Lecture Script for the Tenth Lecture of the Special Lecture Series of the Standing Committee of the 14th National People's Congress
The Development of Artificial Intelligence and Intelligent ComputingSun Ninghui
Chairman, Vice Chairpersons, Secretary-General, and Members:
Artificial intelligencehas in recent years been experiencing an explosive development led bygenerative artificial intelligencelarge models. On November 30, 2022, OpenAI launched an artificial intelligence conversational chatbotChatGPT, whose outstanding natural language generation capabilities attracted widespread attention worldwide, surpassing 100 million users in 2 months, and immediately triggering a wave of large models at home and abroad,Gemini、ERNIE Bot、Copilot、LLaMA、SAM、SORAand various other large models emerged like bamboo shoots after a spring rain, and 2022 has also been hailed as the first year of large models. The current information age is accelerating into the development stage of intelligent computing, breakthroughs in artificial intelligence technology are emerging endlessly, gradually empowering all industries in depth, and promoting artificial intelligence and data elements to become typical representatives of new quality productive forces. General Secretary Xi Jinping pointed out that the new generation of artificial intelligence should be used as a driving force for promoting leapfrog development of science and technology, optimization and upgrading of industries, and overall elevation of productivity, striving to achieve high-quality development. Since the 18th National Congress of the Communist Party of China, the Party Central Committee with Comrade Xi Jinping at its core has attached great importance to the development of the intelligent economy, promoted the deep integration of artificial intelligence and the real economy, and injected strong momentum into high-quality development.
I. Brief Introduction to the Development of Computing Technology
The history of the development of computing technology can be roughly divided into four stages. The appearance of the abacus marked humanity's entry into the first generation—the era of mechanical computing. The second generation—electronic computing—was marked by the emergence of electronic devices and electronic computers. The emergence of the Internet brought us into the third generation—network computing. At present, 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 Napier's bones (1612) and the wheel adder (1642). By 1672, the first computing device that automatically completed the four arithmetic operations—the stepped reckoner—was born.
During the mechanical computing period, some basic concepts of modern computers had already appeared. Charles Babbage proposed the design concepts of the Difference Engine (1822) and the Analytical Engine (1834), supporting automatic mechanical computation. During this period, the concepts of programming and programs basically took shape. The concept of programming originated from the Jacquard loom, which controlled printing patterns through punched cards, and eventually evolved into storing all mathematical calculation steps in the form of computing instructions. The first programmer in human history was Ada, the daughter of the poet Byron. She wrote a set of calculation instructions for Babbage's Difference Engine to solve the Bernoulli sequence. This set of instructions was also the first set of computer algorithm programs in human history. It separated hardware and software, and the concept of a program appeared for the first time.
It was not until the first half of the twentieth century that the four scientific foundations of modern computing technology emerged: Boolean algebra (mathematics), the Turing machine (computational model), the von Neumann architecture (architecture), and the transistor (device). Among them, Boolean algebra is used to describe the underlying logic of programs and hardware such as CPUs; the Turing machine is a general computational model that transforms complex tasks into automatic computation, an automated process requiring no manual intervention; the von Neumann architecture proposed three basic principles for constructing computers: adopting binary logic, stored-program execution, and the computer being composed of five basic units: the arithmetic unit, the control unit, the memory, the input device, and the output device; the transistor is a semiconductor device that constitutes basic logic circuits and memory circuits, and is the "brick" for building the tower of modern computers. Based on the above scientific foundations, computing technology has been able to develop rapidly and form a huge-scale industry.
From the birth of ENIAC, the world's first electronic computer, in 1946 to the twenty-first century today, five types of successful platform-based computing systems have been formed. Various types of applications in various fields today can all be supported by these five types of platform-based computing devices. The first type is the high-performance computing platform, which solves scientific and engineering computing problems for the core departments of the state; the second type is the enterprise computing platform, also called a server, used for enterprise-level data management and transaction processing. Currently, the computing platforms of internet companies such as Baidu, Alibaba, and Tencent all belong to this category; the third type is the personal computer platform, which appears in the form of desktop applications, through which people interact with personal computers; the fourth type is the smartphone, whose main feature is mobile portability. Phones connect to data centers through networks and are mainly based on internet applications, distributedly deployed in data centers and mobile phone terminals; the fifth type is the embedded computer, embedded in industrial equipment and military equipment, ensuring the completion of specific tasks within a determined time through real-time control. These five types of devices cover almost every aspect of our information society. The sixth type of platform-based computing system, centered on intelligent computing applications, which people have long pursued, has not yet been formed.
