Tao Lake Think Tank | Sun Ninghui: The Development of Artificial Intelligence and Intelligent Computing

On the morning of April 26, the ninth meeting of the Standing Committee of the 14th National People’s Congress concluded at the Great Hall of the People in Beijing.

After the closing meeting, the Standing Committee of the 14th National People’s Congress held a special lecture, presided over by Chairman Zhao Leji. Academician Sun Ninghui from the Chinese Academy of Engineering and a researcher at the Institute of Computing Technology, Chinese Academy of Sciences, delivered a lecture titled “The Development of Artificial Intelligence and Intelligent Computing”.

Recently, the official website of the National People’s Congress of China published the full text of Academician Sun Ninghui’s ten-thousand-word manuscript, let’s take a look.

The Development of Artificial Intelligence and Intelligent Computing

Sun Ninghui

Chairman, Vice Chairmen, Secretary General, and all members:

The field of artificial intelligence has recently witnessed an explosive development led by generative artificial intelligence large models. On November 30, 2022, OpenAI launched an AI conversational chatbot, ChatGPT, whose outstanding natural language generation capabilities attracted widespread attention worldwide, surpassing 100 million users in just two months. This sparked a wave of large models both domestically and internationally, with various models such as Gemini, Wenxin Yiyan, Copilot, LLaMA, SAM, and SORA emerging like mushrooms after rain, making 2022 known as the year of large models. The current information age is rapidly entering the development stage of intelligent computing, with continuous breakthroughs in artificial intelligence technology gradually empowering various industries, making artificial intelligence and data elements typical representatives of new productive forces. General Secretary Xi Jinping pointed out that the new generation of artificial intelligence should be regarded as a driving force for promoting leapfrog development in science and technology, optimizing industrial upgrades, and enhancing overall productivity, striving to achieve high-quality development. Since the 18th National Congress of the Communist Party, the Central Committee with Comrade Xi Jinping at its core has attached great importance to the development of the intelligent economy, promoting the deep integration of artificial intelligence and the real economy, injecting strong momentum into high-quality development.

1. Overview of the Development of Computing Technology

The history of the development of computing technology can be roughly divided into four stages. The invention of the abacus marked the entry of humanity into the first generation—the mechanical computing era. The second generation—electronic computing was marked by the emergence of electronic devices and electronic computers. The advent of the internet ushered us into the third generation—network computing, and currently, human society is entering the fourth stage—intelligent computing.

Early computing devices were manual auxiliary computing devices and semi-automatic computing devices. The history of human computing tools began with the Chinese abacus in 1200 AD, followed by the invention of Napier’s bones (1612) and the wheel-based adder (1642), leading to the birth of the first automatic computing device capable of performing four arithmetic operations—the stepper calculator in 1672.

During the mechanical computing period, some basic concepts of modern computers had already emerged. Charles Babbage proposed the designs for the difference engine (1822) and the analytical engine (1834), supporting automatic mechanical computation. During this period, the concepts of programming and programs were basically formed. The concept of programming originated from the Jacquard loom, which controlled the printing pattern through punched cards, eventually evolving into the form of storing all mathematical computation steps through computational instructions; the first programmer in human history was Ada, the daughter of poet Byron, who wrote a set of computational instructions for solving Bernoulli numbers for Babbage’s difference engine. This set of instructions is also the first computer algorithm program in human history, separating hardware from software and introducing the concept of a program for the first time.

It was not until the first half of the 20th century that four scientific foundations of modern computing technology emerged: Boolean algebra (mathematics), Turing machines (computational models), von Neumann architecture (architecture), and transistors (devices). Among them, Boolean algebra is used to describe the underlying logic of programs and hardware such as CPUs; Turing machines are a universal computational model that transforms complex tasks into automated computations without human intervention; the von Neumann architecture proposed three basic principles for constructing computers: using binary logic, storing and executing programs, and that computers consist of five basic units: arithmetic units, controllers, memory, input devices, and output devices; transistors are semiconductor devices that form the basic logic circuits and storage circuits, serving as the “bricks” for building modern computers. Based on these scientific foundations, computing technology has developed rapidly, forming a large-scale industry.

