Suddenly going viral: China's "mysterious power" shakes up the U.S. tech community

After the release of this model, it sparked discussions among many technology leaders in the overseas AI circle. For example, Jim Fan, a senior research scientist at Nvidia, publicly tweeted on his personal social platform: "We are at a historical moment: a non-American company is continuing OpenAI's original mission - empowering all of humanity through truly open cutting-edge research. It seems unreasonable, but the most interesting endings are often the most likely to come true."

Why did DeepSeek suddenly become popular?

Bonus

Expert Interpretation: Why China's New AI Big Model Is So Popular on the Internet

Recently, a Chinese AI startup called DeepSeek has become a hot topic in the field of artificial intelligence (AI) big models at home and abroad. In less than 30 days, DeepSeek has released two big models, DeepSeek-V3 and DeepSeek-R1. Their costs are low compared to foreign big model projects that cost hundreds of millions or even tens of billions of dollars, and their performance is comparable to that of foreign top big models. At the same time, DeepSeek is different from the closed-source path of foreign big model giants and adopts an open-source model. The development model and achievements of this Chinese company have attracted great attention from Silicon Valley. Many Western mainstream media have published articles lamenting that "Chinese AI models shocked Silicon Valley", and even triggered a "trend" of many well-known manufacturers and institutions at home and abroad trying to reproduce DeepSeek's achievements overnight. What are the characteristics of DeepSeek's development? Does it bring some inspiration to the development path and innovative ideas of domestic large models? The Global Times reporter interviewed several experts in the field of artificial intelligence on the 26th.

 "OpenAI o1 economical and open competitor"

DeepSeek released the large model R1 on the 20th of this month, and said that "in tasks such as mathematics, code, and natural language reasoning, the performance is comparable to the official version of OpenAI o1", which has attracted great attention from foreign media, especially American media, to this Chinese company and its latest large model results.

"China's cheap and open artificial intelligence model DeepSeek makes scientists excited." "Nature" magazine said on the 24th that the large language model DeepSeek-R1 developed by China has excited scientists. It is considered to be an economical and open competitor to "reasoning" models such as OpenAI o1.

The New York Times reported on the 24th with the title "How Chinese AI Startup DeepSeek Competes with Silicon Valley Giants" that achieving the above results is already a milestone, but the team behind the DeepSeek-V3 large model described a greater improvement. They used only a small fraction of the highly specialized computer chips used by advanced AI companies to train the system. Chinese engineers said they only spent about $6 million and about 2,000 Nvidia dedicated chips to complete the training of the new model, which is far less than the world's leading AI companies in terms of both funding and chip usage.

"This is not a question of China catching up with the United States, but a question of open source catching up with closed source"

DeepSeek has attracted much attention, in addition to its high cost-effectiveness, there is another reason: open source. In recent days, a wave of DeepSeek reproduction has appeared on the Internet. The University of California, Berkeley, Hong Kong University of Science and Technology, and the well-known artificial intelligence company HuggingFace have all successfully reproduced the model, using only reinforcement learning, without supervised fine-tuning, and even with a cost of only tens of dollars.

The US Reddit said on the 25th that China's DeepSeek's model is open source, which is the real reason for excitement. Basically, they make the knowledge of making these things available to the world for free, ensuring that no one can really monopolize it. Chinese companies are basically doing the exact opposite of what American companies do. Can you see OpenAI, Anthropic or Google open source any powerful models? So far, all we have from them is fur. Meta is the only major Western company that has made a significant contribution to open source large models, but they may not open source their best models in the future. Yann LeCun, known as one of the "three giants of deep learning", said on the social platform X that this is not a question of China catching up with the United States, but a question of open source catching up with closed source.

Liu Wei, director of the Human-Computer Interaction and Cognitive Engineering Laboratory of Beijing University of Posts and Telecommunications, said in an interview with the Global Times that the three core elements of a large model are data, algorithms, and computing power. Deepseek uses less data and less computing power, and achieves equivalent or even better results than well-known foreign large models through algorithm optimization, which is very worthy of recognition. At the same time, it should be seen that it is open source and can be used and reproduced by users around the world who want to use this large model.

Shen Yang, a professor at the School of Journalism and the School of Artificial Intelligence of Tsinghua University, told the Global Times on the 26th that DeepSeek's large model is a very good one among the global open source large models. It is an innovative breakthrough that surpasses traditional pre-training technology by mixing multiple advanced technologies. He talked about several advantages of this large model based on his own experience. First, it combines the current methods of improving the capabilities of AI large models with micro-innovations in engineering. Second, DeepSeek has published relevant papers, and the whole process can be reproduced by everyone. This is the power of open source. Third, DeepSeek's reasoning process has its own innovations. As a researcher in the field of AI, Shenyang has used AI more than 30,000 times. He believes that compared with American AI, DeepSeek still has many Chinese elements in it, such as some hot words in the Chinese network.

Improving reasoning ability

As for what important inspirations Deepseek's development model provides for the development and innovation of large models in China, Liu Wei believes that "innovation is not planned, and it requires the market and professional institutions to find new ways through long-term research. In particular, some commercial companies that have long focused on vertical fields can find better innovation points through reflection on technical paths and a sense of market development. The initial development process of OpenAI is also like this. It was not planned by the US government and technology giants with heavy investments."

Recently, OpenAI, SoftBank and other companies announced the "Stargate" plan to spend $500 billion in four years to accelerate the development of artificial intelligence in the United States. Liu Wei emphasized that this development path that concentrates human, financial and material resources and then gives policy inclinations has certain uncertainties in future research directions and research results. "We should encourage more domestic commercial companies and scientific research institutes to focus on their own research fields and find innovation and development paths that suit them." Shenyang said that in the history of AI development, new breakthroughs are often driven by inconspicuous engineering innovations and scientific exploration. This trend is deeply reflected in the achievements of DeepSeek, which not only breaks through the traditional training methods, but also brings a new perspective to the improvement of reasoning ability. "Although its achievements are still at a stage level, its engineering contributions and theoretical innovations have laid an important foundation for the future development of AI." Shenyang believes that the contribution of the DeepSeek team in pre-training basic models is not only a breakthrough at the technical level, but also lies in the refinement and efficiency of its engineering methods. This engineering innovation of DeepSeek marks a new stage in AI model training, which not only reduces development costs, but also provides a reference path for other companies. At the same time, DeepSeek's core innovation is also reflected in the improvement of reasoning ability, especially through the innovation of related algorithms to promote the natural reasoning ability of the model, proving a potential in the field of AI-without a large number of expensive thought chain annotations, the model can still emerge with reasoning ability.

Shenyang believes that the success of DeepSeek also allows us to see the future development direction of the AI ​​industry: more open source innovation, deep collaboration between hardware and software, and continuous optimization of model development costs and reasoning capabilities. At the same time, we must also see that although DeepSeek has achieved significant interim results, it still needs to face many deep-level challenges in order to achieve further breakthroughs in the future development path, such as more original training data and algorithm innovation.