The development of modern computing technology can be roughly divided into three eras. IT1.0, also called the electronic computing era (1950-1970), had the basic characteristic of being centered on the "machine." The basic architecture of computing technology was formed. With advances in integrated circuit processes, the scale of basic computing units shrank rapidly, and transistor density, computing performance, and reliability continued to improve. Computers were widely used in scientific and engineering computing and enterprise data processing.
IT2.0, also called the network computing era (1980-2020), is centered on "people." The internet connects the terminals used by people with back-end data centers, and internet applications interact with people through intelligent terminals. Internet companies represented by Amazon and others proposed the idea of cloud computing, encapsulating back-end computing power as a public service and renting it to third-party users, forming the cloud computing and big data industry.
IT3.0, also called the intelligent computing era, began in 2020. Compared with IT2.0, it adds the concept of "things," namely various end-side devices in the physical world, which are digitized, networked, and made intelligent, realizing the trinity integration of "human-machine-thing." In the intelligent computing era, in addition to the internet, there is also data infrastructure, supporting various terminals in realizing the Internet of Everything through device-edge-cloud. Terminals, things, edges, and clouds all embed AI, providing large-model intelligent services similar to ChatGPT, ultimately realizing that wherever there is computing, there is AI intelligence. Intelligent computing has brought massive amounts of data, breakthroughs in artificial intelligence algorithms, and explosive demand for computing power.
II. Brief Introduction to the Development of Intelligent Computing
Intelligent computing includes artificial intelligence technology and its computing carriers, and has roughly gone through four stages, namely general-purpose computing devices, logical reasoningExpert systems、Deep learningComputing systems, large-model computing systems.
The starting point of intelligent computing was the general-purpose automatic computing device (1946). Scientists such as Alan Turing and John von Neumann initially hoped to simulate the process by which the human brain processes knowledge and to invent machines that think like the human brain. Although this was not achieved, it did solve the problem of automating computation. The emergence of general-purpose automatic computing devices also drove the birth of the concept of artificial intelligence (AI) in 1956. Since then, all developments in artificial intelligence technology have been built on a new generation of computing devices and stronger computing power.
The second stage in the development of intelligent computing was the logical reasoning expert system (1990). Scientists of the symbolic intelligence school, such as Edward Albert Feigenbaum, took the automation of logical and reasoning capabilities as their main goal and proposed expert systems capable of performing logical reasoning on knowledge symbols. Human prior knowledge entered the computer in the form of knowledge symbols, enabling the computer to assist humans in certain logical judgments and decisions in specific domains, but expert systems relied heavily on manually generated knowledge bases or rule bases. Typical representatives of this kind of expert system were Japan's Fifth Generation Computer and the 306 intelligent computer theme supported by China's 863 Program. In logical expert systems, Japan adopted dedicated computing platforms and knowledge reasoning languages such as Prolog to complete application-level reasoning tasks; China took a different technical route from Japan, using general-purpose computing platforms as the foundation, turning intelligent tasks into artificial intelligence algorithms, connecting both hardware and system software to general-purpose computing platforms, and giving rise to a group of backbone enterprises such as Sugon, Hanwang, and iFlytek.
The limitation of symbolic computing systems lies in their explosively high computational space-time complexity, that is, symbolic computing systems can only solve problems of linear growth and cannot solve high-dimensional complex space problems, thereby limiting the size of the problems they can handle. At the same time, because symbolic computing systems are established based on knowledge rules, and we cannot exhaustively enumerate all common sense by brute force, their scope of application is greatly limited. With the arrival of the second AI winter, the first generation of intelligent computers gradually withdrew from the stage of history.