From the birth of the world’s first electronic computer, ENIAC, in 1946 to today in the 21st century, five types of successful platform computing systems have been formed. Currently, various types of applications in different fields can be supported by these five types of platform computing devices. The first type is high-performance computing platforms, which solve scientific and engineering computing problems for core national departments; The second type is enterprise computing platforms, also known as servers, used for enterprise-level data management and transaction processing. Currently, computing platforms of internet companies like Baidu, Alibaba, and Tencent belong to this category; The third type is personal computer platforms, which appear in the form of desktop applications, allowing people to interact with personal computers through desktop applications; The fourth type is smartphones, characterized by mobility and portability, connecting to data centers via the internet, primarily focused on internet applications, distributedly deployed in data centers and mobile terminals; The fifth type is embedded computers, embedded in industrial equipment and military devices, ensuring the completion of specific tasks within a determined time through real-time control. These five types of devices cover almost all aspects of our information society, while the long-sought sixth type of platform computing system centered on intelligent computing applications has yet to be formed.

The development of modern computing technology can be roughly divided into three eras. IT 1.0, also known as the electronic computing era (1950-1970), is characterized by being “machine-centered.” The basic architecture of computing technology was formed, and with the advancement of integrated circuit technology, the scale of basic computing units rapidly shrank, with transistor density, computing performance, and reliability continuously improving, leading to widespread application of computers in scientific engineering computing and enterprise data processing.

IT 2.0, also known as the network computing era (1980-2020), is centered around “people.” The internet connects the terminals used by people with backend data centers, and internet applications interact with people through intelligent terminals. Internet companies like Amazon proposed the idea of cloud computing, encapsulating backend computing power as a public service rented to third-party users, forming the cloud computing and big data industry.

IT 3.0, also known as the intelligent computing era, began in 2020, adding the concept of “things” compared to IT 2.0, referring to various edge devices in the physical world that are digitized, networked, and intelligent, achieving a triadic integration of “people-machine-things.” In the intelligent computing era, in addition to the internet, there is also data infrastructure supporting various terminals to achieve the Internet of Everything through edge-cloud integration, with terminals, things, edges, and clouds all embedded with AI, providing intelligent services similar to large models like ChatGPT, ultimately achieving AI intelligence wherever there is computation. Intelligent computing has brought massive data, breakthroughs in artificial intelligence algorithms, and explosive demand for computing power.

Alan Turing (left) and John von Neumann

2. Overview of Intelligent Computing Development

Intelligent computing includes artificial intelligence technology and its computing carriers, which have roughly gone through four stages: general computing devices, logical reasoning expert systems, deep learning computing systems, and large model computing systems.

The starting point of intelligent computing is the general automatic computing device (1946). Scientists like Alan Turing and John von Neumann initially hoped to simulate the process of the human brain processing knowledge, inventing machines that think like the human brain. Although this was not achieved, it solved the problem of automation in computation. The emergence of general automatic computing devices also promoted the birth of the concept of artificial intelligence (AI) in 1956, and all subsequent developments in artificial intelligence technology have been built on the new generation of computing devices and stronger computing capabilities.

The second stage of intelligent computing development is logical reasoning expert systems (1990). Scientists from the symbolic AI school, such as Edward Albert Feigenbaum, aimed primarily at automating logical reasoning capabilities, proposing expert systems capable of performing logical reasoning with knowledge symbols. Human prior knowledge enters the computer in the form of knowledge symbols, enabling the computer to assist humans in making certain logical judgments and decisions in specific fields. However, expert systems heavily rely on manually generated knowledge bases or rule bases. A typical representative of this type of expert system is Japan’s fifth-generation computer and China’s 863 Program-supported 306 intelligent computer theme. Japan adopted dedicated computing platforms and knowledge reasoning languages like Prolog to complete application-level reasoning tasks in logical expert systems; China took a different technical route, based on general computing platforms, transforming intelligent tasks into artificial intelligence algorithms, integrating both hardware and system software into general computing platforms, giving rise to a number of backbone enterprises such as Shuguang, Hanwang, and iFlytek.

