2024 NPC Lecture: AI and China — History, Prospects, Challenges, Strategies and Legislation

Posted on 06/20/2024 by 高大伟 David Cowhig

Sun Ninghui’s lecture to the Standing Committee of the PRC National People’s Congress reflects Chinese hopes and concerns on AI, its applications, and its role in strengthening China in international commercial competition. Sun also refers to AI’s role as a leading new quality productive force and enabler of China’s high quality development, two themes that Chinese Communist Party General Secretary Xi Jinping often mentions.

This is one in a series of lectures give to the NPC Standing Committee. Several 2023 NPC lectures have been translated on this translation blog:

See also the analysis of Professor Sun’s lecture by Larry Catá Backer on his Law at the End of the Day blog.

Development of Artificial Intelligence and Intelligent Computing

人工智能与智能计算的发展

Recently, the National People’s Congress of China published Chinese Academy of Sciences Institute of Computing Technology Director Sun Ninghui‘s 孙凝晖 lecture manuscript “Development of Artificial Intelligence and Intelligent Computing” delivered at the special lecture of the Standing Committee of the 14th National People’s Congress. Here is the full text, allowing us to delve into the profound world of artificial intelligence together.

Chairman, Vice Chairpersons, Secretary-General, and Committee Members:

In recent years, the field of artificial intelligence (AI) has been experiencing an explosive growth led by generative AI models. On November 30, 2022, OpenAI launched an AI conversational chatbot called ChatGPT. Its outstanding natural language generation capabilities attracted worldwide attention, reaching 100 million users in just two months. This sparked a wave of large models globally.  The year  2022 has been hailed as Year Zero of Large Language Models (LLMs) with models such as Gemini, Wenxin Yiyan [Ernie Bot] 文心一言, Copilot, LLaMA, SAM, and SORA popping up like bamboo after a spring rain. The current information era is accelerating into the intelligent computing development stage, with continual breakthroughs in AI technology deeply empowering various industries and promoting AI and data elements as classic representatives of new quality productive forces 新质生产力. 

Communist Party General Secretary Xi Jinping has emphasized making the new generation of AI a driving force for leapfrog development of science and technology, industrial optimization, and overall productivity improvement, as China strives to achieve high-quality development 高质量发展. Ever since the 18th National Congress of the Communist Party of China, the central leadership with Xi Jinping at its core, has made the development of the intelligent economy a top priority. This has promoted the deep integration of AI with the real economy thereby creating a strong impetus for high-quality development.

1. Overview of Computing Technology Development

The history of computing technology can be roughly divided into four stages: 

  1. The first generation – the abacus and the mechanical computing era; 
  2. The second generation—the electronic computing era marked by the emergence of electronic devices and computers; 
  3. The third generation began with the advent of the internet —network computing; 
  4. The fourth generation which human society is now entering—intelligent computing.

Early computing devices were manual and semi-automatic calculation aids. The history of human computing tools started with the Chinese abacus around 1200 AD, followed by Napier’s bones (1612) and Pascal’s calculator – a wheel-based adding machine – in 1642. The first device to automatically perform the four basic arithmetic operations, the stepped reckoner, was born in 1672.

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 calculation. During this period, the concepts of programming and programs were basically formed, originating from the Jacquard loom, which controlled patterns using punched cards, eventually evolving into storing all mathematical computation steps in the form of computational instructions. The first programmer in human history was Ada Lovelace, the daughter of poet Byron, who wrote a set of computational instructions for Babbage’s difference engine to solve the Bernoulli number series. This set of instructions was also the first computer algorithm program in human history, separating hardware and software and introducing the concept of a program for the first time.

In the first half of the twentieth century, four scientific foundations of modern computing technology emerged: 

  1. Boolean algebra (mathematics), 
  2. Turing machine (computational model), 
  3. von Neumann architecture (architecture), and 
  4. Transistors (devices). 

Boolean algebra describes the underlying logic of programs and hardware such as the CPU;

The Turing machine is a universal computational model that transforms complex tasks into automated processes without human intervention; 

The von Neumann architecture proposed three basic principles for constructing computers: binary logic, stored program execution, and the computer comprising five basic units—arithmetic unit, control unit, memory, input device, and output device; 

The transistor is a semiconductor device that constitutes basic logic circuits and storage circuits, forming the “bricks” of the modern computer tower. 

