China's AI disrupter DeepSeek bets on low-key team of 'young geniuses' to beat US giants

Chinese AI startup DeepSeek has stunned the industry by developing an advanced large language model (LLM), DeepSeek V3, using limited resources and a “young geniuses” approach to talent. Rather than recruiting long-established AI researchers or overseas-educated PhDs, the Hangzhou-based firm prefers bright newcomers who have excelled in fields like physics and engineering at top universities in China. Founder and hedge-fund manager Liang Wenfeng, himself an AI graduate of Zhejiang University, spun DeepSeek off from High Flyer-Quant in 2023 and cultivates a low-key, mentor-like atmosphere, guiding junior engineers and researchers with subtle suggestions. DeepSeek’s team of around 150 researchers and engineers, plus a dedicated data automation wing, managed to train the V3 model in about two months using only 2,000 Nvidia H800 chips for roughly US$6 million—significantly less than typical training budgets for similarly capable models.

Despite its “shoestring” development, DeepSeek V3 has matched or even surpassed certain U.S. rivals, including systems from Meta and OpenAI. Much of this success rests on DeepSeek’s novel training architectures and techniques, such as Multi-head Latent Attention and DeepSeekMoE, which allow the firm to achieve major gains in performance without the usual financial and computing power requirements. High-profile engineers like “AI prodigy” Luo Fuli have helped raise DeepSeek’s visibility, reportedly drawing generous employment offers from other tech giants. As competition between Chinese and U.S. AI developers intensifies—and with China’s limited access to top-tier chips—DeepSeek’s resourceful approach and unconventional hiring methods have gained attention as a possible blueprint for closing the AI gap.