The limitations of symbolic computing systems lie in their explosive computational time-space complexity, meaning that symbolic computing systems can only solve linear growth problems and are unable to address high-dimensional complex spatial problems, thus limiting the size of problems they can handle. Additionally, since symbolic computing systems are based on knowledge rules, we cannot enumerate all common sense through exhaustive methods, significantly restricting their application range. With the arrival of the second AI winter, the first generation of intelligent computers gradually exited the historical stage.

It was not until around 2014 that intelligent computing advanced to the third stage—deep learning computing systems. Represented by Geoffrey Hinton and others from the connectionist school, aiming for automation of learning capabilities, new AI algorithms such as deep learning were invented. Through the automatic learning of deep neural networks, the model’s statistical induction capabilities were significantly enhanced, achieving tremendous breakthroughs in applications such as pattern recognition, with recognition accuracy in certain scenarios even surpassing that of humans. Taking facial recognition as an example, the entire training process of the neural network is equivalent to a process of adjusting network parameters, inputting a large number of labeled facial image data into the neural network, and then adjusting parameters between networks to make the output probability of the neural network approach the real result. The greater the probability of the neural network outputting the true situation, the larger the parameters, thus encoding knowledge and rules into network parameters. As long as there is enough data, it can learn various common sense, greatly enhancing its generality. The applications of connectionist intelligence are more extensive, including speech recognition, facial recognition, and autonomous driving. In terms of computing carriers, the Institute of Computing Technology, Chinese Academy of Sciences proposed the world’s first deep learning processor architecture in 2013, and internationally renowned hardware manufacturer NVIDIA has continuously released several leading general-purpose GPU chips, both of which are typical representatives of deep learning computing systems.

The fourth stage of intelligent computing development is large model computing systems (2020). Driven by the technology of large models in artificial intelligence, intelligent computing has reached new heights. In 2020, AI shifted from “small models + discriminative” to “large models + generative,” upgrading from traditional facial recognition, object detection, and text classification to today’s text generation, 3D digital human generation, image generation, speech generation, and video generation. A typical application of large language models in dialogue systems is OpenAI’s ChatGPT, which uses the pre-trained base large language model GPT-3, incorporating a training corpus of 300 billion words, equivalent to the total of all English text on the internet. Its basic principle is: by giving it an input, it predicts the next word to train the model, improving prediction accuracy through extensive training, ultimately allowing it to generate an answer to a question posed to it in real-time dialogue. Based on the base large model, it is further fine-tuned with supervised instruction through human to gradually teach the model how to engage in multi-turn dialogue with humans; finally, through manually designed and automatically generated reward functions, reinforcement learning iterations are conducted to gradually align the large model with human values.

The characteristics of large models are defined by their “size,” which has three layers of meaning: (1) large parameters, with GPT-3 having 170 billion parameters; (2) large training data, with ChatGPT using approximately 300 billion words and 570GB of training data; (3) large computing power requirements, with GPT-3 requiring thousands of V100 GPUs for training. To meet the explosive demand for intelligent computing power from large models, new intelligent computing centers are being built on a large scale both domestically and internationally, and NVIDIA has also launched a large model intelligent computing system composed of 256 H100 chips and 150TB of massive GPU memory.

The emergence of large models has brought about three transformations. First, the scale law in technology, which states that the accuracy of many AI models rapidly improves once the parameter scale exceeds a certain threshold, the reasons for which are still not very clear in the scientific community and are highly debated. The performance of AI models has a “logarithmic linear relationship” with the model parameter scale, dataset size, and total computing power, thus increasing the model’s scale can continuously improve its performance. Currently, the most advanced large model, GPT-4, has reached a parameter scale of trillions to tens of trillions and continues to grow; second, the explosive growth of computing power demand in the industry, where training large models with hundreds of billions of parameters typically requires training on thousands or even tens of thousands of GPU cards for 2-3 months, leading to a rapid increase in computing power demand, driving related computing power companies to develop at super high speed, with NVIDIA’s market value approaching $2 trillion, a phenomenon never seen before in chip companies; third, the impact on the labor market in society, with a report released by Peking University’s National Development Research Institute and Zhaopin Recruitment indicating that among the 20 occupations most affected, accounting, sales, and clerical positions rank at the top, while labor-intensive jobs that require interaction with people, such as human resources, administration, and logistics, are relatively safer.