Based on these scientific foundations, computing technology developed rapidly, forming a large-scale industry.

From the birth of the world’s first electronic computer ENIAC in 1946 to today, five types of successful platform-based computing systems have formed. Various applications across different fields can be supported by these five types of platform-based computing devices. 

  • The first type is high-performance computing platforms, solving 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, with computing platforms from companies like Baidu, Alibaba, and Tencent belonging to this type; 
  • The third type is personal computer platforms, appearing in the form of desktop applications where people interact with personal computers through desktop applications; 
  • The fourth type is smartphones, characterized by mobility and connectivity to data centers via the network, mainly focusing on internet applications, distributed across data centers and phone terminals; 
  • The fifth type is embedded computers, embedded in industrial equipment and military devices to ensure the completion of specific tasks within a certain time through real-time control. 

These five types of devices cover almost all aspects of our information society. However, the sixth type of platform-based computing system, centered on intelligent computing applications, has yet to form.

The development of modern computing technology can be roughly divided into three eras.

IT1.0, also known as the electronic computing era (1950-1970), was characterized by a focus on the “machine.” The basic architecture of computing technology was formed, and with the advancement of integrated circuit technology, the scale of basic computing units rapidly shrank, while the density, performance, and reliability of transistors continued to improve. Computers were widely used in scientific and engineering computations and enterprise data processing.

IT2.0, also known as the network computing era (1980-2020), put people at the center 以“人”为中心。The internet connected human-used terminals with back-end data centers, and internet applications interacted with humans through smart terminals. Internet companies like Amazon proposed the concept of cloud computing, encapsulating back-end computing as a service rented to third-party users, forming the cloud computing and big data industries.

IT3.0, also known as the intelligent computing era, began in 2020. Compared to IT2.0, it added the concept of the “Internet of Things,” where various end-side devices in the physical world are digitized, networked, and intelligent, achieving the “human-machine-things” triple integration. In the intelligent computing era, besides the internet, there are also data infrastructures supporting various terminals through edge-cloud integration, with AI embedded in terminals, edges, and clouds to provide large model intelligent services similar to ChatGPT. Eventually, wherever there is computing, there is AI intelligence. Intelligent computing brings a vast amount of data, breakthroughs in AI algorithms, and explosive demand for computing power.

2. Overview of Intelligent Computing Development

Intelligent computing includes AI technology and its computational carriers and has roughly undergone four stages: general-purpose computing devices, logic reasoning expert systems, deep learning computing systems, and large model computing systems.

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

The second stage of intelligent computing development is logic reasoning expert systems (1990). Scientists of the symbolic intelligence school, such as Edward Albert Feigenbaum, aimed to automate logic and reasoning capabilities and proposed expert systems capable of logical reasoning with knowledge symbols. Human prior knowledge entered the computer in the form of knowledge symbols, enabling the computer to assist humans in specific fields with logical judgments and decisions. However, expert systems heavily relied on manually generated knowledge bases or rule bases. The typical representatives of such expert systems are Japan’s Fifth Generation Computer Project and China’s 863 Program-supported 306 intelligent computer project. Japan used dedicated computing platforms and knowledge reasoning languages like Prolog to perform application-level reasoning tasks in logic expert systems. In contrast, China took a different technical route, based on general computing platforms, transforming intelligent tasks into AI algorithms, integrating hardware and system software into general computing platforms, leading to the emergence of core enterprises such as Sugon 曙光, Hanwang 汉王, and iFlytek 科大讯飞.。

The limitations of symbolic computing systems lie in the explosive increase in computing time required as spatiotemporal complexity increases. This means that symbolic computing systems can only solve problems that increase linearly in complexity.  They cannot solve high-dimensional complex space problems. This limits the problem size they can handle. Additionally, because symbolic computing systems are based on knowledge rules, we cannot enumerate all common sense exhaustively. This also significantly restricts the scope of their applications. With the arrival of the second AI winter, the first generation of intelligent computers gradually exited the stage of history.

