The emergence of the Chinese application Deep Seek has raised questions about the boundaries of global competition over the development of artificial intelligence technology. Over the past decade, large language models (LLMs) have seen a significant qualitative leap in their capabilities. In 2017, a research paper titled “Attention is All You Need” was published by eight researchers at Google, which introduced the attention mechanism relied upon in developing large language models. This became clearly evident to the public with the unveiling of the ChatGPT application in late November 2022. With the emergence of several American chatbots, questions began to arise about China’s position in these developments, especially as AI advancements became a central issue in the competition between the U.S. and China. Despite the presence of various discussions, data interpretations, and writings regarding the advancement in AI research and applications, with many doubting Beijing’s ability to achieve the same level of progress as the U.S., the launch of Deep Seek confirmed that China had made significant strides in the field of AI, creating a sense of chaos in the American AI market.
Key Features
Some experts in artificial intelligence technology, such as Aravind Srinivasan, CEO of Perplexity AI, described the Chinese application “Deep Seek”—especially its logic-based model R1—as the biggest technological transformation humanity has seen since the advent of the internet. This is attributed to several distinguishing features of the application compared to others in the field, which can be summarized as follows:
Prioritizing Open Source Models: The Chinese application relies on a different research model from the Western approaches to large language models. The release of the “Deep Seek V3” and “Deep Seek R1” models has opened many people’s eyes globally to what is actually happening in AI in China. This model demonstrates the research team’s reliance on various studies and techniques distinct from those used in the U.S. Interestingly, the reference standards for this model are accessible to everyone, being open source, which means any developer can download the model, host this API on their own server, and test it to verify its claimed capabilities and features, including comparing its results with those of other available models.
Developed with Relatively Low Costs: Another striking aspect of this model is its relatively low development cost, which pales in comparison to the production costs incurred by similar Western models, representing another radical shift in the AI field as a whole. While the exact cost of launching and developing this application is not precisely known, experts affirm that it does not approach the development costs of Western models, with its cost estimated at around $10 million, whereas similar models in the U.S. can cost hundreds of millions of dollars.
Expansion through Distillation Technology: In addition to cost challenges, large language models employed in the West consume substantial amounts of computing power and infrastructure. Open-source language models, whether LLaMA models or others in the U.S., rely on a process known as distillation. This process simply utilizes a very large model to assist smaller models aiming to perform specific tasks (such as developing a language model to aid lawyers in understanding all laws and regulations related to various cases). This is highly cost-effective, allowing these teams to use minimal APIs, downloading the open-source model and training smaller models with it. “Deep Seek” applied this logic at a broader and deeper level.
Utilizing the Expert Mixture Mechanism: Chinese research teams involved in developing the application employed a new engineering method called the “expert mixture.” In this approach, China divided its large language model into different specialized segments, with the main model receiving user commands and redistributing these commands to the relevant specialized segments. For example, if a question pertains to law, it is quickly directed to the legal specialist model for a response.
Incorporating Distinct Software Services: The features of the Chinese AI model “Deep Seek” extend beyond its exceptionally low production cost compared to other American language models and its free availability to all users. Its advantages are particularly evident in technical aspects related to programming and data analysis. For instance, when asked to create a website interface using HTML, the program, similar to its closest competitor “ChatGPT,” writes the complete code for the site, allowing users to test the code and see the final result. In contrast, ChatGPT does not provide this feature and requires users to utilize an external text editor or platform to test the code and display the final result.
Employing Reinforcement Learning: Some reports suggest that Deep Seek’s models rely on a method known as reinforcement learning, enabling them to learn through trial and error, interact with their environment, and incorporate user feedback and assessments. This approach allows them to evolve and gain experiences mimicking human behavior, unlike other models that depend on developer-directed learning based on their instructions and data inputs.
Major Implications
The U.S. and China are engaged in a significant technological competition often referred to as a war. Among its prominent elements is the U.S. imposing a strict ban on transferring all American technology involved in semiconductor manufacturing, a prohibition that Japan and the Netherlands also participated in. With the revelation of this application, China underscores the ineffectiveness of American measures to prevent Beijing from accessing advanced technologies, especially in AI, which has emerged as a top priority for the new U.S. administration. The key short-term and long-term implications of the unveiling of this application include the following:
Sharp Decline in the U.S. Stock Market: On January 27, the U.S. stock market experienced a sharp decline in shares, particularly in technology companies. For instance, shares of Nvidia, the company behind the advanced chips dominating many AI investments, dropped by 17%, losing nearly $600 billion in market value and being the most affected. Investors rushed to sell their shares to reassess their investment decisions following China’s revelation of a superior and lower-cost large language model. Although American tech stocks stabilized the following day (January 28), with Nvidia shares rebounding by over 6% by midday trading, experts noted that the wave of selling AI stocks in the U.S. could be an overreaction; however, the impact of the application on the American tech market and its leaders remains difficult to quantify at present.
Increased Doubts About Washington’s AI Dominance: American President Donald Trump described the moment the Chinese application was revealed as a significant wake-up call for the American tech industry but also indicated it might ultimately be a “positive” for the United States. He justified this by stating, “If you can do the same thing at a lower cost and get the same outcome, I think that’s good for us.” He added that he is not concerned about this breakthrough, emphasizing that the U.S. would remain a dominant player in this field.
Accusations Against the Chinese Application of Copying OpenAI Models: On January 29, OpenAI, the owner of the ChatGPT application, accused Chinese companies and others of attempting to copy its AI model, calling for enhanced cooperation with American authorities and security measures. OpenAI emphasized that such actions contradict its usage terms and that it is working to find ways to identify and prevent future attempts in this regard.
Skepticism About the Effectiveness of U.S. Sanctions on China’s Tech Sector: The specific nature and type of technology that China relied upon to develop this model remain unclear, especially since graphic processing units (GPUs), fundamental to the operation of large language models, are primarily produced by Nvidia, which is under export bans to China. Experts differ in their conclusions on this point; some believe the Chinese model may have utilized different GPUs available in China, noting that AI research labs sometimes gain access to the latest advanced units. Others argue that China may have stockpiled American GPUs before the imposition of the export bans. Nonetheless, the Chinese model suggests that experts there have managed to use the available devices and technologies more efficiently, indicating the inefficacy of American restrictions on China’s tech industry and threatening Washington’s dominance in this area.
Heightened Global Competition in AI: Naturally, the launch of Deep Seek and its various models, which rely on the principles of open source and free availability, will escalate global competition in the field of artificial intelligence. The United States will likely push to counter these Chinese advancements, especially as Trump aims to enhance U.S. standing in AI. It is also expected that American companies will ramp up their investments in this area. Conversely, China will continue its initiatives to bolster its technological influence, potentially prompting other powers to develop new applications.
In conclusion, this revelation is anticipated to enhance technological competition between China and the United States; however, it may also support the efforts of several companies aiming to develop their existing AI technologies lacking sufficient funding and capabilities for such advancements. This model also raises questions about the level of development in this field in other countries, especially Europe and Russia, and whether these nations will expend more effort and funds to develop related research or merely remain spectators while China and the U.S. dominate the field entirely.
thank you