The artificial intelligence community is intensively discussing Deepseek R1, a new open -source reasoning model developed by the Chinese venture Deepseek. Deepseek argues that the R1 is equal in many basic criteria with the OpenAI’s chatgpt O1, even leaving it behind in some areas, and it is realized at a much lower cost.
Hancheng Cao, an assistant professor at the University of Information Systems at the University of Emory, states that this innovation can create an equal opportunity for researchers and developers with limited resources.
Deepseek’s success becomes even more remarkable considering the effects of the US strict controls on high -tech chip exports on Chinese artificial intelligence companies. However, the existing indicators reveal that these restrictions do not have the expected consequences. Instead of preventing China’s artificial intelligence capabilities, sanctions seem to have directed initiatives such as Deepseek to more efficient use and cooperation -oriented innovations.
Deepseek had to optimize the educational processes in order to alleviate the burden of Nvidia for the Chinese market, while educating the R1 model. Former Deepseek employee, a PhD student at Northwestern University, elaborated the difficulties in the process.
The R1 is praised by researchers for its superior success in complex logic problems such as mathematics and coding. The model draws attention with its ability to solve problems step by step by using an approach called “chain of thought ..
Dimitris Papailopoulos, the chief researcher of Microsoft’s AI Frontiers research laboratory, says the engineering simplicity in the R1 model is the most affected. “Deepseek focused on directing the right solutions, rather than focusing on processing each stage in detail, and this has decreased significantly reduced with high efficiency,” he says.
Deepseek has also released six different R1 versions, which are small enough to work on laptops. The company claims that one of these versions has crossed the O1-Mini model of OpenAI in certain tests. Perplexity CEO Aravind Srinivas argues that Deepseek has largely copied O1-Mini and makes it open. However, Deepseek did not comment on this claim.
Liang Wenfeng, the founder of Deepseek, studied information and electronic engineering at the University of Zhejiang and founded a company in Hangzhou in 2023. Liang had previously established a hedge fund called High-Flyer in 2015. Now, it aims to create artificial general intelligence (AGI), such as Sam Altman, CEO of OpenAI.
According to Latepost, a Chinese media organization, Liang has a forward -thinking behind Deepseek’s success. Liang, who made a large stock from the Nvidia A100 chips before the US sanctions, was able to optimize these resources with low -power chips. According to some estimates, the company’s stock has between 10,000 and 50,000 chips.
Deepseek is one of the rare initiatives that survived in a market dominated by giant technology companies Alibaba and ByTedance. It also differs with its continuation of its activities without receiving foreign investment.
Mind Wang states that he has access to plenty of information processing during his work in Deepseek and that he can do freely experiments. This is a rare luxury, especially for new graduates.
Liang acknowledges that Chinese companies face inefficiency in artificial intelligence engineering. However, with his team, he says that they are trying to close these gaps by making improvements in memory usage and calculation speed.
Chinese companies have begun to give importance to open source principles as well as productivity in recent years. Alibaba Cloud has released more than 100 open -source AI models for 29 languages and coding and mathematics. Other initiatives made their models openly.
These developments allowed China to take second place in the Global Artificial Intelligence race. Artificial Intelligence Researcher Matt Sheehan says the US’s restrictions on chip exports have pushed Chinese companies to use resources more efficiently.
Deepseek draws attention as an example that continues to produce innovation even under these pressures.