Around 2014, intelligent computing advanced to the third stage—deep learning computing systems. Represented by scientists like Geoffrey Hinton of the connectionist school, the goal was to automate learning capabilities, inventing new AI algorithms such as deep learning algorithms. Through the automatic learning of deep neural networks, the model’s statistical inductive capabilities were significantly improved, achieving substantial breakthroughs in applications such as pattern recognition, with recognition accuracy in some scenarios even surpassing humans. 

For example, in facial recognition, the entire neural network training process is akin to adjusting network parameters, inputting a large amount of labeled facial image data into the neural network, adjusting the parameters between networks, making the network output’s result probability arbitrarily close to the actual result. The greater the probability that the neural network outputs the actual situation, the greater is the weighting of the parameters, thus encoding knowledge and rules into the network parameters. As long as the data is adequate, various common sense matters can be learned, greatly enhancing generality. Connectionist applications are more extensive than conventional applications. These include speech recognition, facial recognition, and autonomous driving. Regarding computing devices, the Institute of Computing Technology of the Chinese Academy of Sciences proposed the world’s first deep learning processor chip architecture in 2013. Internationally renowned hardware manufacturers like NVIDIA have continuously released multiple high-performance general Graphic Processing Unit (GPU) chips, representing deep learning computing systems.

The fourth stage of intelligent computing development is large model computing systems (2020). Driven by AI large model technology, intelligent computing has reached new heights. In 2020, AI transitioned from “small + discriminative models” to “large + generative models” 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, with 300 billion words of training data equivalent to all the English text on the internet. Its basic principle is: given an input, it predicts the next word to train the model, improving prediction accuracy through massive training, ultimately allowing real-time dialogue by asking it a question. Based on the base large model, it then provides some prompt words for supervised instruction fine-tuning, gradually learning how to conduct multi-turn dialogues with humans through human-instructed <commands, replies>; finally, reinforcement learning iteration is performed through manually designed and automatically generated reward functions, gradually aligning the large model with human values.

The large model’s characteristics are its “large” scale, with three implications:

  • Large parameters, GPT-3 has 170 billion parameters; 
  • Large training data, ChatGPT used about 300 billion words, 570GB of training data; 
  • Large computing power demand, GPT-3 used tens of thousands of V100 GPUs for training. 

To meet the explosive increase in intelligent computing power demand brought by large models, large-scale new intelligent computing centers are being constructed domestically and internationally, with NVIDIA launching large model intelligent computing systems comprising 256 H100 chips and 150TB of massive GPU memory.

The advent of large models has brought three transformations.

First, the Neural Scaling Law of this technology, where the accuracy of many AI models rapidly improves once the parameter scale exceeds a certain threshold. The scientific reasons for this are still much debated. The performance of AI models is log-linear  with the model parameter scale, dataset size, and total computing power. Therefore, increasing the model’s scale continuously improves performance. Currently, the most advanced large model, GPT-4, has parameters reaching trillions to tens of trillions and is still growing.

Second, explosive growth in computing power demand, training large models with hundreds of billions of parameters typically requires training on thousands to tens of thousands of GPU cards for 2-3 months, leading to the rapid development of companies that provide computing power. NVIDIA’s market value approaches two trillion dollars, unprecedented for a chip company.

Third, the impact on the labor market. A report by the National School of Development at Peking University and Zhaopin.com “A Study of the Potential Impact of the AI Megamodel on Our Labor Market”《AI大模型对我国劳动力市场潜在影响研究》indicated that the top 20 occupations most affected include accounting, sales, and clerical work, while jobs involving physical labor and providing services, such as human resources, administration, and logistics, are relatively insulated from the impact of AI.

AI technology is expected to develop on four frontiers.

The first frontier is multi-modal large models. From a human perspective, human intelligence is inherently multi-modal, with humans possessing eyes, ears, nose, tongue, body, and mouth (language). From an AI perspective, vision and hearing can also be modeled as token sequences, which can be learned using the same methods as large language models and further aligned with semantics in language to achieve multi-modal aligned intelligence.

The second frontier is video generation large models. On February 15, 2024, OpenAI released the SORA text-to-video model, significantly increasing video generation duration from a few seconds to a minute, with notable improvements in resolution, realism, and temporal consistency. SORA’s greatest significance is its basic characteristics of a world model, i.e., the ability to observe and further predict the world, based on basic physical common sense (e.g., water flows downhill). Although SORA still faces many challenges in becoming a model of the real world, it has learned imaginative and minute-scale future prediction capabilities. These are foundational characteristics for a real-world model.