The technological frontier of artificial intelligence will develop in the following four directions. The first frontier direction is multimodal large models. From the perspective of human beings, human intelligence is inherently multimodal, possessing eyes, ears, nose, tongue, body, and mouth (language). From the AI perspective, vision, hearing, etc., can also be modeled as sequences of tokens, and the same methods used for large language models can be applied for learning, further aligning with the semantics in language to achieve intelligent capabilities of multimodal alignment.

The second frontier direction is video generation large models. OpenAI released the text-to-video model SORA on February 15, 2024, significantly increasing video generation duration from a few seconds to one minute, with significant improvements in resolution, visual realism, and temporal consistency. The greatest significance of SORA is that it possesses the basic characteristics of a world model, which is the ability to observe and predict the world as humans do. A world model is built on understanding the basic physical knowledge of the world (e.g., water flows downhill) and then observing and predicting what event will happen next. Although SORA still faces many issues to become a world model, it can be considered that SORA has learned the ability of visual imagination and minute-level future prediction, which are foundational characteristics of a world model.

The third frontier direction is embodied intelligence. Embodied intelligence refers to intelligent agents with bodies that support interaction with the physical world, such as robots and unmanned vehicles, processing various sensor data inputs through multimodal large models, generating motion instructions to drive the intelligent agents, replacing traditional rule-based or mathematical formula-based motion driving methods, achieving deep integration of the virtual and real worlds. Therefore, robots with embodied intelligence can integrate the three major schools of artificial intelligence: connectionism represented by neural networks, symbolism represented by knowledge engineering, and behaviorism related to cybernetics, which is expected to bring new technological breakthroughs.

The fourth frontier direction is AI4R (AI for Research) becoming the main paradigm for scientific discovery and technological invention. Currently, scientific discovery mainly relies on experiments and human intelligence, with humans making bold conjectures and careful verifications, where information technology, whether in computation or data, only plays a supportive and validating role. Compared to humans, artificial intelligence has significant advantages in memory, high-dimensional complexity, full vision, reasoning depth, and conjecture. Whether AI can take the lead in some scientific discoveries and technological inventions, significantly improving the efficiency of human scientific discovery, such as actively discovering physical laws, predicting protein structures, designing high-performance chips, and efficiently synthesizing new drugs, is a question. Because large models of artificial intelligence possess complete data and have a god-like perspective, they can look further ahead than humans. If they can achieve a leap from inference to reasoning, artificial intelligence models have the potential to possess imagination and scientific conjecture abilities akin to Einstein, greatly enhancing the efficiency of human scientific discovery and breaking human cognitive boundaries. This is the true disruption.

Finally, general artificial intelligence (AGI) is a highly challenging and controversial topic. There was once a bet between a philosopher and a neuroscientist: whether researchers could reveal how the brain achieves consciousness in 25 years (i.e., by 2023). At that time, there were two schools of thought regarding consciousness: one called integrated information theory, which believed that consciousness is formed by specific types of neuron connections in the brain, and the other called global workspace theory, which pointed out that consciousness arises when information spreads through interconnected networks to brain regions. In 2023, people conducted adversarial experiments through six independent laboratories, and the results did not completely match either theory, with the philosopher winning and the neuroscientist losing. This bet illustrates that people always hope artificial intelligence can understand the mysteries of human cognition and the brain. From the perspective of physics, physics is about thoroughly understanding the macroscopic world, starting from quantum physics to understand the microscopic world. The intelligent world, like the physical world, is also a research object of immense complexity. AI large models still study the macroscopic world through data-driven methods to improve machine intelligence levels, and directly seeking answers in the microscopic world of the nervous system is challenging. Since its inception, artificial intelligence has always carried humanity’s dreams and fantasies about intelligence and consciousness, inspiring continuous exploration.