Until around 2014, intelligent computing advanced to its third stage—deep learning computing systems. The connectionist intelligence school, represented by Geoffrey Hinton and others, took the automation of learning ability as its goal and invented new AI algorithms such as deep learning. Through the automatic learning of deep neural networks, the ability of models to perform statistical induction was greatly improved, and huge breakthroughs were achieved in the effectiveness of applications such as pattern recognition①, with recognition accuracy in some scenarios even surpassing that of humans. Taking face recognition as an example, the entireneural networktraining process is equivalent to a process of adjusting network parameters. A large amount of annotated face image data is input into the neural network, and then parameters between the networks are adjusted so that the probability of the results output by the neural network infinitely approaches the true result. The greater the probability that the neural network outputs the true situation, the larger the parameters, thereby encoding knowledge and rules into network parameters. In this way, as long as there is enough data, various kinds of common sense can be learned, and generality is greatly improved. Connectionist intelligence is applied more widely, includingspeech recognition, face recognition, autonomous driving, and so on. In terms of computing carriers, the Institute of Computing Technology of the Chinese Academy of Sciences proposed the world's first deep learning processor architecture in 2013, and the internationally renowned hardware manufacturer NVIDIA continuously released multiple general-purpose GPU chips with leading performance, both of which are typical representatives of deep learning computing systems.
The fourth stage in the development of intelligent computing is large model computing systems (2020). Driven by large-model artificial intelligence technology, intelligent computing has reached new heights. In 2020, AI shifted from "small models + discriminative" to "large models + generative," from traditional facial recognition,object detection, and text classification to today'stext generation, 3D digital human generation, image generation, speech generation, andvideo generation. A typical application of large language models in the field of dialogue systems is OpenAI's ChatGPT, which uses the pretrained foundation large language model GPT-3 and introduces 300 billion words of training corpus, equivalent to the sum of all English text on the internet. Its basic principle is: by giving it an input and letting it predict the next word to train the model, and through extensive training to improve prediction accuracy, it ultimately reaches the point where, when asked a question, the large model produces an answer and engages in real-time conversation with a person. On the basis of the foundation large model, it is further given some prompt words forsupervised instruction fine-tuning, and through human <instruction, response> pairs the model gradually learns how to conduct multi-turn conversations with people; finally, through artificially designed and automatically generated reward functions,reinforcement learningiteration is carried out to gradually align the large model with human values.
The characteristic of large models is that they win by being "large," which has three meanings: (1) large parameters—GPT-3 has 170 billion parameters; (2) large training data—ChatGPT used about 300 billion words, 570GB of training data; (3) large computing power demand—GPT-3 used about tens of thousands of V100 GPUs for training. To meet the explosively increasing demand of large models for intelligent computing power, both domestically and internationally, massive new intelligent computing centers costing enormous sums are being built on a large scale, and NVIDIA has also launched a large-model intelligent computing system composed of 256 H100 chips, 150TB of massive GPU memory, and more.
The emergence of large models has brought three transformations. First is the technical Scaling Law, namely that the accuracy of many AI models improves rapidly after the parameter scale exceeds a certain threshold, and the reason for this is still not very clear in the scientific community and is highly controversial. The performance of AI models has a "log-linear relationship" with three variables: model parameter scale, dataset size, and total computing power, so model performance can be continuously improved by increasing model scale. At present, the parameter count of the most cutting-edge large model, GPT-4, has already reached the order of trillions to tens of trillions, and is still continuously growing. Second is the explosive growth of computing power demand in industry. Training a large model with a scale of hundreds of billions of parameters usually requires training on thousands or even tens of thousands of GPU cards for 2-3 months. The sharply increasing computing power demand has driven related computing power enterprises to develop at ultra-high speed, and NVIDIA's market value is approaching two trillion US dollars, which has never happened before for chip companies. Third is the impact on the labor market in society. The report "Research on the Potential Impact of AI Large Models on China's Labor Market," jointly released by Peking University's National School of Development and Zhaopin, points out that among the 20 most affected occupations, accounting, sales, and clerical work are at the forefront, while physical labor jobs that require dealing with people and providing services, such as human resources, administration, and logistics, are instead relatively safer.