The third frontier is embodied intelligence. This refers to intelligent agents with bodies supporting interaction with the physical world, such as robots and autonomous vehicles. These agents process multimodal sensory data inputs using large models to generate motion instructions, replacing traditional rule-based or mathematical formula-based motion driving methods, achieving deep integration of virtual and real worlds. Robots with embodied intelligence can integrate the three main AI schools: connectionism represented by neural networks, knowledge representation represented by knowledge engineering, and behaviorism related to cybernetics. These three schools, which can simultaneously influence an intelligent agent, are expected to bring new technological breakthroughs.

The fourth frontier is AI4R (AI for Research) which is becoming a main paradigm for scientific discovery and technical invention. Current scientific discovery mainly relies on experiments and human intelligence, with humans making bold hypotheses and careful verifications. Information technology, whether computing or data, only plays a supporting and verification role. Compared to humans, AI has significant advantages in memory, handling complex high-dimensional tasks, full-view analysis, deep reasoning, and hypothesis-making. Whether AI can primarily conduct scientific discoveries and technological inventions, greatly improving human scientific discovery efficiency, is a key question, such as proactively discovering physical laws, predicting protein structures, designing high-performance chips, and efficiently synthesizing new drugs. Since AI large models have full data and a god’s-eye view, they can look further ahead than humans. If AI models can achieve a leap from inference to reasoning, they might possess the imagination and scientific hypothesis generating abilities of an Einstein. This would significantly enhance the efficiency of human scientific discovery and overcome human cognitive limitations. That would truly be a disruptive innovation. 

Lastly, Artificial General Intelligence (AGI) is a highly challenging and controversial topic. A philosopher and a neuroscientist once made a bet on whether scientists could reveal how the brain achieves consciousness in 25 years (i.e., by 2023). At the time, there were two main theories about consciousness: Integrated Information Theory and Global Workspace Theory. The former suggested consciousness is a “structure” formed by specific types of neuron connections in the brain, while the latter proposed consciousness arises when information spreads through an interconnected network to brain regions. In 2023, adversarial experiments conducted by six independent labs could not fully confirm either theory and so the philosopher won and the neuroscientist lost. This bet reflected the hope that AI can understand human cognition and the mysteries of the human brain. 

From a physics perspective, understanding the macroscopic world led to the understanding of the microscopic world, starting with quantum physics. The intelligent world, like the physical world, is a complex research object. AI large models are still driven by research, including data-driven studies, of the macro-scale world to improve machine intelligence. These models have limited understanding of the macro-scale world of intelligence. This makes it difficult to find answers directly in the micro-scale world of the neural networks. Since its birth, AI has brought dreams and fantasies about the nature of human intelligence and consciousness. These dreams and fantasies continually inspire further exploration. 

3. AI Security Risks

AI development has promoted technological progress worldwide but also brought many security risks, requiring responses from both technical and regulatory aspects.

First, the proliferation of false information on the internet. Here are several scenarios:

  • Digital doppelgangers. AI Yoon was the first official “candidate” synthesized using deep fake technology, based on South Korean People Power Party candidate Yoon Suk-yeol. Using 20 hours of Yoon’s audio and video clips and over 3,000 sentences specially recorded by Yoon for researchers, a virtual image of AI Yoon was created by a local deep fake technology company and quickly went viral online. The content expressed by AI Yoon was written by the campaign team, not the candidate himself.
  • Fake videos, especially of leaders, causing international disputes, disrupting election orders, or triggering sudden public opinion incidents, such as fake videos of Nixon announcing the first moon landing failure or Ukrainian President Zelensky announcing “surrender,” leading to a decline in social trust in the news media industry.
  • Fake news, generated automatically for illegal profit, using ChatGPT to create trending news and earn traffic. As of June 30, 2023, there were 277 fake news websites globally, seriously disrupting social order.
  • Face-swapping and voice-changing for fraud. For instance, using AI voice imitation of a company executive, a Hong Kong international company was scammed out of $35 million.
  • Generating obscene images, especially targeting public figures, such as making pornographic videos of movie stars, causing adverse social impacts. Therefore, it is urgent to develop technology to detect fake information on the internet.

