Xue Lan: most LLMs in China are "encapsulated" and computing power is also being "choked"

On April 1st, during the recent Academic Annual Conference on the Development and Governance of China's Digital Economy (2024) （中国数字经济发展和治理学术年会）, Professor Xue Lan (薛澜), the Dean of Schwarzman College at Tsinghua University and the Dean of the Institute for International Governance of Artificial Intelligence, delivered a keynote speech titled "The Governance Challenges of Artificial Intelligence Development."

In his speech, Xue Lan mentioned that the latest development trend of AI in China is generally stable, with over 130 large language models currently in the country. Although there has been significant progress in terms of quantity, he pointed out that there are still many issues with China's LLMs. He noted that many of these models are constructed through a process of "encapsulation"（套壳） and assembly（拼装）, and that they also face significant challenges in computing power, colloquially referred to as being "choked" in terms of their operational capabilities.

"Currently, many foreign LLMs are open-source. Thus, by encapsulating these open-source bases, a 'shell' LLM can be created, and then assembling some of these LLMs together forms an even larger language model. However, the originality behind these LLMs created in this manner is limited. Additionally, our computing power is being 'choked.' For instance, the sale of NVIDIA's GPUs, A100 and H100, has been banned by the United States, affecting many companies. Furthermore, the quality of data in China is relatively low, and it has not truly been industrialized. There are relatively few standardized data service providers, and issues such as short-sightedness in Chinese capital are all problems that need to be thought about and addressed in the next step of the development of China's LLM industry.”

This is a rare recent mention by Xue Lan of some of the issues such as encapsulation, computing power, and technological gaps between China and the United States in the AI industry.

Xue Lan noted that since the emergence of ChatGPT in November 2022, technological innovation in AI has accelerated, particularly among leading American technology companies such as OpenAI, which has created a competitive environment. This year, OpenAI launched a very unique, cross-modal video generation tool called Sora, which can produce videos up to 60 seconds long, astonishing the world. This innovation is expected to have a significant impact on the short video and many other media entertainment industries. At the same time, different sectors like autonomous driving, advertising, education, healthcare, and security are also greatly affected and could become potential markets for Sora.

"Moreover, the United States has made many new breakthroughs in core hardware for AI computing power, such as the Groq LPU, a new large model inference chip. Although its manufacturing process is still at 14nm, its specific processing capacity is very fast due to the TSA architecture, with inference speed 10 times faster than GPUs, but with only 1/10 of the energy consumption."

Xue Lan frankly stated that as US companies like Groq, NVIDIA, and OpenAI continue to announce breakthroughs in multiple new AI technologies, the gap in LLM technology between China and the US has widened, posing a significant challenge to China's innovation system.

"The current innovation model in China still follows a relatively traditional linear pattern, starting from academic research to engineering technology, and finally to products. This model fails to create an inseparable ecosystem among research, engineering, and the market, which is an urgent issue that China's innovation system needs to address." "The latest development trend of AI in China is generally stable, with China publishing the most AI papers globally, followed by the United States, and there are also many papers co-authored by China and the US. In terms of AI large models, there are now more than 130 large models. If we talk about progress in terms of quantity, it is significant, but in reality, our large models still have many problems, as many are constructed through 'encapsulation' and assembly methods. Moreover, our computing power is being 'choked'.

At the same time, the low quality of data in China is also a problem. China has a large volume of data, but it has not truly been industrialized, and there are relatively few standardized data service providers. This is because big data services are not profitable, public data enterprises are unwilling to clean the data, and customized services generally charge higher fees. Therefore, how to construct the data market is also an issue that needs to be addressed.

Another problem is short-sightedness in capital. There are various reasons for capital short-sightedness, one of which is concerns about the instability of China's capital market policies. Additionally, there are challenges in commercialization and so on. All these are issues that need to be considered and addressed in the next step of development for China's large model industry.

