is watching, and the answer won’t be found in distillation alone.