Second, AI large models face serious trust issues. These include: 

  1. Factual errors spoken with confidence; 
  2. Outputting political biases and erroneous statements based on Western values
  3. Easily induced to output incorrect knowledge and harmful content; 
  4. Increased data security issues, with large models becoming traps for sensitive data, as ChatGPT incorporates user input into its training database for improvement. This technology would enable the U.S. to obtain Chinese language material from non-public channels, mastering “China knowledge” that we might not even possess ourselves. Hence, it is urgent to develop large model safety regulatory technology and trusted large models.

Besides technical measures, AI safety assurance requires relevant legislation. 

Western countries have also enacted regulations

. 

China should expedite the introduction of an “Artificial Intelligence Law” to establish an AI governance system, to 

  • Ensure that AI development and applications adhere to shared human values, 
  • Promote human-machine harmony and friendship 促进人机和谐友好 ; 
  • Create a favorable policy environment for AI technology research, development, and applications; Establishing reasonable disclosure and audit mechanisms to understand AI system mechanism, principles and decision-making processes; 
  • Clarifying AI system safety responsibilities and accountability mechanisms, 
  • Ensure traceable responsibility and remediation; and 
  • Promote fair, reasonable, open, and inclusive international AI governance rules.

4. Development Challenges of Intelligent Computing in China

AI technology and the intelligent computing industry are at the focus of China-U.S. technological competition. Although China has made significant achievements in recent years, it still faces many development challenges, particularly those arising from U.S. technology trade suppression policies.

The first challenge is the U.S. long-term lead in AI core capabilities, with China in follower mode. There is a certain gap between China and the U.S. in terms of high-end AI talent, foundational AI algorithm innovation, base large model capabilities (large language models, text-to-image models, text-to-video models), training data for base large models, and computing power for base large models. This gap is expected to persist for a long time.

The second challenge is the ban on high-end computing power products and long-term restrictions on advanced chip manufacturing processes. High-end smart computing chips like A100, H100, and B200 are banned from being sold to China. Companies such as Huawei, Loongson, Cambricon, Sugon, and Hygon are on the entity list, constraining their use of advanced manufacturing processes. Domestic manufacturing processes for large-scale production lag behind international advanced levels by 2-3 generations, and the performance of core computing power chips lags behind by 2-3 generations.

The third challenge is the weak domestic intelligent computing ecosystem and insufficient penetration of AI development frameworks. NVIDIA’s CUDA (Compute Unified Device Architecture) ecosystem is complete and has formed a defacto monopoly. The domestic ecosystem is weak, specifically in:

  • Shortage of R&D personnel, with nearly 20,000 developers in the NVIDIA CUDA ecosystem, 20 times the total of all Chinese domestic intelligent chip companies; 
  • Shortage of development tools, with CUDA having 550 SDKs (Software Development Kits), hundreds of times more than those of Chinese domestic companies; 
  • Inadequate funding, with NVIDIA investing $5 billion annually, dozens of times more than domestic companies; 
  • AI development frameworks like TensorFlow dominate the industrial market and PyTorch the research market, while domestic frameworks like Baidu’s PaddlePaddle have only one-tenth the developers of foreign frameworks. 

An even more serious problem is the fragmentation of Chinese domestic companies.  They are unable to form a cohesive force. Although each layer has related products, from intelligent applications, development frameworks, system software, to intelligent chips, there is no deep integration between layers and so they are unable to form a competitive technical system.

The fourth challenge is the high cost and barriers to AI application in industries. Currently, AI applications in China are mainly concentrated in the internet and some defense fields. When promoting AI technology across various industries, especially transitioning from the internet to non-internet industries, extensive customization work is required, with high migration difficulty and single-use costs. Additionally, the number of AI workers in China is small compared to actual demand.