The technological frontier of artificial intelligence will develop in the following four directions. The first frontier direction ismultimodallarge models. From the human perspective, human intelligence is naturally multimodal, as humans have eyes, ears, nose, tongue, body, and mouth (language); from the AI perspective, vision, hearing, and so on can also all be modeled astoken② sequence, learning can be conducted using the same methods as large language models, and further aligned with semantics in language to achieve multimodal alignment intelligence capabilities.
The second frontier direction is large models for video generation. On February 15, 2024, OpenAI released the text-to-video model SORA, greatly extending video generation duration from a few seconds to one minute, with significant improvements in resolution, image realism, temporal consistency, and other aspects. The greatest significance of SORA is that it possesses the basic characteristics of a world model, namely the ability of humans to observe the world and further predict it. A world model is built on understanding the basic physical common sense of the world (such as water flowing from high to low, etc.), and then observing and predicting what event will happen in the next second. Although many problems remain for SORA to become a world model, it can be considered that SORA has learned visual imagination and minute-level future prediction capabilities, which are foundational characteristics of a world model.
The third frontier direction is embodied intelligence. Embodied intelligence refers to having a body and supporting interaction with the physical world.agents, such as robots, autonomous vehicles, etc., which process multiple types of sensor data input through multimodal large models, and use large models to generate motion commands to drive the agents, replacing traditional motion-driving methods based on rules or mathematical formulas, and achieving deep integration of the virtual and the real. Therefore, robots with embodied intelligence can bring together the three major schools of artificial intelligence: connectionism represented by neural networks, symbolism represented by knowledge engineering, and behaviorism related to cybernetics. The three schools can act simultaneously on one agent, 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 brainpower, with humans making bold conjectures and careful verification, while information technology, whether computation or data, only plays a supporting and verifying role. Compared with humans, artificial intelligence has greater advantages in memory, high-dimensional complexity, full-field vision, reasoning depth, conjecture, and other aspects. Whether AI can be the main force in conducting some scientific discoveries and technological inventions, greatly 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. Because large artificial intelligence models possess the full set of data and have a god's-eye view, through deep learning capabilities, they can look more steps ahead than humans. If they can achieve the leap from inference toreasoning(reasoning), artificial intelligence models have the potential to possess imagination and scientific conjecture abilities like Einstein, greatly improving the efficiency of human scientific discovery and breaking through the boundaries of human cognition. This is where the true disruption lies.
Finally,artificial general intelligence③ (Artificial General Intelligence, abbreviated as AGI) is an extremely challenging and highly controversial topic. There was once a philosopher and a neuroscientist who made a bet: in 25 years (that is, 2023), would researchers be able to reveal how the brain achieves consciousness? At that time, there were two schools of thought about consciousness, one called integrated information theory, and the other called global workspace theory. The former believed that consciousness is a "structure" formed by connections among specific types of neurons in the brain, while the latter pointed out that consciousness arises when information propagates through interconnected networks to brain regions. In 2023, people conducted adversarial experiments through six independent laboratories, and the results did not fully match either theory. The philosopher won, and the neuroscientist lost. Through this bet, it can be seen that people always hope artificial intelligence can understand human cognition and the mysteries of the brain. From the perspective of physics, physics began with a thorough understanding of the macroscopic world and then started from quantum physics to open up the understanding of the microscopic world. The intelligent world, like the physical world, is an object of study with enormous complexity. AI large models still rely on data-driven and other methods for studying the macroscopic world to improve the intelligence level of machines, and their understanding of the intelligent macroscopic world is not sufficient. Going directly to the microscopic world of the nervous system to find answers is difficult. Since its birth, artificial intelligence has always carried humanity's various dreams and fantasies about intelligence and consciousness, and has also inspired people to keep exploring.
III. Security Risks of Artificial Intelligence
While the development of artificial intelligence has promoted scientific and technological progress in today's world, it has also brought many security risks, which must be addressed from both technical and regulatory aspects.
