The Rise of DeepSeek, China’s AI Star

Liang Wenfeng, once the head of a quantitative hedge fund in China, transformed his career by betting on artificial intelligence research. He committed substantial resources, amassed 10,000 Nvidia chips and assembled a team of talented young researchers. Two years later, this ambitious project led to the launch of DeepSeek, an initiative that quickly attracted worldwide attention from the technology industry.

On 20 January, the DeepSeek laboratory released an open-source model that caught Silicon Valley's attention. According to a company paper, DeepSeek-R1 outperforms leading models such as those from OpenAI in mathematics and reasoning, posing a serious challenge to Western AI giants.

DeepSeek's success highlights an unexpected consequence of US technology restrictions. With access to advanced chips limited, many Chinese companies focused on applications rather than foundation models. DeepSeek, however, took a different approach: improving AI model architecture to make more efficient use of limited resources.

“DeepSeek stands out for its software-driven optimisation of resources, an approach that encourages collaborative innovation,” says Marina Zhang, associate professor at the University of Technology Sydney.

The Origins of DeepSeek

DeepSeek began as Fire-Flyer, a research arm of High-Flyer, a successful Chinese quantitative hedge fund. Founded in 2015, High-Flyer accumulated GPUs for financial analysis until 2023, when Liang decided to redirect these resources towards developing advanced AI models.

DeepSeek adopted a distinctive recruitment strategy, targeting young graduates from prestigious universities such as Peking University and Tsinghua who were eager to prove themselves. “We chose researchers without industry experience but with an innovative mindset,” Liang said.

Innovation Under Pressure

US restrictions on advanced chips pushed DeepSeek to develop more efficient training methods. “They optimised communication between chips and implemented mix-of-models strategies,” says Wendy Chang of the Mercator Institute for China Studies. Their latest model is so efficient that it reportedly required only one tenth of the computing power needed to train Meta's Llama 3.1.

Sharing DeepSeek's innovations as open source has strengthened its global reputation. “They demonstrate that advanced models can be built with fewer resources by optimising training methods,” Chang concludes.

US restrictions may prove ineffective in curbing advances in Chinese AI, as alternative strategies such as DeepSeek's continue to emerge successfully.

Source: Wired