5. Path Choices for Developing Intelligent Computing in China

The choice of path for AI development is crucial for China. It affects sustainability and the ultimate shape of the international competitive landscape. The current cost of using AI is very high, with Microsoft Copilot suite requiring a monthly fee of $10, ChatGPT consuming 500,000 kWh of electricity daily, and NVIDIA’s B200 chip priced over $30,000. Overall, China should develop affordable, safe, and trusted AI technology, eliminate information poverty in China and share benefits with the Belt and Road countries. This will enable companies operating in industries with a low barrier to survive, enable Chinese companies in industries where they have a comparative advantage to remain competitive, and significantly reduce the competitiveness gap for Chinese companies in operating in industries where China is relatively backward.

Choice One: Adopt a closed, proprietary technological system or an open-source, open path system?

The intelligent computing industry is supported by a tightly coupled technological system, an integrated whole of technical standards and materials, devices, processes, chips, machines, system software, application software, etc. that embody intellectual property. 

China has three alternative paths it could follow to develop an intelligent computing technology system:

1. Catching up by being compatible with the U.S.-dominated alternative A system. Most Chinese internet companies are following the General-Purpose Graphics Processing Unit (GPGPU)/CUDA compatible path, and many startup companies in the chip field are also trying to be compatible with CUDA in ecosystem construction. This path is relatively realistic. However, due to U.S. restrictions on our process and chip bandwidth in computing power, fragmented domestic ecosystems that are difficult to mesh, and a shortage of high-quality Chinese language data, it is challenging to narrow the gap between followers and leaders. Sometimes the gap even yawns wider.

2. Constructing a proprietary, closed alternative B system. In specialized fields like military, meteorology, and judiciary, build a closed ecosystem based on domestically mature processes, focusing more on specific vertical models than base large models, and using specialized high-quality data for training. This path can form a complete, controllable technical system and ecosystem. Some large backbone enterprises in China follow this path. Its disadvantage is being closed, unable to absorb benefits from most domestic sources and challenging to globalize.

3. Globally building an open-source alternative C system. Use open source to break ecosystem monopolies, lowering the threshold for enterprises to own core technologies, allowing each company to make its chips at low costs, creating a vast sea of intelligent chips to meet ubiquitous intelligent needs. Form an open unified technical system, where Chinese enterprises unite with global forces to co-build a unified intelligent computing software stack based on international standards. Establish pre-competitive sharing mechanisms for enterprises, sharing high-quality databases and open-source general base large models. In the internet era, Chinese enterprises benefited greatly from global open-source ecosystems as users and participants. In the intelligent era, Chinese companies should become major contributors to the RISC-V + AI open-source technology system, becoming leading forces in global open-source and shared ecosystems.

Choice Two: Compete in algorithm models or in new infrastructure?

AI technology needs to empower various industries, exhibiting typical long-tail effects. China’s 80% of small and medium-sized enterprises require low-threshold, low-cost intelligent services. Thus, China’s intelligent computing industry must be based on new data space infrastructure, with the key being the comprehensive infrastructure of AI elements, namely data, computing power, and algorithms. This task can be compared to the historical impact of the U.S. information superhighway plan (information infrastructure construction) on the internet industry in the early 20th century.

The core productive force of the information society is cyberspace. The evolution process of cyberspace is from the computing space formed by machine connections, to the information space formed by human-machine information connections, and then to the data space formed by human-machine-Internet of Things device data connections. From the data space perspective, the essence of AI is refining vast quantities of data into a finished product, with large models being the product of deep processing of the entire Internet’s data. In the digital age, information flows transmitted on the internet are structured abstractions of data processed by computing power. In the intelligent age, intelligent flows transmitted on the internet are model abstractions of data deeply processed and refined by computing power. A core feature of intelligent computing is processing massive collections of data in a computing power pool using numerical computation, data analysis, and AI algorithms, to obtain intelligent models embedded in various processes of the information and physical worlds.

The Chinese government has proactively laid out new infrastructure, gaining an edge in global competition.

First, data has become a national strategic information resource. Data attributes include resource elements and value processing, encompassing production, acquisition, transmission, aggregation, circulation, transaction, ownership, assets, and security. China should continue to build national data hubs and data circulation infrastructure.

Second, AI large models are a type of algorithmic infrastructure in the data space. Constructing large model R&D and application infrastructure based on general large models, supporting enterprises in developing specialized large models for industries such as robotics, autonomous driving, wearables, smart homes, and smart security and for covering long-tail applications.

