Despite China's limitations in purchasing advanced chips, increasing evidence indicates that Chinese AI startups are catching up with America’s leading AI models at an unexpectedly fast pace, surpassing many industry insiders' expectations. Although some models have not yet been officially released and tested, experts who have seen these models in operation say that the advancements of these AI models are "impressive."
According to a report by The Wall Street Journal, a startup called DeepSeek, which is invested in by one of China's most successful hedge fund managers, released a preview of its latest large language model in November. The company claims that the capabilities of this large language model are on par with OpenAI's reasoning model known as o1.
Meanwhile, in recent weeks, other Chinese companies have made similar statements. Moonshot AI, a startup backed by Chinese internet giants Alibaba and Tencent, claims to have developed a model specializing in mathematics whose capabilities approach those of OpenAI's o1. On the other hand, Alibaba claims that one of its experimental research models in mathematics outperforms the preview version of OpenAI's o1.
The report notes that although these companies have yet to publish papers describing their models' contents, many experts have expressed their admiration for these models. Former OpenAI researcher and current AI entrepreneur Carl stated that China is "catching up faster." Carl believes that DeepSeek researchers attempting to replicate OpenAI's reasoning models "figured it out within months," and he admits that many colleagues were surprised.
DeepSeek claims its model beat OpenAI at the American Invitational Mathematics Examination (AIME). However, The Wall Street Journal conducted an experiment using 15 questions from this year's AIME, finding that OpenAI's o1 preview model faster answered than the models of DeepSeek, Moonshot, and Alibaba's experiment.
The report points out that since the end of last year, AI developers have increasingly used a technique called "Mixture of Experts," which reduces dependence on chips.
Anthropic's co-founder Clark wrote on his blog, "One way China bypasses export controls is by building very good software and hardware training with the hardware it can access." He states, "Chinese manufacturing will become a part of AI models, just like electric vehicles, drones, and other technologies." Nonetheless, according to the Chinese higher echelon, the lack of cutting-edge chips is painful for Chinese startups, and the gap will continue to widen.