China Plays the Long Game in AI While US Chases Superintelligence: Brookings

A new report from the Brookings Institution indicates that China's artificial intelligence industry is advancing through a focus on efficiency, global adoption, and integration into physical machines, contrasting with the United States' primary pursuit of artificial general intelligence (AGI). While U.S. firms invest heavily in AGI development, Chinese companies are concentrating on deploying AI across a wide range of devices, manufacturing processes, and international markets.

Competing AI Strategies

The report highlights a divergence in AI development philosophies. U.S. tech companies are reportedly investing billions in data centers to create AI systems capable of human-level cognitive performance. In contrast, Chinese firms are prioritizing making AI products accessible to a broad user base and embedding the technology throughout their economy. Hamza Chaudhry, AI and National Security Lead at the Future of Life Institute, noted that the U.S. focus on AGI is largely an obsession of a few Silicon Valley companies, whereas China's strategy emphasizes widespread product distribution and application.

Efficiency and Deployment

Chinese AI developers are reportedly excelling at maximizing performance from limited computational resources, a strategy that contrasts with the U.S. approach of building massive compute clusters. This focus on efficiency and deployment is evident in China's rapid integration of AI into physical products like vehicles, smartphones, wearables, and robotics. The country is also expanding the use of autonomous systems, including robotaxis and humanoid robots, rather than solely focusing on achieving superintelligence.

Open-Source Models and Security Concerns

The Brookings report also points to China's utilization of open-source AI models. Chaudhry expressed concerns that this approach allows governments and militaries to access these models, citing reports of the Chinese military using open-source AI. He suggests a need for a broader AI strategy that re-evaluates how the global community interacts with these technologies.

Furthermore, the report raises questions about the role of model distillation, a technique where AI systems learn from more advanced models. Chaudhry suggests that the Brookings report underemphasizes the significance of distillation attacks, where AI models' capabilities are effectively extracted by querying them and using the outputs to train competing models. This technique has been highlighted in reports by companies like Anthropic, OpenAI, and DeepMind, with Anthropic specifically claiming that several Chinese AI labs used fraudulent accounts to extract data from its Claude AI for training purposes.

Potential for Arms-Control Agreements

The differing priorities in AI development between the U.S. and China could create an opportunity for new agreements, potentially similar to arms-control treaties. Chaudhry suggests that this divergence could lead to the establishment of 'red lines' by both nations on specific types of AI development, focusing on what should not be built in the future.