Lastly, the construction of a nationwide integrated computing power network leading the way in creating a strong computing power infrastructure. China’s computing power infrastructure plan should significantly reduce the cost and threshold of using computing power, providing high-throughput, high-quality intelligent services to the broadest population. 

China’s computing power infrastructure plan needs “two lows and one high”: 

  • On the supply side, significantly reducing the total cost of computing power devices, equipment, network connections, data acquisition, algorithm model calls, power consumption, operation and maintenance, and development and deployment, enabling small and medium enterprises to afford high-quality computing power services and develop computing power network applications; 
  • On the consumption side, significantly lowering the threshold for using computing power, with public services being easily accessible and usable, like water and electricity, and computing power services being as easily customizable as writing web pages. 
  • In service efficiency, China’s computing power services should achieve low entropy and high throughput, with high throughput meaning meeting high concurrency service needs while maintaining acceptable end-to-end service response times. Low entropy ensures the system’s throughput does not plummet in high-concurrency load situations with disordered resource competition. This will ensure the “computing power” that is crucial for China.

Choice Three: Emphasize enabling the virtual economy or focus on the real economy?

“AI+” success is the litmus test for AI’s value. After its subprime mortgage crisis, the U.S. saw its manufacturing value-added as a proportion of GDP drop from 28% in 1950 to  just11% in 2021 while manufacturing employment dropped from 35% in 1979 to 8% in 2022. This shows that the U.S. prefers higher-return virtual economies but overlooks high-investment, low-return real economies. China prefers simultaneous development of both the real and the virtual economies, emphasizing equipment manufacturing, new energy vehicles, photovoltaic power generation, lithium batteries, high-speed rail, and 5G.

Accordingly, U.S. AI primarily applies to virtual economies and IT infrastructure tools, with AI technology “moving from reality to virtuality,” reflecting trends since 2007 of Silicon Valley hyping virtual reality, the metaverse, blockchain, Web3.0, deep learning, and large language models for AI.

China’s strength lies in its real economy, with the most comprehensive and complete manufacturing industry globally, characterized by diverse scenarios and abundant private data. China should select representative industries for significant investment to form models easily promotable across industries with low entry thresholds, such as equipment manufacturing to hold on to its advantages there and the pharmaceutical industry to quickly narrow the gap with foreign companies. The technical difficulties of supporting the real economy lies in integrating AI algorithms with physical mechanisms.

The key to AI technology success is whether it can significantly reduce costs in an industry or in producing a product, expanding user numbers and enable companies to greatly scale up their production. This can be transformative  just as the steam engine was for the textile industry and smartphones have been for the internet industry.

China should pursue a high-quality development path for AI enabling the real economy.

Notes:

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

2. Token refers to symbols representing words or phrases in natural language processing. Tokens can be single characters or sequences of multiple characters.

3. AGI refers to AI with intelligence comparable to or surpassing humans. AGI can perceive, understand, learn, and reason like humans and flexibly apply, quickly learn, and think creatively in various fields. AGI research aims to seek a unified theoretical framework to explain various intelligent phenomena.

4. Chip manufacturing process refers to the process of making CPUs or GPUs, i.e., the size of transistor gate circuits, measured in nanometers. The most advanced mass-produced process internationally is TSMC‘s 3nm. More advanced processes can integrate more transistors within CPUs and GPUs, providing more functions and higher performance, with smaller size and lower cost.

5. CUDA 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 write programs for CUDA architecture using the C language, running at ultra-high performance on CUDA-supported processors.

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

7. Long-tail effect refers to the phenomenon where products or services with small sales but numerous varieties, which were previously neglected, accumulate total revenue greater than mainstream products due to their massive total volume. The long-tail effect is particularly evident in the internet field.