Generative AI has a very strong permeability in terms of comprehensive social benefits, and its impact on society is comprehensive. For the manufacturing industry, AI is conducive to promoting the digital and intelligent transformation and upgrading of the industry, with great potential in advancing new productive forces and high-quality development. China has already made many explorations in this area. For example, after applying AI technology, the manufacturing efficiency of Sany Heavy Industry's factory increased by 185%, and the production cycle was reduced from 30 days to 7 days. At the same time, AI has also played a huge role in promoting the scalability of knowledge-intensive service industries.

However, the potential risks of AI are manifested on multiple levels. At the individual level, AI may lead to algorithmic discrimination, intellectual property disputes, misinformation, etc., which are related to the technology itself. There are also issues in the application and development process, including data privacy protection and the prevention of forgery. At present, we are more concerned about these direct impacts, but there are some longer-term major issues at the social level, such as the adjustment of the labor force structure and industrial structure. Meanwhile, due to the adjustment of the labor force structure, there may be a significant impact on income distribution, thereby exacerbating the digital divide, social injustice, and other issues, which all need to be carefully considered. In addition, the energy consumption of training large models is very severe, and the greenhouse gas emissions added now seem to be non-negligible. It is estimated that the carbon emissions of ChatGPT during its entire training cycle are equivalent to the total amount of driving to the moon and back to Earth.

The issues at the national security level are equally worrying, including the impact on social stability and military applications.

In terms of higher-level risks, what we are all very concerned about is whether General Artificial Intelligence (AGI) could become uncontrollable one day? Could it threaten the survival of human society? That is, what conflicts might arise between silicon-based and carbon-based civilizations in the future? The risks at this level concern the survival of human society. With the emergence of ChatGPT, the entire society, especially the tech community, has been impacted, with many experts beginning to pay close attention to this issue. In 2020, some experts in the tech field had differing views on AI governance, believing it was too early to discuss governance as AI technology was just beginning to develop. Besides, we weren't as concerned about the potential risks of so-called AGI at that time. However, the situation changed significantly after the appearance of ChatGPT. Recently, I participated in a joint effort with some renowned AI technology and governance experts to write a paper submitted to 'Science'. This paper points out that we have reached a critical stage where it is imperative to push for global governance of AI. The reason these experts think the situation has changed is that the emergence of ChatGPT has led them to believe that the potential arrival of AGI could be much sooner than expected."

AI has numerous applications in fields such as e-commerce, healthcare, education, and scientific research, potentially offering tremendous social benefits. However, it also poses many risks that we need to explore how to effectively mitigate. Risks generated by the technology itself are typical examples. For instance, when we ask ChatGPT for the lyrics of the Tsinghua University anthem, it might fabricate an incorrect answer, confusing information between Peking University and Tsinghua University. Besides, there are many other risks including data misuse, privacy breaches, deep fakes, etc. In the specific context of e-commerce, some platforms may engage in false advertising and "price discrimination" based on big data. Furthermore, the future development of the AI field also needs to comprehensively consider its social effects, including its role in shaping new forms of human civilization.

New Generation AI Development Plan" in 2017, China has formulated a three-stage strategy for AI development, within which AI governance is a crucial component. The plan explicitly calls for attention to AI's potential risks and the task of building an effective governance system. After 2020, China accelerated its pace in AI governance.

China's AI has strong accumulation and early-mover advantages, forming a governance framework that combines top-down and bottom-up approaches. AI technology also has a very broad range of application scenarios. However, challenges such as creating a fair competitive market environment for different companies, building an industry ecosystem, and attracting top global AI talents must be addressed to promote the healthy and orderly development of the AI industry."

In addition to the aforementioned comments, Xue Lan recently also stated at the "China Development Forum 2024" during the seminar on AI development and governance that AI governance faces four major challenges. He suggested strengthening international cooperation and addressing geopolitical issues as part of the efforts to navigate these challenges.

"The discrepancy between policy formulation and technological development pace poses a challenge to AI governance. Development of AI is very rapid, but changes to the governance system are relatively slower—a law or regulation needs to go through a series of arguments and considerations from various aspects.

