The recent DeepSeek can be built with low-cost, non-advanced process acceleration hardware.Large-scale natural language models with performance comparable to most technology companiesIBM Chairman and CEO Arvind Krishina responded earlier that DeepSeek's approach also confirms IBM's earlier research, which believes that the cost of artificial intelligence technology optimized for specific needs can be reduced by up to 30 times. It is also expected that this approach can make artificial intelligence more easily popularized and drive higher computing efficiency.
Amazon CEO Andy Jassy also pointed out that as DeepSeek's approach has attracted market attention, it is expected that the cost of applying artificial intelligence technology will be greatly reduced in the future, allowing more companies to easily introduce artificial intelligence technology into their application services, while also increasing market demand for artificial intelligence technology.
At the same time, Andy Jassy believes that even if the cost of applying AI technology decreases, it does not mean that companies will reduce their investment in AI development. On the contrary, it can restart development projects that were previously restricted by the high cost of introducing AI technology, which is expected to increase the overall market demand for AI technology.
Prior to this, OpenAI CEO Sam Altman praised DeepSeek for its recent attention, but also emphasized thatWe will continue to invest in the development of large-scale natural language models, and emphasized that computing power remains important in promoting the development of artificial intelligence technology.
Other opinions on DeepSeek include former Google CEO Eric Schmidt's belief that the emergence of DeepSeek means that Chinese companies can compete with large American technology companies with fewer resources. He also called on the United States to expand its development of open source AI models to gain an advantage at a turning point in the global AI technology competition. Former Intel CEO Pat Gelsinger also mentioned the importance of open source in his views on DeepSeek, and believed that DeepSeek would help reshape the increasingly closed AI model building process.
Furthermore, AI guru Fei-Fei Li's team recently used Alibaba Tongyi Qianwen's Qwen2.5-32B-Instruct model as a foundation and performed a distillation process using Google DeepMind's Gemini 2.0 Flash Thinking experimental version to create the more lightweight S1 model. The entire training process utilized only 16 NVIDIA H100 accelerators and took a total of 26 minutes. Furthermore, the team emphasized that leasing the relevant computing infrastructure through a cloud service platform cost less than US$20, bringing the total cost of the entire training process to under US$50.
This means that the future development of artificial intelligence technology will be more clearly stratified. Large technology companies will continue to invest in large-scale natural language models with higher performance and continue to move towards the goal of creating general artificial intelligence technology (AGI). More applications will tend to introduce artificial intelligence technology produced through customization and optimization, thereby enabling artificial intelligence to be operated and used in a more efficient and affordable manner, while also accelerating the overall popularization and development of artificial intelligence technology.
However, DeepSeek's rapid popularity has also brought with it some problems. In addition to attracting the attention of governments in countries such as the United States, there are even suspicions that its model contains backdoors related to the Chinese government, and therefore may have been ordered to be banned. At the same time, many "counterfeit products" have also appeared, trying to steal users' personal information and money. There has also been a significant increase in the number of cases where open source models are "re-skinned" and claimed to be self-made high-performance artificial intelligence models to attract users.