The first is the proliferation of false information on the internet. Here are several scenarios: first, digital avatars. AI Yoon is the first official "candidate" synthesized using DeepFake technology. This digital person is based on South Korean People Power Party candidate Yoon Suk-yeol, and with the help of 20 hours of Yoon Suk-yeol's audio and video clips, as well as more than 3,000 sentences he specially recorded for researchers, a local DeepFake technology company created the virtual image AI Yoon, which quickly became popular online. In fact, the content expressed by AI Yoon was written by the campaign team, not by the candidate himself.
The second is forged videos, especially forged videos of leaders that cause international disputes, disrupt election order, or trigger sudden public opinion incidents, such as forging Nixon announcing the failure of the first moon landing, and forging Ukrainian President Zelensky announcing "surrender." These actions have led to a decline in the news media industry's social trust.
The third is forged news, mainly through automatic generation of fake news to seek illegal profits, using ChatGPT to generate trending news and earn traffic. As of June 30, 2023, the number of fake news generation websites worldwide had reached 277, seriously disrupting social order.
The fourth is face swapping and voice changing, used for fraud. For example, because AI voice imitated the voice of a corporate executive, an international company in Hong Kong was defrauded of 35 million US dollars.
The fifth is generating indecent images, especially targeting public figures. For example, the production of pornographic videos of film and television stars causes adverse social impact. Therefore, there is an urgent need to develop forgery detection technology for false information on the internet.
Second, large AI models face serious credibility problems. These problems include: (1) factual errors of "confidently talking nonsense"; (2) narrating from Western values, outputting political bias and erroneous statements; (3) being easily induced to output erroneous knowledge and harmful content; (4) data security issues intensifying, with large models becoming traps for important sensitive data; ChatGPT incorporates user input into its training database to improve ChatGPT, and the US side can use large models to obtain Chinese-language corpora that public channels do not cover, and master "China knowledge" that even we ourselves may not possess. Therefore, there is an urgent need to develop large model security regulatory technology and our own trustworthy large models.
In addition to technical means, AI security assurance requires related legislative work. In 2021, the Ministry of Science and Technology issued the "New GenerationArtificial Intelligence EthicsCode", in August 2022, the National Information Security Standardization Technical Committee issued the "Information Security Technology Machine Learning Algorithm Security Assessment Specification", and in 2022-2023, the Cyberspace Administration of China successively issued the "Provisions on the Administration of Algorithm Recommendations for Internet Information Services", the "Provisions on the Administration of Deep Synthesis of Internet Information Services", and the "Administrative Measures for Generative Artificial Intelligence Services", among others. European and American countries have also successively introduced regulations: on May 25, 2018, the European Union issued the "General Data Protection Regulation"; on October 4, 2022, the United States issued the "Blueprint for an AI Bill of Rights"; and on March 13, 2024, the European Parliament passed the EU "Artificial Intelligence Act".
Our country should accelerate the promulgation of the "Artificial Intelligence Law", build an artificial intelligence governance system, ensure that the development and application of artificial intelligence follow the common values of humanity, and promote harmonious and friendly human-machine relations; create a policy environment conducive to artificial intelligence technology research, development, and application; establish reasonable disclosure mechanisms and audit and evaluation mechanisms to understand the mechanisms and decision-making processes of artificial intelligence; clarify the security responsibilities and accountability mechanisms of artificial intelligence systems, so that responsible entities can be traced and remedies made; and promote the formation of fair, reasonable, open, and inclusive international artificial intelligence governance rules.
IV. Dilemmas in China's Intelligent Computing Development
Artificial intelligence technology and the intelligent computing industry are at the focus of Sino-US technological competition. Although our country has made great achievements in the past few years, it still faces many development dilemmas, especially the difficulties brought by the US technology suppression policy.
Dilemma one is that the United States has long held a leading position in AI core capabilities, while China is in a tracking mode. China has a certain gap with the United States in the number of high-end AI talents, AI basic algorithm innovation, foundation large model capabilities (large language models, text-to-image models, text-to-video models), foundation large model training data, foundation large model training computing power, etc., and this gap will continue for a long period of time.
