Currently, DeepSeek is the only superior company company in China that does not count on funding from technology giants like Baidu, or Bytedance. Eager Young Genius group proves self-esteem, when he join the DeepSeek research team. , they are not looking for experiences that have experience to build products facing consumers. Instead, he focuses on the PhD student from the highest university in China, including the University of Peking and Tsinghua University, who wanted to prove himself. Many have published in the highest journal and winning the award in international academic conference, but do not have industrialized community of technology QBitai. , “Liang to 36KR in 2023. An orthodox strategy is a lot of free companies’ strategies. Job is a very different operation of the Internet company located China, in Where teams often compete for the resource. Most people, when they are young, they can adopt the mission without the experimental consideration of the world. “The fact that this young research is almost all study in China adding the drive , said the experts. “This generation is also containing the paintism, especially when passing the US limits and a critical point of technology,” Zhang’s obvious software. “The determination to overcome the obstacle does not only reflect the personal ambition but also a vast commitment to raise Chinese positions as global innovation leadership.” The company of accessing the cutting-edge chips like NVIDIA H100. The step gives you a problem for deepseek. The company begins with Stockpile 10,000 H100, but it takes more to compete with the company like the opening and meta. “Our problems are never funding, but the export control over the advanced chips,” pursuing 36KR in the second interview of 2024. DeepSeek must come up with a more efficient way to sport its model. “She optimized the roller architecture using the trick-sloma battery technique among the chips, reduced the size of the model, and the use of innovative mixed model,” said Wendy Chang, software engineer that is a policy. Analyst in Mercute Institute for China Studies. “Many of these approaches are not new ideas, but consciously join successfully to produce advanced models as an incredible achievement.” DeepSeek has also made significant progress on multi-head lantront attention (MLA) and Mixture-of-Experts, two technical design that creates more expenses to require fewer computing resources. In fact, DeepSeek latest models are definitely efficiently so requires a sefficity of the Llama model computing power 3.1 Meta number of trained, according to the epoch ai research institution. Global Research AI Community. For many AI China companies, developing the Open Source model as the only way to play with western partners, as they can attract more users and contributors, which then helps the model grow. “They have now showed that the latest model is built by low money, even if the model number of models today make a lot of space for optimizing,” says Chang. “We will always see more attempts in this direction.” The news can cause problems to the US export control now that focuses on creating computing resource resources. “Estimates are now about the power of ai computation owning China, and what can be replaced, can be changed,” said Chang.
