Advancements in AI: A Comparison Between the U.S. and China
Overview of the Global AI Landscape as of December 30, 2024
Global AI in 2024: Innovation, Competition, and the Road Ahead
American AI companies have long been considered leaders in the field due to their access to top global talent and the latest high-performance chips, which are often deemed essential for developing advanced models. However, in recent weeks, Chinese AI companies have made notable claims of progress, showcasing models that they assert rival leading Western AI technologies.
Breakthroughs in Chinese AI Training Methods
Several Chinese AI firms have introduced cutting-edge developments:
- DeepSeek: Supported by a leading hedge fund in China, DeepSeek unveiled a demonstration of its large language model (LLM) in November, claiming it matches the capabilities of OpenAI’s most advanced models.
- Moonshot AI: Backed by Alibaba and Tencent, Moonshot AI revealed a model specialized in solving mathematical problems, reportedly rivaling OpenAI’s capabilities in this domain.
- Alibaba: The tech giant stated that its experimental models outperform OpenAI in specific benchmarks.
Despite these bold claims, independent evaluations remain challenging due to the lack of a universally accepted benchmark for assessing AI performance. One possible measure, the American Invitational Mathematics Examination (AIME), tests advanced mathematical abilities with a rigorous three-hour exam. While DeepSeek asserts its model outperformed OpenAI’s, independent tests by The Wall Street Journal showed OpenAI’s model solving all 15 questions faster than the Chinese models. However, all models, including OpenAI’s and those from DeepSeek, Moonshot, and Alibaba, provided correct answers—a significant milestone compared to earlier AI systems that struggled with basic arithmetic.
Navigating U.S. Export Restrictions
The achievements of Chinese companies are particularly impressive given their lack of access to the most advanced AI chips. Since October 2022, the Biden administration has imposed strict limitations on the export of high-performance AI chips to China, aiming to curb the country’s ability to develop advanced AI models and related technologies, including military systems. These restrictions have tightened further over time, with additional constraints on chip performance and even investments by U.S. funds in Chinese AI firms.
Despite these obstacles, Chinese companies have adapted by developing alternative training methods. For instance:
- Reinforcement Learning: Moonshot AI employs this approach, which mimics human trial-and-error learning, requiring less computational power to improve performance.
- Mixture of Experts (MoE): This strategy involves using multiple specialized models rather than one large general model. Tasks are routed to the most relevant model, reducing computational demands.
Tencent’s MoE model, launched in November, reportedly matches the performance of Meta’s Llama 3.1 but was trained with only a fraction of the computing power.
Innovative Use of Available Resources
Chinese companies have also optimized the use of less advanced hardware. DeepSeek, for example, built a cluster of 10,000 Nvidia A100 chips, achieving performance comparable to larger, more energy-intensive setups. This ingenuity highlights their ability to overcome hardware limitations.
As Jack Clark, founder of Anthropic, noted, “Chinese companies are building exceptional hardware and software systems with the resources they have. Just as they’ve succeeded with electric cars and drones, they’ll create Chinese-made AI models.”
The Road Ahead
While these alternative methods have allowed Chinese firms to stay competitive, the landscape is expected to shift in 2025 as new computing systems leveraging next-generation chips come online. For instance, xAI, founded by Elon Musk, has raised $6 billion to build data centers with 100,000 Nvidia Blackwell chips, the company’s most advanced AI hardware. Similarly, Amazon plans to construct a supercomputer using hundreds of thousands of custom-developed chips.
In contrast, Chinese companies face uncertainty that affects their fundraising and valuations. Zhipu AI, for example, raised funds at a valuation of $3 billion—only a fraction of xAI’s and significantly below OpenAI’s estimated $175 billion valuation. The company also postponed its IPO plans due to challenges in achieving its desired valuation.
Howard Hang, a former senior executive at a Chinese AI firm, described the industry as “dancing with shackles.” He emphasized that focusing on strengths and resourcefulness is the only way forward for these companies.
Conclusion
Despite significant challenges, including export restrictions and limited access to advanced chips, Chinese AI companies are proving resilient. Their innovative approaches, including reinforcement learning and the Mixture of Experts model, demonstrate their ability to adapt and compete on the global stage. However, as next-generation hardware becomes available, the gap between Chinese and U.S. firms may widen, presenting both challenges and opportunities for the future of AI.
