DeepSeek is in the virus.
After the CHATBOT app climbed to the top of the Apple App Store chart (and Google Play), the Chinese AI laboratory Deepseek broke into the mainstream this week. The AI model of the DEEPSEEK, which was trained using Compute efficient technology, questioned whether Wall Street analysts and technicians could maintain the initiative in the AI race and whether the demand for AI chips would be maintained.
But where did Deepseek come from and how did they rise so quickly?
DeepSeek’s Merchant origin
DeepSeek is supported by the Flyer Capital Management, a Chinese quantitative hedge fund that uses AI to provide information on trading decisions.
Liang Wenfeng, an AI lover, co-founded High-Flyer in 2015. Wenfeng is a hedge fund that focuses on developing and distributing AI algorithms in Zhejiang University in 2019. It is known that he started trading while starting high capital management.
In 2023, High-Flyer began DEEPSEEK as a laboratory dedicated to investigating AI tools separated from financial business. The laboratory is one of the investors and is a own company, also known as DeepSeek through a high flight.
From the first day, DeepSeek has built its own data center cluster for model training. But like other AI companies in China, Deepseek was influenced by US exports to hardware. In order to train one of the recent models, the company had to use the NVIDIA H800 chip, a chip version of the chip available by the US company.
DeepSeek’s technical team is called young. The company is reportedly actively hiring a Ph.D. researcher at China’s best university. DeepSeek also hires people without computer science backgrounds to help you understand the extensive topics according to the New York Times.
DeepSeek’s powerful model
In November 2023, Deepseek unveiled DEEPSEEK CODER, DeepSeek LLM and DeepSeek Chat (DeepSeek Coder, Deepseek LLM, and DEEPSEEK CHAT. The AI industry began notification until the DEEPSEEK-V2 model product was announced.
DeepSeek-V2, a general-purpose text and image analysis system, was well performed in various AI benchmarks and was much cheaper than similar models at the time. DEEPSEEK’s domestic competition, including Bytedance and Alibaba, has reduced some models and made other models completely free.
The Deepseek-V3, which began in December 2024, was added to the notoriousness of Deepseek.
According to the internal benchmark test of DeepSeek, DeepSeek V3 surpasses downloadable public models such as “Closed” models that can only be accessible through APIs such as META’s LLAMA and Openai GPT-4O.
The R1 “reasoning” model of Deepseek is also impressive. The DEEPSEEK, released in January, insists that R1 does not only do the O1 model of Openai on the main benchmark.
Since it is an reasoning model, the R1 effectively checks the facts and confirms the itself, which usually helps to avoid the traps that start the model. The reasoning model is slightly longer (usually a few seconds to a few minutes longer) to reach the solution compared to the general non -rational model. Conversely, there is a tendency to be more reliable in areas such as physics, science and mathematics.
However, other models of R1, Deepseek V3 and DeepSeek have disadvantages. Because of the AI developed by China, they allow China’s Internet regulatory agencies to benchmark the reaction “realize core socialism.” For example, in DeepSeek’s chatbot app, R1 does not answer questions about Tiananmen Square or Taiwan’s autonomy.
Destructive approach
If you have a business model in Deepseek, you can’t know exactly what the model is. The company sets products and services that are much lower than market value at prices and offer free to others. It also does not receive investor funds despite the interest of VC.
DeepSeek’s ways of talking and efficiency innovation allowed you to maximize cost competitiveness. However, some experts have a challenge to the value provided by the company.
Whatever the case, the developer took it to the model of the DeepSeek, which is generally understood, so it is not an open source but a commercially available allowable license. According to Clem Delangue, the CEO of HUGGing Face, one of the platforms that host DeepSeek’s model, developers of HUGGING FACE have created more than 500 “derivatives” models of 2.5 million downloads.
DEEPSEEK’s success for larger and more competitors is described as “upending AI” and “over hyped”. The company’s success was at least partially responsible for the NVIDIA’s share price decreased by 18% in January and to lead the public response to Openai CEO Sam Altman.
Microsoft announced that DeepSeek can be used by Azure AI Foundry Service, a Microsoft platform that provides AI services for companies under a single banner. Mark Zuckerberg, CEO of Mark Zuckerberg, said that when he asked about the effect of Meta’s AI expenditure during the first quarter, the spending on AI infrastructure would continue to be Meta’s “strategic advantage.” In March, Openai was called DeepSeek “National Subsidies” and “National Control”, and the US government recommends that the ban on DeepSeek is considered.
During NVIDIA’s four -quarter import call, Jensen Huang’s CEO emphasized DeepSeek’s “excellent innovation.” This says that the IT and other “reasoning” models are suitable for NVIDIA because they need much more computing.
At the same time, some companies prohibit Deepseek, and there are some nationwide and governments including Korea. The New York State also banned the use of Deepseek in the government device.
It is not clear about what the future of DeepSeek is. Improved models are given. However, the US government seems to be careful about recognizing harmful foreign influences. In March, the Wall Street Journal reported that the United States would ban Deepseeks on government devices.
This story was originally published on January 28, 2025 and is regularly updated.