China: AI enterprises are mainly composed of national teams and entrepreneurial unicorns, where national teams often comprehensively layout technology and application fields, while entrepreneurial unicorns focus on vertical technology and application markets.The United States: AI manufacturers are numerous, with both software and hardware strengths, large manufacturers often comprehensively layout the AI industry, and startups are spread across the foundational, technical, and application layers.

3. Security Risks of Artificial Intelligence

The development of artificial intelligence has promoted technological progress in today’s world, but it has also brought many security risks that need to be addressed from both technical and regulatory perspectives.

First is the rampant spread of false information on the internet. Here are several scenarios: First, digital avatars. AI Yoon is the first official “candidate” synthesized using DeepFake technology, based on South Korean National Power Party candidate Yoon Suk-yeol, created by a local DeepFake technology company using 20 hours of audio and video clips of Yoon, along with over 3,000 sentences specifically recorded for researchers. In reality, the content expressed by AI Yoon was written by the campaign team, not the candidate himself.

Second, forged videos, especially those of leaders, can cause international disputes, disrupt election order, or trigger sudden public opinion events, such as the forged announcement of the first moon landing failure by Nixon or the forged announcement of “surrender” by Ukrainian President Zelensky. These actions have led to a decline in social trust in the news media industry.

Third, forged news, primarily generated through automated false news to gain illegal profits, using ChatGPT to generate trending news to earn traffic. As of June 30, 2023, there were 277 global websites generating fake news, severely disrupting social order.

Fourth, face-swapping and voice imitation used for fraud. For instance, due to AI mimicking the voice of a corporate executive, an international company in Hong Kong was defrauded of $35 million.

Fifth, generating inappropriate images, especially targeting public figures, such as the production of pornographic videos of film stars, causing adverse social impacts. Therefore, there is an urgent need to develop detection technologies for the forgery of false information on the internet.

Secondly, AI large models face serious trust issues. These issues include: (1) factual errors that are “seriously misleading”; (2) narratives based on Western values, outputting political biases and erroneous statements; (3) susceptibility to inducement, outputting incorrect knowledge and harmful content; (4) exacerbated data security issues, with large models becoming significant traps for sensitive data, as ChatGPT incorporates user inputs into its training database to improve itself, allowing foreign entities to utilize large models to access Chinese language data that may not be available through public channels, gaining knowledge about China that even we may not possess. Therefore, there is an urgent need to develop security regulatory technologies for large models and establish our own trustworthy large models.

In addition to technical measures, the security of artificial intelligence also requires relevant legislative work. In 2021, the Ministry of Science and Technology released the “Ethical Norms for the New Generation of Artificial Intelligence,” and in August 2022, the National Information Security Standardization Technical Committee released the “Safety Assessment Specifications for Machine Learning Algorithms.” Between 2022 and 2023, the Central Cyberspace Administration of China successively released regulations such as the “Regulations on Algorithm Recommendation Management for Internet Information Services,” “Regulations on Deep Synthesis Management for Internet Information Services,” and “Management Measures for Generative Artificial Intelligence Services.” European and American countries have also successively introduced regulations, with the European Union issuing the “General Data Protection Regulation” on May 25, 2018, and the United States releasing the “Blueprint for the AI Bill of Rights” on October 4, 2022, while the European Parliament passed the “AI Act” on March 13, 2024.

China should accelerate the introduction of the “Artificial Intelligence Law,” establish an artificial intelligence governance system, ensure that the development and application of artificial intelligence adhere to common human values, promote harmonious and friendly human-machine interactions; create a policy environment conducive to the research, development, and application of artificial intelligence technology; establish reasonable disclosure and audit assessment mechanisms to understand the principles and decision-making processes of artificial intelligence mechanisms; clarify the security responsibilities and accountability mechanisms of artificial intelligence systems, trace back responsible parties and provide remedies; and promote the formation of fair, reasonable, open, and inclusive international governance rules for artificial intelligence.

