China faces significant challenges in advancing in the field of artificial intelligence. While U.S. tech giants continue to acquire the latest-generation GPUs to train their models, the Asian country faces restrictions that hinder access to these critical chips. Despite these barriers, Chinese researchers are finding ways to navigate the limitations, leading to impressive advancements in AI models.
DeepSeek. One of the latest developments comes from DeepSeek, a research lab in China that has recently introduced DeepSeek-R1. According to its creators, this model is the first AI capable of competing with OpenAI’s o1 in terms of “reasoning” abilities.
Rethinking responses. TechCrunch reports that DeepSeek-R1 employs a reasoning process similar to that of o1. Instead of providing an immediate answer, it reviews multiple potential responses before selecting the most suitable one. This approach results in longer wait times for users seeking a response.
As good as o1. Performance tests indicate that DeepSeek-R1 exhibits behavior comparable to o1, which OpenAI launched in September 2024.
Far from perfect. However, some users have noted that DeepSeek-R1 struggles with logical problems and even simple games like tic-tac-toe, which also pose challenges for o1, given its own limitations. Additionally, DeepSeek-R1 can be susceptible to manipulation, with jailbreaking techniques allowing users to bypass the security measures designed to prevent the generation of harmful content.
No questions about China. In other tests, users observed that the model tends to apologize and refrain from answering inquiries about its opinions on the political situation in China. The Chinese government has likely worked to prevent these kinds of interactions. In fact, according to the Financial Times, China has censorship mechanisms in place to ensure that chatbots align with the country’s values and policies.
Who’s behind DeepSeek? The startup is funded by High-Flyer Capital Management, a Chinese investment fund that utilizes AI to make investment decisions. The company operates its own server clusters to train AI models. Its most recent setup features 10,000 Nvidia GPUs, which cost approximately $138 million.
Similar to the major AI players in the U.S., High-Flyer’s goal is to successfully develop artificial general intelligence through DeepSeek. Interestingly, the creators of DeepSeek have announced their intention to publish their model as open source and offer an API.
China is making rapid strides. The launch of DeepSeek-R1 showcases the significant progress the country is making in AI development. Earlier this year, DeepSeek released a model called DeepSeek-V2 and prompted local competitors such as ByteDance, Baidu, and Alibaba to lower the prices of their AI models. Some even became free to use.
The AI race features prominent players in the U.S., including OpenAI, Google, Anthropic, and Meta. DeepSeek’s latest release proves the competition is also gaining traction in China.
Image | Julien Tromeur
View 0 comments