On January 20, DeepSeek, a lesser-known AI research lab from China, released an open-source model that quickly gained attention in Silicon Valley. The company's paper claims that DeepSeek-R1 surpasses leading models like OpenAI's in various math and reasoning benchmarks, demonstrating significant capabilities at a lower cost and with an emphasis on openness. DeepSeek's success illustrates a new path for Chinese tech firms amid the tech cold war; facing US export restrictions, many have shifted from upstream model-building to downstream applications. DeepSeek, however, has focused on optimizing AI model structures and using resources efficiently. Founded by Liang Wenfeng, a computer science master's graduate, DeepSeek emerged from High-Flyer, a prominent quantitative hedge fund that initially concentrated on deep learning research. The hedge fund amassed resources for AI development, leading to DeepSeek's establishment, reflecting a shift toward long-term technological goals rather than immediate profit. Liang stated that scientific curiosity drove this venture, highlighting a focus on advancing foundational AI research rather than commercialized applications. DeepSeek's team is primarily composed of recent PhD graduates from top Chinese universities, fostering a collaborative environment focused on groundbreaking research.

This approach differs from typical practices in established Chinese tech firms, where competition for resources often hampers innovation. Liang notes that young researchers are more suited for high-investment projects due to their commitment and lack of utilitarian considerations. The firm faces challenges from the US government’s export controls, which limit access to advanced AI chips. Initially possessing a stockpile of 10, 000 Nvidia H100 chips, DeepSeek's ability to compete necessitated more efficient model training methods. The team implemented engineering strategies and innovative designs like Multi-head Latent Attention (MLA) and Mixture-of-Experts, significantly reducing the computing power required to train their latest model compared to Meta's Llama 3. 1. DeepSeek's open-source approach has garnered goodwill in the global AI research community, allowing it to attract contributors and users necessary for enhancing its models. This shift in the existing AI landscape may challenge the effectiveness of current US export controls by demonstrating that cutting-edge models can be developed with optimized resource use, leading to potential changes in perceptions of Chinese AI capabilities.