4. Dilemmas in the Development of Intelligent Computing in China

The technology of artificial intelligence and the intelligent computing industry are at the center of Sino-U.S. technological competition. Although China has made significant achievements in recent years, it still faces many development dilemmas, particularly difficulties brought about by U.S. technological suppression policies.

The first dilemma is that the U.S. has long been in a leading position in AI core capabilities, while China is in a tracking mode. China lags behind the U.S. in the number of high-end AI talents, innovation in AI foundational algorithms, capabilities of foundational large models (large language models, text-to-image models, text-to-video models), training data for foundational large models, and training computing power for foundational large models, and this gap is expected to persist for a long time.

The second dilemma is the ban on the sale of high-end computing products, with high-end chip processes being long restricted. High-end intelligent computing chips such as A100, H100, and B200 are banned from being sold to China. Companies like Huawei, Loongson, Cambricon, Shuguang, and Haiguang have all been placed on the entity list, and their advanced chip manufacturing processes are restricted, with domestic processes capable of meeting large-scale production lagging 2-3 generations behind international advanced levels, and the performance of core computing chips lagging 2-3 generations behind international advanced levels.

The third dilemma is the weak domestic intelligent computing ecosystem, with insufficient penetration of AI development frameworks. NVIDIA’s CUDA (Compute Unified Device Architecture) ecosystem is complete, forming a de facto monopoly. The domestic ecosystem is weak, specifically manifested in: (1) insufficient R&D personnel, with nearly 20,000 developers in the NVIDIA CUDA ecosystem, which is 20 times the total number of personnel in all domestic intelligent chip companies; (2) insufficient development tools, with CUDA having 550 SDKs (Software Development Kits), which is hundreds of times more than those of domestic companies; (3) insufficient funding, with NVIDIA investing $5 billion annually, which is dozens of times more than domestic companies; (4) the AI development framework TensorFlow occupies the industrial market, while PyTorch dominates the research market, with domestic AI development frameworks like Baidu’s PaddlePaddle having only 1/10 the number of developers compared to foreign frameworks. More seriously, domestic enterprises are fragmented, unable to form a cohesive force. Although there are related products at each layer, such as intelligent applications, development frameworks, system software, and intelligent chips, there is no deep adaptation between layers, making it impossible to form a competitive technical system.

The fourth dilemma is that the cost and threshold for applying AI in industries remain high. Currently, AI applications in China are mainly concentrated in the internet industry and some defense fields. When promoting AI technology applications across various industries, especially when migrating from the internet industry to non-internet industries, a large amount of customization work is required, making migration difficult and the cost of single-use high. Finally, there is also a significant shortage of talent in the AI field compared to actual demand in China.

5. Path Choices for the Development of Intelligent Computing in China

The path choices for the development of artificial intelligence are crucial for China, as they relate to the sustainability of development and the final international competitive landscape. Currently, the cost of using artificial intelligence is extremely high. The Microsoft Copilot suite requires a monthly fee of $10, ChatGPT consumes 500,000 kilowatt-hours of electricity daily, and the price of NVIDIA’s B200 chip exceeds $30,000. Overall, China should develop affordable, safe, and trustworthy artificial intelligence technology to eliminate information poverty among its population and benefit countries along the “Belt and Road”; empower various industries with low thresholds, maintain the competitiveness of China’s advantageous industries, and significantly narrow the gap for relatively lagging industries.

Choice 1: Should we unify the technical system to follow a closed-source approach or an open-source approach?

The intelligent computing industry is supported by a tightly coupled technical system, which is a technical whole closely linking materials, devices, processes, chips, complete machines, system software, and application software through a series of technical standards and intellectual property. China has three paths for developing an intelligent computing technology system:

First, to catch up and be compatible with the U.S.-led A system. Most of China’s internet companies are following the GPGPU/CUDA compatibility path, and many startups in the chip field are also trying to build ecosystems compatible with CUDA, which is a more realistic path. Due to U.S. restrictions on China’s processes and chip bandwidth in terms of computing power, and the difficulty of forming a unified domestic ecosystem in terms of algorithms, the maturity of the ecosystem is severely limited, and the lack of high-quality Chinese data will make it difficult for the gap between followers and leaders to narrow, and at times may even widen.

