The day after Christmas, China's small emerging company, DeepSeek, has announced a new AI system comparable to the functions of state -of -the -art chatbots of companies such as Openai and Google.
That alone would have been a breakthrough event. However, the team behind a system called Deepseek-V3 explained a bigger step. DeepSeek engineers use only a small part of the highly specialized computer chips that major AI companies are using for system training in a research paper describing how to build this technology. I mentioned.
These chips are the center of tense technical competition between the United States and China. The U.S. government is trying to maintain the initiative in global AI competition, and is trying to limit the number of powerful chips that can be sold to China and other rivals, such as Silicon Valley's corporate Nvidia.
However, the performance of the DeepSeek model questions the unexpected results of the US government's trade restrictions. These regulations have forced Chinese researchers to use a wide range of tools available on the Internet to demonstrate their creativity.
According to a benchmark test used by US AI companies, DeepSeek chatbots answered questions, solved logical issues, and created a unique computer program that is as competent as those that have already appeared in the market.
And it was created based on inexpensive and challenging things that only the largest company in the high -tech industry (all -based companies) can afford to create the most advanced AI system. Ta. Chinese engineers said that the computing ability required to build a new system was only about $ 6 million. This is about one -tenth of the amount spent by Meta Meta Meta Meta's latest AI technology.
“The number of companies with $ 6 million is much more than the number of companies with $ 100 million or $ 1 billion,” said the venture capital company Page One Ventures Investor Chris V. Nicholson. He says. AI technology.
Since Openai released the AI boom in 2022 with Chatgpt's release, many experts and investors have concluded that there is no company that can compete with market leaders without spending hundreds of millions of dollars on special chips. 。
The world's leading AI companies train chatbots using supercomputers that use more than 16,000 chips. Meanwhile, DeepSeek engineers needed only about 2,000 NVIDIA's special computer chips.
Due to the restrictions on chips in China, Deep Shik engineers said, “It was necessary to train chips more efficiently so that we could maintain competitiveness,” said George Washington University, which specializes in international relations and international relations. According to Assistant Professor Jeffrey Din.
Earlier this month, the Biden administration has announced new rules to prevent China from obtaining advanced AI chips through other countries. This rule has been built based on previous multiple regulations, which prevent Chinese companies from purchasing and manufacturing the most advanced computer chips. President Trump has not yet shown whether to maintain or withdraw.
The U.S. government has strive not to reach a Chinese company, given the concern that advanced chips could be used for military purposes. In response, some Chinese companies have stocked thousands of chips, and other companies are procuring chips from the prosperous underground market.
DeepSeek is operated by a quantitative stock trading company called High Flyer. By 2021, the company poured thousands of NVIDIA chips and used it for initial model training. Although the company did not respond to comments, it is known in China by scouting human resources who have just graduated from a top university, promising to the high -paid and most interesting research topic.
According to computer engineer Zihan Wang, who was involved in the early DeepSeek model, the company could understand technology and generate poetry, and to generate an ace problem for the famous Chinese university entrance exam. He also hires human resources without computer science background.
Deepseek does not manufacture any consumer products, and engineers are completely focused on research. This means that the company's technology is not tied to the harshest aspect of China's AI regulations, in which consumer technology requires government information management.
In December, a major US company has evolved the most advanced AI, Openai announced a new “reasoning” system called O3, which exceeds existing technology performance, but is still widely used outside the company. No. However, DeepSeek keeps showing that it is not so late. This month, the company has released its own impressive reasoning model.
(The New York Times sued Openai and his partner Microsoft for copyright infringement of the AI system related to the AI system. Openai and Microsoft denied these claims.)
An important part of this rapidly changing world market is the old idea of open source software. Like many other companies, DeepSeek has opened the latest AI system. This means that the basic code is shared with other companies and researchers. This allows others to build and distribute their own products using the same technology.
The employees of major technology companies in China are limited to collaboration with colleagues, but “If you are working on open source, you will work with talented people around the world.” According to Unen Chan, the chief software engineer working on open source SGLANG. project. He supports other people and companies to build products using the DeepSeek system.
AI's open source ecosystem gained momentum in 2023, and Meta freely shared the AI system called LLAMA. Many people thought that this community would prosper only if a company like META (a major technology with a large data center full of special chips) continues to open its own technology. 。 However, DeepSeek and other companies have indicated that they can expand the power of open source technology. “
Many managers and experts have argued that large US companies should not open their own technology. This is because technology can be used to spread fake information and cause other serious harm. Some US members of the United States are looking for the possibility of preventing or suppressing this practice.
However, some claim that China will get a significant advantage if regulatory authorities suppress the progress of open source technology in the United States. They argue that if the best open source technology comes from China, US developers will build a system on those technology. In the long term, China may be the center of AI R & D.
“The center of gravity of the open source community has been shifted to China,” said Aeon Steica, a professor at the University of California Berkeley School. “This can be a major danger for the United States,” as China will be able to accelerate the development of new technology.
A few hours after taking office, Trump canceled the Biden administration's presidential decree, which threatens open source technology.
Dr. Steka and his students have recently built an AI system called SKY-T1, which is comparable to the latest Openai O1 system in a specific benchmark test. The required computing ability was only $ 450.
They have realized this based on two open source technology released by Alibaba in China.
Their $ 450 systems are not as powerful as Openai technology or DeepSeek's new systems. And it is unlikely that the technology used will create a system that exceeds the performance of major technology. However, the project has indicated that even a very small resource can build a competitive system.
REUVEN COHEN, a Toronto technology consultant, has been using Deepseek-V3 since late December. According to him, this is comparable to the latest system of Openai, Google, and San Francisco's emerging companies and is much cheaper.
“Deepseek is a way to save money,” he said. “This is a kind of technology that humans like me want to use.”
