Governance of AI: The Need for Sovereign AI

Click the blue text

Follow us

Shen Xiangyang

Chairman of the Board of Hong Kong University of Science and Technology, Foreign Member of the National Academy of Engineering, USA

Governance of AI: The Need for Sovereign AI
On September 5, 2024, the internationally renowned Inclusion Bund Conference was grandly held in Shanghai. In the keynote speech at this conference, Shen Xiangyang, Chairman of the Board of Hong Kong University of Science and Technology and Foreign Member of the National Academy of Engineering, USA, delivered an important speech on artificial intelligence (AI), especially regarding AI agents. With the continuous development of AI technology, Shen pointed out that AI agents have become an important focus in the industry, and their disruptive power is gradually becoming evident.
However, technological advancement does not merely mean convenience; Shen emphasized the importance of AI governance in his speech. With the widespread adoption of AI systems, it is essential to recognize their potential risks and ethical issues. In his view, in the face of such disruptive power, building a responsible AI governance system is particularly urgent.
This requires countries around the world to work together to develop governance frameworks that are suitable for different national conditions and regions. Shen proposed that understanding the changes brought about by technology and its impact on society is key to achieving effective governance. He called for a combination of technological means and forward-thinking to achieve effective governance across borders.
Governance of AI: The Need for Sovereign AI