Second, to build a dedicated closed B system. In specialized fields such as military, meteorology, and justice, to construct a closed ecosystem based on domestic mature processes for chip production, focusing more on vertical large models in specific fields rather than foundational large models, and training large models using proprietary high-quality data from the field. This path is easier to form a complete and controllable technical system and ecosystem, and some large backbone enterprises in China are following this path. Its drawback is that it is closed, making it difficult to gather the majority of domestic forces and achieve globalization.

Third, to globally co-build an open-source C system. Breaking the ecological monopoly with open-source, lowering the threshold for enterprises to possess core technologies, allowing every enterprise to create its own chips at low cost, forming a vast ocean of intelligent chips to meet ubiquitous intelligent demands. Using openness to form a unified technical system, Chinese enterprises should unite with global forces to co-build a unified intelligent computing software stack based on international standards. Establishing a shared mechanism for enterprise competition, sharing high-quality databases, and sharing open-source general foundational large models. In the global open-source ecosystem, Chinese enterprises have benefited greatly during the internet era, and in the intelligent era, Chinese enterprises should become the main contributors to the RISC-V + AI open-source technology system, becoming the leading force in global open sharing.

Choice 2: Should we focus on algorithm models or invest in new infrastructure?

Artificial intelligence technology must empower various industries, exhibiting a typical long-tail effect. 80% of China’s small and medium-sized enterprises need low-threshold, low-cost intelligent services. Therefore, China’s intelligent computing industry must be built on a new data space infrastructure, where the key is to achieve comprehensive infrastructure for intelligent elements, namely data, computing power, and algorithms. This work can be compared to the historical role of the U.S. information superhighway plan (i.e., information infrastructure construction) for the internet industry in the early 20th century.

The core productivity of the information society is cyberspace. The evolution of cyberspace is from a computing space formed by unidirectional connections of machines, evolving into an information space formed by bidirectional connections of humans and machines, and further evolving into a data space formed by triadic connections of humans, machines, and things. From the perspective of data space, the essence of artificial intelligence is the refinement of data, and large models are products of deep processing of all data on the internet. In the digital age, what is transmitted over the internet is information flow, which is a structured abstraction of data processed roughly by computing power; in the intelligent age, what is transmitted over the internet is intelligent flow, which is a modeled abstraction of data processed and refined deeply by computing power. A core feature of intelligent computing is to process massive data items using numerical computation, data analysis, and artificial intelligence algorithms in a computing power pool to obtain intelligent models, which are then embedded in various processes of the information world and the physical world.

The Chinese government has proactively laid out new infrastructure, seizing the initiative in global competition. First, data has become a national strategic information resource. Data has both resource element and value processing attributes, including production, acquisition, transmission, aggregation, circulation, trading, ownership, assets, and security. China should continue to strengthen the construction of national data hubs and data circulation infrastructure.

Second, AI large models are a type of algorithmic infrastructure in the data space. Based on general large models, infrastructure for the research and application of large models should be constructed to support enterprises in developing specialized large models for industries such as robotics, unmanned driving, wearable devices, smart homes, and intelligent security, covering long-tail applications.

Finally, the construction of a national integrated computing power network has played a pioneering role in promoting the infrastructure of computing power. The Chinese solution for computing power infrastructure should significantly reduce the cost and threshold of using computing power while providing high-throughput and high-quality intelligent services to the widest range of people. The Chinese solution for computing power infrastructure needs to achieve “two lows and one high,” meaning that on the supply side, it should significantly reduce the total costs of computing power devices, computing power equipment, network connections, data acquisition, algorithm model invocation, power consumption, operation maintenance, and development deployment, making high-quality computing power services affordable for a wide range of small and medium-sized enterprises, encouraging them to actively develop computing power network applications; on the consumption side, it should significantly lower the threshold for the general public to use computing power, ensuring that public services must be easy to access and use, like water and electricity, and easily customizable like web page creation. In terms of service efficiency, China’s computing power services should achieve low entropy and high throughput, where high throughput refers to the ability to provide high concurrency services while maintaining a high satisfaction rate for end-to-end service response times; low entropy means ensuring system throughput does not sharply decline under high concurrency loads due to resource competition. Ensuring “more computing power” is especially important for China.

