Deepseek has spread.
The Chinese AI Laboratory Deepseek introduced the top (and Google Play) of the Apple App Store this week, which broke into the mainstream consciousness this week. DeepSeek's AI model uses effective technical training. It has led Wall Street analysts and technicians to question whether the United States can maintain its leading position in the AI competition and whether the demand for AI chips can be maintained.
But where does DeepSeek come from, and how does it be so quickly ascended into the international reputation?
DeepSeek was supported by high flight capital management. This is China's quantitative hedge fund, which uses AI to inform them of transaction decisions.
AI liang wenfeng jointly founded High-Flyer in 2015. According to reports, Wenfeng began to get involved in transactions, and a student of Zhejiang University launched a high -level capital management as hedge fund in 2019, focusing on development and deployment of AI algorithms.
In 2023, High-Flyer started with DeepSeek and is a laboratory dedicated to studying the separation of AI tools from financial business. The laboratory uses High-Flyer as one of its investors and rotates into its own company, also known as Deepseek.
Starting on 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 the US export ban. In order to train one of its latest models, the company was forced to use the NVIDIA H800 chip, which is a small version that American companies can use the chip H100 of the US company.
It is said that Deepseek's technical team is biased towards Young. According to reports, the company has actively recruited PhDs from AI researchers from top universities in China. Deepseek also hires people with no computer scientific background to help their technology better understand the New York Times.
DeepSeek launched its first set of models-Deepseek encoders, Deepseek LLM and Deepseek Chat in November 2023. Start paying attention.
DeepSeek-V2 is a general text and image analysis system that performs well in various AI benchmark tests-and running is much cheaper than those comparable models at the time. It forces DeepSeek's domestic competition, including ByTedance and Alibaba, reducing the use price of certain models and making others completely free.
DeepSeek-V3 was launched in December 2024, and only added DeepSeek's infamous.
According to Deepseek's internal benchmark test, Deepseek V3 is better than available, open and available models, such as Meta's LLAMA and "closed" models, and can only be accessed by API, such as OPENAI's GPT-4O.
Equally impressive is the R1 "reasoning" model of DeepSeek. DeepSeek claims that R1 was released in January, and the O1 model of Openai was tested in critical benchmarks.
As a reasoning model, R1 effectively conducts factual inspections, which helps it avoid some traps that usually trip models. Compared with typical non -disputed models, the reasoning model takes longer (usually longer to minutes) to achieve solutions. The advantage is that they tend to be more reliable in the fields of physics, science and mathematics.
However, other models of R1, Deepseek V3 and Deepseek have a disadvantage. As the development of China's development, they are tested by the Chinese Internet regulatory agency to ensure that they react "reflect the core socialist values." For example, in DeepSeek's chat robot application, R1 will not answer questions about Tiananmen Square or Taiwan's autonomy.
If Deepseek has a business model, it is unclear what the model is. The company's price of its products and services is much lower than market value and provides other people for free.
DeepSeek's way of telling is that efficiency breakthroughs allow it to maintain high cost competitiveness. However, some experts raised the number provided by the company.
In any case, developers use DeepSeek model. This model is not open source because the phrase is usually understood, but it can be obtained under the loose permission of allowing commercial use. Clem Delangue, chief executive of Hugging Face, said that one of the platforms for hosting DeepSeek model, developers on Hugging Face have created more than 500 R1 "derivative" models, which have added 2.5 million downloads.
Deepseek is described as "raising artificial intelligence" and "super propaganda" for the greater and more mature competitors. The company's success was at least part of the company's stock price fell 18 % on Monday and caused public response from OPENAI CEO Sam Altman.
Microsoft announced that DeepSeek can be provided on its Azure Ai Foundry Service (Microsoft platform), which summarizes AI services as corporate services under a single banner. When asked about the impact of DeepSeek on the AI expenditure of Meta in the first quarter, CEO Mark Zuckerberg said that the expenditure on AI infrastructure will continue to be META's "strategic advantages".
As for the future of Deepseek, it may be unclear. The improved model is given. However, the US government seems to be more and more vigilant about its harmful foreign influence.
TechCrunch has a news communication focusing on AI! Register here to obtain it in the box every Wednesday.
This story was originally released on January 28, and more information will be updated continuously.