Deepseek has become a viral.
After the chatbot apps were on the top of the Apple App Store chart (Google Play), China’s AI Lab Deepseek entered the mainstream consciousness this week. DeepSeek’s AI model is trained using high -calculation -efficient technologies, and whether the US analyst and engineer can maintain a lead in AI races, and maintains the AI chip. I wondered if it was.
But where did DeepSeek come from, and how did it rose quickly to international fame?
Deepseek trader origin
Deepseek is supported by HIGHFLYER CAPITAL MANAGEMENT, a quantitative hedge fund in China, which uses AI to notify transaction decisions.
AI enthusiast Liang Wenfeng jointly established high flyers in 2015. Wenfeng began to work on transactions while a student at Zhijiang university started HighFlyer Capital Management in 2019 and focused on the development and development of AI algorithms.
In 2023, HIGHFLYER started Deepseek as a lab that focuses on investigating AI tools different from the financial business. As one of the investors, there was a high flyer, so the lab rushed to our company, also known as DeepSeek.
From the first day, DeepSeek has built a unique data center cluster for model training. However, as with other AI companies in China, DeepSeek is affected by the ban on export in the United States. To train one of the recent models, the company has to use the H100 NVIDIA H800 chips, a very powerful version of the chips available by US companies.
Deepseek’s technical team is said to distort young teams. The company has been reportedly adopted a doctorate researcher from a top university in China. Deepseek also hires people who do not have a computer science background without computer science background, according to the New York Times, so that a wide range of subjects can be understood.
Deepseek’s powerful model
Deepseek announced in November 2023 the first set of the first model (Deepseek Coder, Deepseek LLM, Deepseek Chat). I started to attract attention.
Deepseek-V2, a general-purpose text and image analysis system, worked well with various AI benchmarks and was much cheaper than the same model at the time. It has been forced to reduce the price of some models of DeepSeek, including bytedance and Alibaba, and make others completely free.
The Deepseek-V3, released in December 2024, was added to the notorious of Deepseek.
According to Deepseek’s internal benchmark test, DeepSeek V3 exceeds both downloadable and open-available models, such as “Meta LLAMA”, which can only be accessed through APIs such as Openai GPT-4O.
Similarly, the Deepseek’s R1 “Progress” model is a model. Deepseek, released in January, claims that R1 will execute the O1 model of O1NAI on the key benchmark.
Since R1 is a reasoning model, there is actually a check itself. This helps avoid some pitfalls that normally stumble. Progress models take a little longer (usually a few seconds to a few minutes) to reach the solution compared to the normal non -rational model. The advantage is that it tends to be more reliable in domains such as physics, science, and mathematics.
However, other models of R1, Deepseek V3, and Deepseek have drawbacks. Because it is an AI developed by China, they are subject to benchmarks by China’s Internet regulatory authorities, guarantee that their response will “embodies core socialist value.” For example, in DeepSeek’s ChatBot app, R1 does not answer questions about Tiananmen Square or Taiwan’s autonomy.
Destroyed approach
If DeepSeek has a business model, it is not clear what the model is exactly. The company offers the price of products and services far below the market value and provides free to others.
According to DeepSeek, the efficiency breakthrough has maintained extreme cost competitiveness. However, some experts are dismissing the numbers provided by the company.
In any case, the developer adopts the model of DeepSeek. This is not an open source because the phrase is generally understood, but can be used under a generous license that enables commercial use. According to Clem Delangue, one of the platforms that hosts Deepseek models, CLEM DELANGUE, the CLEM DELANGUE, a CLEM DELANGUE, created more than 500 R1s of R1, which combines 2.5 million downloads. 。
Deepseek’s success in more established rivals is said to be “maintaining AI” and “overhip”. The company’s success was at least partially responsible for at least a decrease in NVIDIA’s stock prices on Monday by 18 % and Openai CEO’s public response from Sam Altman.
Microsoft has announced that DeepSeek can be used in Azure AI Foundry Service. This is a Microsoft platform that connects AI services for companies under a single banner. When asked about DeepSeek’s influence on AI spending in meta in the first quarter, CEO’s Mark Zuckerberg stated that spending on AI infrastructure will continue to be a “strategic advantage.”
It is not clear how DeepSeek’s future is kept. The improved model is given. However, the US government seems to be alert to what is recognized as harmful foreign influence.
TechCrunch has a newsletter focusing on AI! Sign up here and get it on the reception tray every Wednesday.
This story was originally released on January 28 and will continue to update more information.