Choice 3: Should we focus on empowering the virtual economy with AI or strengthen the real economy?

The effectiveness of “AI +” is a touchstone for the value of artificial intelligence. After the subprime mortgage crisis, the share of manufacturing value added in the U.S. GDP dropped from 28% in 1950 to 11% in 2021, and the proportion of employment in manufacturing in the U.S. dropped from 35% in 1979 to 8% in 2022, indicating that the U.S. tends to favor the virtual economy with higher returns, neglecting the real economy with high investment costs and low economic returns. China, on the other hand, tends to develop both the real economy and the virtual economy simultaneously, placing greater emphasis on developing equipment manufacturing, new energy vehicles, photovoltaic power generation, lithium batteries, high-speed rail, and 5G as part of the real economy.

Correspondingly, AI in the U.S. is mainly applied to the virtual economy and IT infrastructure tools, reflecting a trend of “detaching from reality.” Since 2007, Silicon Valley has continuously hyped virtual reality (VR), the metaverse, blockchain, Web 3.0, deep learning, and AI large models, reflecting this trend.

China’s advantage lies in the real economy, with the most complete and comprehensive global manufacturing industry categories, characterized by diverse scenarios and abundant private data. China should select several industries to increase investment, forming paradigms that can be promoted across all industries with low thresholds, such as choosing equipment manufacturing as a representative industry to continue its advantages and choosing the pharmaceutical industry as a representative industry to rapidly narrow the gap. The technical challenge of empowering the real economy lies in the integration of AI algorithms with physical mechanisms.

The key to the success of artificial intelligence technology is whether it can significantly reduce the costs of an industry or product, thereby expanding the user base and industry scale by tenfold, producing transformative effects similar to those of the steam engine on the textile industry or the smartphone on the internet industry.

China should find a high-quality development path suitable for its own artificial intelligence to empower the real economy.

Notes:

① Pattern recognition refers to the use of computational methods to classify samples based on their features, studying the automatic processing and interpretation of patterns through mathematical methods by computers, with major research directions including image processing and computer vision, speech language information processing, brain networks, and brain-like intelligence.

② Token refers to symbols used to represent words or phrases in the natural language processing process. Tokens can be single characters or sequences composed of multiple characters.

③ General artificial intelligence refers to a type of artificial intelligence that possesses intelligence comparable to or exceeding that of humans. General artificial intelligence can not only perceive, understand, learn, and reason like humans but also flexibly apply knowledge across different fields, learn quickly, and think creatively. The research goal of general artificial intelligence is to seek a unified theoretical framework to explain various intelligent phenomena.

④ Chip manufacturing processes refer to the processes used to manufacture CPUs or GPUs, specifically the size of transistor gate circuits, measured in nanometers. Currently, the most advanced processes in mass production internationally are represented by TSMC’s 3nm. More advanced manufacturing processes can integrate more transistors within CPUs and GPUs, enabling processors to have more functions and higher performance while being smaller and cheaper.

⑤ CUDA is a parallel computing platform and programming model developed by NVIDIA, which includes the CUDA instruction set architecture and the parallel computing engine within GPUs. Developers can use the C language to write programs for the CUDA architecture, and the programs can run at ultra-high performance on processors that support CUDA.

⑥ RISC-V (pronounced “risk-five”) is an open general-purpose instruction set architecture initiated by the University of California, Berkeley, which allows anyone to freely use the RISC-V instruction set to design, manufacture, and sell chips and software, unlike other paid instruction sets.

⑦ The long-tail effect refers to the phenomenon where products or services that were previously overlooked but have small sales volumes, due to their large total number, accumulate total revenues that exceed those of mainstream products. This effect is particularly significant in the internet field.

⑧ High concurrency typically refers to the design ensuring that a system can process many requests in parallel simultaneously.