Thought 1: Computing Power is a Barrier

Since 2010 and 2012, the required computing power for all these models has been growing at an unprecedented scale. In the first few years, there was a growth of 6 to 7 times, but it has stabilized in recent years at about 4 times annually. Shen Xiangyang stated, “Today, to develop large models and deep learning, the most important thing is to have computing power.” Since 2010 and 2012, the computing power required for large models has been increasing at a rate of 6 to 7 times. In recent years, it has stabilized at about 4 times a year. “The growth of computing power is crucial; today’s large models have increasingly larger parameters, and thus the computing power requirements grow not just linearly but more like a square growth. Why? Because as models grow larger, they require more data to train. Therefore, the overall demand for computing power has been very high in recent years.”
“Talking about cards hurts feelings; no cards, no feelings.” Shen Xiangyang lamented, the entire development of the computer chip industry has shifted from Moore’s Law to Huang’s Law. If the GPU increases computing power by 4 times a year, and only doubles every 18 months, over ten years, it would amount to about 4 times 100. A yearly increase of 4 times results in a million times over a decade.
Thought 2: The Importance of Data
According to data, the training data for GPT-3 is 2 trillion tokens, and GPT-4 is around 12 trillion, and it continues to add more training data. Shen Xiangyang speculated that the training data for GPT-5 will be around 200 trillion. Currently, the entire internet cannot provide such a massive dataset, and more data must be mined to meet the model training requirements.Shen Xiangyang stated, “The 40 years of accumulated data on the internet seems to be for this AI moment.” Previously, as the core accumulation of the internet, most data was used by Google for search engines; in the future, this data will be used to train large models.
Thought 3: The Next Step for Large Models
Shen Xiangyang believes the path ahead is very clear, “In the future, we must move towards embodied intelligence, particularly towards robotics, with autonomous driving being a special form within robotics.” From the original language models and text, the next step is to develop multimodal capabilities; today’s multimodal Sora model is not yet very powerful, and its physical properties cannot be guaranteed, making it impossible to create a world model. Technically, we must unify understanding and generation.
Thought 4: The Impact of Large Models on the Industry
Returning to the demand for computing power, a general large model requires at least 10,000 GPUs, while future needs may require 10,000 high-performance GPUs; industry-specific models might require thousands; and a single company or enterprise aiming to create a large model may need hundreds. “However, what excites me is that the future is personal large models; for instance, Lenovo and Microsoft’s AIPC, and Apple’s claimed Apple Intelligence, are all moving towards the direction of Personal Intelligence,” Shen Xiangyang stated.
As of the end of July, the number of registered large models in China is 197, with about 30% being general models and 70% industry-specific models. The number of industry-specific models will continue to grow.
Thought 5: AI Agent – The Super Application
Shen Xiangyang stated, “Super applications are AI agents. Although ChatGPT is impressive and powerful, it is far from being an agent.” An agent significantly enhances human productivity. Up to now, GPT is still relatively a single-point breakthrough; to move forward, we need to establish workflows and perform industry analyses, connecting the entire framework of large model applications to platforms, knowledge skills, tasks, and dialogues to achieve such results. “A workflow must integrate skills and databases with all company data, paired with foundational large models to accomplish remarkable things. I am very confident in the future development of many Chinese companies in this regard.”
Thought 6: Emphasizing AI Governance
AI governance is very important, and countries have differing views. This year’s World Artificial Intelligence Conference (WAIC) focused on AI governance, highlighting the varied perspectives of different nations. The development of AI has had a tremendous impact on the public, companies, government regulation, and social development, raising public concerns about its safe governance.
“I believe that a crucial point in the future development of artificial intelligence, from the perspective of countries worldwide, is the need to establish sovereign artificial intelligence, which must be supported by a sovereign cloud to promote its development,” Shen Xiangyang stated.
Sovereign AI refers to a country’s ability to produce artificial intelligence using its own infrastructure, data, labor force, and business networks. Sovereign AI includes physical and data infrastructures. The latter includes sovereign foundational models, such as large language models. As artificial intelligence and accelerated computing become essential tools for addressing climate change, enhancing energy efficiency, and preventing cybersecurity threats, sovereign AI can play a key role in equipping each country to strengthen its sustainable development efforts.
Thought 7: Rethinking Human-Machine Relations
“How much of the impact brought by GPT is a shock to human-machine interaction, and how much is due to the development of machine intelligence?”Shen Xiangyang believes that we should rethink the relationship between humans and machines. He pointed out that AI provides a new context for humans to coexist with technology, and new ways of human-machine interaction point towards the integration of “AI and IA”. IA (Intelligent Augmentation) represents a human-centered AI development path that focuses on enhancing human capabilities rather than replacing them, emphasizing the collaborative relationship between humans and AI.
“John Markoff, a columnist for The New York Times, mentioned that in the journey of computer development over the past few decades, the real winners are those who excel in human-machine interaction. Regardless of the technology, the ultimate goal should be to help humans better utilize machines,” Shen Xiangyang stated. “In the AI era, the essence of human-machine interaction is dialogue, just like ChatGPT. Will ChatGPT combined with Microsoft become the greatest company of the AI era? I think only time will tell.”
Thought 8: The Essence of Intelligence
Today, the development of GPT is in full swing, but people’s understanding of intelligence is still very limited. Unlike physics, which can explain everything from the vast universe to the smallest quantum with a unified theory; today, many aspects of deep learning are inexplicable and lack robustness.
“The essence of intelligence is a century-long debate between neural networks and symbolic systems,” Shen Xiangyang stated. “Today, although the development of artificial intelligence is still in a relatively early stage, there are already many applications in the industry that are worth committing to. I am confident about the future development.”
Through Shen Xiangyang‘s speech at the Inclusion Bund Conference, we see the opportunities and challenges brought by artificial intelligence, especially AI agents. In the future, as technology continues to advance, AI agents will play an increasingly important role. Humans need to seriously consider our governance methods while enjoying the conveniences brought by these new technologies.
In the AI era, individuals, enterprises, and governments should clarify their roles and responsibilities and actively participate in AI governance. In the face of this world-changing force, building a responsible AI ecosystem will be an important topic for future development.

About Us

Institute for AI International Governance, Tsinghua University (THU I-AIIG) was established in April 2020 as a research institution at Tsinghua University. Relying on Tsinghua University’s existing accumulation and interdisciplinary advantages in artificial intelligence and international governance, the institute conducts research on major theoretical issues and policy needs in AI international governance, aiming to enhance Tsinghua’s global academic influence and policy leadership in this field, providing intellectual support for China’s active participation in international AI governance.

Sina Weibo: @Tsinghua University Institute for AI International Governance

WeChat Video Account: THU-AIIG

Bilibili: Tsinghua University AIIG

Source | This article is reprinted from Trusted AI, click “Read the original text” for more content.

Leave a Comment