Dilemma two is the ban on sales of high-end computing power products, with high-end chip processes long being bottlenecked. High-end intelligent computing chips such as A100, H100, and B200 are banned from sale to China. Enterprises such as Huawei, Loongson, Cambricon, Sugon, and Hygon have all entered the Entity List, and their advanced chip manufacturing processes④ are restricted. The domestic process nodes that can satisfy mass production lag behind the international advanced level by 2-3 generations, and the performance of core computing power chips lags behind the international advanced level by 2-3 generations.
The third predicament is the weakness of the domestic intelligent computing ecosystem and the insufficient penetration rate of AI development frameworks. NVIDIA's CUDA⑤ (Compute Unified Device Architecture) ecosystem is complete and has formed a de facto monopoly. The domestic ecosystem is weak, specifically manifested as follows: First, there is an insufficient number of R&D personnel; NVIDIA's CUDA ecosystem has nearly 20,000 developers, 20 times the total number of personnel at all domestic intelligent chip companies; second, there is an insufficient number of development tools; CUDA has 550 SDKs (Software Development Kits), over a hundred times that of related domestic enterprises; third, there is insufficient funding; NVIDIA invests US$5 billion annually, dozens of times that of related domestic companies; fourth, the AI development framework TensorFlow occupies the industrial market, PyTorch occupies the research market, and domestic AI development frameworks such as Baidu PaddlePaddle have only 1/10 the developers of foreign frameworks. Even more serious is that domestic enterprises are divided into many separate camps and cannot form a joint force. From intelligent applications, development frameworks, system software, and intelligent chips, although each layer has related products, there is no deep adaptation between the layers, and a competitive technology system cannot be formed.
The fourth predicament is that when AI is applied to industries, the cost and threshold remain high. At present, China's AI applications are mainly concentrated in the internet industry and some national defense fields. When AI technology is promoted and applied to all walks of life, especially when migrating from the internet industry to non-internet industries, a large amount of customization work is required, the migration difficulty is great, and the single-use cost is high. Finally, compared with actual demand, China's number of talents in the AI field is also clearly insufficient.
V. China's Path Choices for How to Develop Intelligent Computing
The choice of path for artificial intelligence development is crucial to China and relates to the sustainability of development and the ultimate international competitive landscape. At present, the cost of using artificial intelligence is very high. The Microsoft Copilot suite costs US$10 per month to use, ChatGPT consumes 500,000 kilowatt-hours of electricity per day, and the price of NVIDIA's B200 chip is as high as over US$30,000. Overall, China should develop artificial intelligence technology that is affordable, safe, and trustworthy, eliminate China's information-poor population, and benefit countries along the Belt and Road; empower all walks of life with a low threshold, so that China's advantageous industries can maintain competitiveness and relatively backward industries can substantially narrow the gap.
Choice one: Should we unify the technology system and take the closed-source, closed path, or the open-source, open path?
What supports the intelligent computing industry is a tightly coupled technology system, that is, a technical whole in which a series of technical standards and intellectual property rights closely link materials, devices, processes, chips, complete machines, system software, application software, and so on. There are three paths for China to develop an intelligent computing technology system:
The first is to catch up with and be compatible with the A system led by the United States. Most Chinese internet companies have taken the GPGPU/CUDA-compatible path, and many entrepreneurial enterprises in the chip field also try to be as compatible with CUDA as possible in ecosystem building. This path is relatively realistic. Because in terms of computing power, the United States restricts China's processes and chip bandwidth; in terms of algorithms, the numerous domestic ecosystems make it difficult to form unity; and in terms of data, high-quality Chinese-language data is scarce. These factors will make it very difficult for catch-up players to narrow the gap with leaders, and at times will further widen it.
The second is to build a dedicated, closed B system. In specialized fields such as the military, meteorology, and the judiciary, build closed enterprise ecosystems, produce chips based on mature domestic processes, focus more on vertical large models for specific fields relative to foundation large models, and use more domain-specific high-quality data for training large models, among other things. This path makes it easy to form a complete and controllable technology system and ecosystem. Some of China's large backbone enterprises have taken this path. Its disadvantage is that it is closed, cannot unite most domestic forces, and is also difficult to globalize.