Additionally, the information asymmetry between the government and businesses also adds to the difficulty of AI governance. Traditional governance involves the government regulating businesses, and the relationship is somewhat like a "cat and mouse" game—businesses try to evade government regulations, while the government seeks to identify issues within businesses. However, in reality, both parties have many blind spots. Businesses may not know what issues are of most concern to government regulations, and the government may also be unclear about the potential risks brought by technological development. Thus, information asymmetry, or sometimes mutual ignorance, often occurs.

Another challenge in AI governance is the cost asymmetry in the process of risk regulation. For example, the cost of misusing or abusing AI technology is quite low, but preventing such risks is very difficult, with the required cost far exceeding the potential harm it might cause.

AI governance is a global issue that cannot be resolved by any single country, international organization, or enterprise alone. Professional organizations, international organizations, foundations, and businesses all have a certain interest in governing this issue, and they even have the capability and conditions to do so. However, there is no hierarchical relationship among all these organizations.

Is the United Nations the biggest? But the United Nations may not be able to manage many businesses. The Chinese government can regulate Chinese businesses, but it cannot regulate American businesses. A series of institutions have overlaps, contradictions, and related interests, all wanting to participate in governance. This makes forming a global governance system very difficult.

Currently, governments worldwide are improving the regulation of generative AI. In August 2023, seven departments announced the implementation of the "Interim Measures for the Management of Generative Artificial Intelligence Services." In October 2023, the White House in the United States issued an AI regulatory order, requiring developers of the most powerful AI systems to notify the federal government when training models. In November 2023, the world's first AI Safety Summit was held in Bletchley Park, UK, where 28 countries and regions, including China, the United States, and the European Union, signed the Bletchley AI Safety Declaration. In Europe, the "AI Act" came into effect in March 2024.”

In response to the aforementioned four major challenges, Xue Lan has put forward a series of suggestions. These include strengthening security and technological research and development, particularly emphasizing the need for international cooperation. He also encourages self-regulation by enterprises, suggesting they learn from the self-regulatory mechanisms of industries like international civil aviation and nuclear power stations.

Xue Lan also specifically mentioned geopolitical issues. He pointed out that U.S. Secretary of State Antony Blinken described the U.S. technology policy towards China as "small yards, high fences," but raised the question of what truly constitutes a "small yard" in the context of technology. Xue Lan proposed that to genuinely solve the various problems facing humanity, there must be enhanced cooperation in the AI field between China and the United States.

“In 2023, China proposed the "Global Artificial Intelligence Governance Initiative," systematically expounding China's approach to AI governance in terms of AI development, safety, and governance. This includes advocating for the establishment of an AI risk-level testing and assessment system to continuously improve the safety, reliability, controllability, and fairness of AI technology. It supports forming a global AI governance framework and standards with broad consensus based on full respect for the policies and practices of all countries and supports discussions under the United Nations framework to establish an international AI governance institution.

Overall, China's AI has a solid foundation and early-mover advantages, forming a general plan for AI development. In the past development process, it has also formed a governance framework that combines top-down and bottom-up approaches, and AI technology has a very broad range of application scenarios. At the same time, the future development of China's AI also faces some challenges.

The first is how to form a market environment of fair competition among different enterprises, including private enterprises, state-owned enterprises, foreign-funded enterprises, and enterprises of all sizes. This diverse and fair competitive market environment is crucial for the development of the AI field; the second is the industry ecosystem issue, how to establish an industry ecosystem where enterprises, funds, talents, and other subjects and resources can circulate and coordinate effectively. The third is governance issues, how to form a predictable, inclusive, prudent, and agile governance framework to lay the institutional foundation for a market environment and industry ecosystem that encourages AI development; finally, how to gather top global AI talents. OpenAI has developed from the gathering of excellent talents from different countries around the world, with a high degree of talent diversity. China needs to increase its openness in this regard.

With the joint efforts of all parties, China's AI industry can seize development opportunities, overcome numerous difficulties, meet challenges from various aspects, effectively prevent and control various risks, and let the development of AI serve China's modernization construction and the peaceful development of the world.”