Third, jointly building an open-source and open C system globally. Use open source to break ecological monopolies, lower the threshold for enterprises to possess core technologies, and enable every enterprise to make its own chips at low cost, forming a vast ocean of intelligent chips and meeting ubiquitous intelligent needs. Use openness to form a unified technical system; Chinese enterprises and global forces should join together to build a unified intelligent computing software stack based on international standards. Form a pre-competitive sharing mechanism for enterprises, sharing high-quality databases and sharing open-source general foundation large models. As for the global open-source ecosystem, Chinese enterprises have benefited greatly in the Internet era, and China has mostly been a user and participant. In the intelligent era, Chinese enterprises should become more of the main contributors on the RISC-V⑥+AI open-source technology system, becoming the leading force in global open sharing.
Option Two: Compete on algorithm models, or compete on new-type infrastructure?
Artificial intelligence technology must empower all industries and has a typical long-tail effect⑦. Eighty percent of China's small, medium, and micro enterprises need low-threshold, low-price intelligent services. Therefore, China's intelligent computing industry must be built on a new data space infrastructure, and the key is that China should take the lead in fully infrastructuralizing intelligent elements, namely data, computing power, and algorithms. This work can be compared to the historic role of the early twentieth-century American Information Superhighway plan (that is, information infrastructure construction) for the Internet industry.
The most core productivity of the information society is cyberspace. The evolution of cyberspace is: from the computing space formed by the unary connection of machines, to the information space formed by the binary connection of humans and machine information, and then to the data space formed by the ternary connection of humans, machines, and object data. From the perspective of data space, the essence of artificial intelligence is the tempering of data into steel, and large models are the product after deep processing of the full volume of Internet data. In the digital era, what is transmitted on the Internet is the information stream, a structured abstraction after rough processing of data by computing power; in the intelligent era, what is transmitted on the Internet is the intelligence stream, a model-based abstraction after deep processing and refinement of data by computing power. A core feature of intelligent computing is to use algorithms such as numerical computation, data analysis, and artificial intelligence to process massive data pieces in a computing power pool, obtain intelligent models, and then embed them into various processes of the information world and the physical world.
The Chinese government has already laid out new-type infrastructure with foresight in advance, seizing the initiative in competition among countries around the world. First, data has become a national strategic information resource. Data has the dual attributes of resource element and value processing. The resource element attributes of data include all links such as production, acquisition, transmission, aggregation, circulation, transaction, ownership, assets, and security. China should continue to intensify efforts to build national data hubs and data circulation infrastructure.
Second, AI large models are a type of algorithmic infrastructure in data space. With general large models as the base, build the infrastructure for large model research and development and application, support the vast number of enterprises in developing domain-specific large models, serve industries such as robots, unmanned driving, wearable devices, smart homes, and intelligent security, and cover long-tail applications.
Finally, the construction of a nationally integrated computing power network has played a pioneering role in promoting the infrastructuralization of computing power. China's approach to the infrastructuralization of computing power should, while greatly reducing the cost and threshold of using computing power, provide high-throughput, high-quality intelligent services to the widest possible population. China's approach to computing power infrastructure needs to have "two lows and one high," namely: on the supply side, greatly reducing the total cost of computing power devices, computing power equipment, network connections, data acquisition, algorithm model invocation, electricity consumption, operations and maintenance, and development and deployment, so that the vast number of small and medium-sized enterprises can afford high-quality computing power services and are motivated to develop computing power network applications; on the consumption side, greatly reducing the threshold for the vast number of users to use computing power, with public services for the general public required to be easy to obtain and easy to use, ready to use like water and electricity, and as easy to customize computing power services and develop computing power network applications as writing a web page. On the service efficiency side, China's computing power services need to achieve low entropy and high throughput, where high throughput means that while achieving high-concurrency⑧ services, the response time of end-to-end services can have a high satisfaction rate; low entropy means that under high-concurrency loads where resources compete in a disorderly manner, the system throughput is ensured not to decline sharply. Ensuring "computing more" is especially important for China.
Option three: Should AI+ focus on empowering the virtual economy, or devote its efforts to the real economy?
The effectiveness of "AI+" is the touchstone of the value of artificial intelligence. After the subprime mortgage crisis, the share of U.S. manufacturing value added in GDP fell from 28% in 1950 to 11% in 2021, and the share of U.S. manufacturing employment in total industry employment fell from 35% in 1979 to 8% in 2022, showing that the United States is more inclined toward the virtual economy with higher returns and looks down on the real economy with high investment costs and low economic returns. China tends to develop the real economy and the virtual economy in parallel, and places greater emphasis on developing the real economy, including equipment manufacturing, new energy vehicles, photovoltaic power generation, lithium batteries, high-speed rail, and 5G.
Correspondingly, U.S. AI is mainly applied to the virtual economy and IT foundational tools, and AI technology has also become "detached from the real and oriented toward the virtual." Since 2007, Silicon Valley has continuously hyped virtual reality (VR), the metaverse, blockchain, Web3.0, deep learning, AI large models, and so on, which is a reflection of this trend.
China's advantage lies in the real economy. Its manufacturing industry has the most complete range of industrial categories and the most complete system in the world, characterized by many scenarios and abundant private data. China should select several industries and increase investment to form a paradigm that can be promoted across the entire industry with a low threshold, such as choosing the equipment manufacturing industry as a representative industry for sustaining its advantages and choosing the pharmaceutical industry as a representative industry for rapidly narrowing the gap. The technical difficulty in empowering the real economy is the integration of AI algorithms with physical mechanisms.
The key to the success of artificial intelligence technology is whether it can greatly reduce the cost of an industry or a product, thereby expanding the number of users and the scale of the industry by 10 times, producing a transformative effect similar to that of the steam engine on the textile industry and the smartphone on the internet industry.
China should embark on a high-quality development path of artificial intelligence empowering the real economy that suits its own conditions.
(The speaker is an academician of the Chinese Academy of Engineering, a researcher at the Institute of Computing Technology, Chinese Academy of Sciences, and director of its Academic Committee.)
Notes:
① Pattern recognition refers to using computational methods to classify samples into certain categories according to their features. It is the study of automatic processing and interpretation of patterns by computers using mathematical methods, and environments and objects are collectively referred to as "patterns." Its main research directions include image processing and computer vision, speech and language information processing, brain networks, and brain-inspired intelligence.
② Token can be translated as ciyuan, referring to symbols used to represent words or phrases in the process of natural language processing. A token can be a single character or a sequence composed of multiple characters.
③Artificial general intelligence refers to a type of artificial intelligence that possesses intelligence comparable to or even exceeding that of humans. Artificial general intelligence can not only perform basic thinking abilities such as perception, understanding, learning, and reasoning like humans, but also flexibly apply them across different domains, learn quickly, and think creatively. The research goal of artificial general intelligence is to seek a unified theoretical framework to explain various intelligent phenomena.
④Chip manufacturing process refers to the process used to manufacture CPUs or GPUs, that is, the size of transistor gate circuits, measured in nanometers. At present, the most advanced process achieving mass production internationally is represented by TSMC's 3nm. More advanced manufacturing processes can integrate more transistors into CPUs and GPUs, giving processors more functions, higher performance, smaller area, lower cost, and so on.
⑤CUDA is a parallel computing platform and programming model designed and developed by NVIDIA, including the CUDA instruction set architecture and the parallel computing engine inside the GPU. Developers can use the C language to write programs for the CUDA architecture, and the programs written can run with ultra-high performance on processors that support CUDA.
⑥RISC-V (pronounced "risk-five") is an open general-purpose instruction set architecture initiated by the University of California, Berkeley. Compared with other paid instruction sets, RISC-V allows anyone to freely use the RISC-V instruction set to design, manufacture, and sell chips and software.
⑦The long tail effect refers to the phenomenon in which products or services that were originally overlooked, with low sales but many varieties, due to their huge total volume, accumulate total revenue that exceeds that of mainstream products. In the Internet field, the long tail effect is particularly significant.
⑧High concurrency usually refers to ensuring through design that a system can process many requests in parallel at the same time.
Editor: Zhang Junming
Responsible Editor: Yu Hao
Source: National People's Congress of China website
April 30, 2024 11:24
Original link:http://www.npc.gov.cn/npc/c2/c30834/202404/t20240430_436915.html