8. High concurrency usually refers to designing systems to handle many requests simultaneously.

人工智能与智能计算的发展

 

作者：中国储能网新闻中心

  中国储能网讯：中国人大网近日刊登孙凝晖在十四届全国人大常委会专题讲座上的讲稿《人工智能与智能计算的发展》，现将全文转载如下，让我们一同走进高深莫测的人工智能世界。

  委员长、各位副委员长、秘书长、各位委员：

  人工智能领域近年来正在迎来一场由生成式人工智能大模型引领的爆发式发展。2022年11月30日，OpenAI公司推出一款人工智能对话聊天机器人ChatGPT，其出色的自然语言生成能力引起了全世界范围的广泛关注，2个月突破1亿用户，国内外随即掀起了一场大模型浪潮，Gemini、文心一言、Copilot、LLaMA、SAM、SORA等各种大模型如雨后春笋般涌现，2022年也被誉为大模型元年。当前信息时代正加快进入智能计算的发展阶段，人工智能技术上的突破层出不穷，逐渐深入地赋能千行百业，推动人工智能与数据要素成为新质生产力的典型代表。习近平总书记指出，把新一代人工智能作为推动科技跨越发展、产业优化升级、生产力整体跃升的驱动力量，努力实现高质量发展。党的十八大以来，以习近平同志为核心的党中央高度重视智能经济发展，促进人工智能和实体经济深度融合，为高质量发展注入强劲动力。

  1 计算技术发展简介

  计算技术的发展历史大致可分为四个阶段，算盘的出现标志着人类进入第一代——机械计算时代，第二代——电子计算的标志是出现电子器件与电子计算机，互联网的出现使我们进入第三代——网络计算，当前人类社会正在进入第四阶段——智能计算。

  早期的计算装置是手动辅助计算装置和半自动计算装置，人类计算工具的历史是从公元1200年的中国算盘开始，随后出现了纳皮尔筹（1612年）和滚轮式加法器（1642年），到1672年第一台自动完成四则运算的计算装置——步进计算器诞生了。

  机械计算时期已经出现了现代计算机的一些基本概念。查尔斯∙巴贝奇（Charles Babbage）提出了差分机（1822年）与分析机（1834年）的设计构想，支持自动机械计算。这一时期，编程与程序的概念基本形成，编程的概念起源于雅卡尔提花机，通过打孔卡片控制印花图案，最终演变为通过计算指令的形式来存储所有数学计算步骤；人类历史的第一个程序员是诗人拜伦之女艾达（Ada），她为巴贝奇差分机编写了一组求解伯努利数列的计算指令，这套指令也是人类历史上第一套计算机算法程序，它将硬件和软件分离，第一次出现程序的概念。

  直到在二十世纪上半叶，出现了布尔代数(数学)、图灵机(计算模型) 、冯诺依曼体系结构(架构) 、晶体管(器件)这四个现代计算技术的科学基础。其中，布尔代数用来描述程序和硬件如CPU的底层逻辑；图灵机是一种通用的计算模型，将复杂任务转化为自动计算、不需人工干预的自动化过程；冯诺依曼体系结构提出了构造计算机的三个基本原则：采用二进制逻辑、程序存储执行、以及计算机由运算器、控制器、存储器、输入设备、输出设备这五个基本单元组成；晶体管是构成基本的逻辑电路和存储电路的半导体器件，是建造现代计算机之塔的“砖块”。基于以上科学基础，计算技术得以高速发展，形成规模庞大的产业。

  从1946年世界上第一台电子计算机ENIAC诞生到二十一世纪的今天，已经形成了五类成功的平台型计算系统。当前各领域各种类型的应用，都可以由这五类平台型计算装置支撑。第一类是高性能计算平台，解决了国家核心部门的科学与工程计算问题；第二类是企业计算平台，又称服务器，用于企业级的数据管理、事务处理，当前像百度、阿里和腾讯这些互联网公司的计算平台都属于这一类；第三类是个人电脑平台，以桌面应用的形式出现，人们通过桌面应用与个人电脑交互；第四类是智能手机，主要特点是移动便携，手机通过网络连接数据中心，以互联网应用为主，它们分布式地部署在数据中心和手机终端；第五类是嵌入式计算机，嵌入到工业装备和军事设备，通过实时的控制，保障在确定时间内完成特定任务。这五类装置几乎覆盖了我们信息社会的方方面面，长期以来人们追求的以智能计算应用为中心的第六类平台型计算系统尚未形成。

  现代计算技术的发展大致可以划分为三个时代。

  IT1.0又称电子计算时代（1950-1970），基本特征是以“机”为中心。计算技术的基本架构形成，随着集成电路工艺的进步，基本计算单元的尺度快速微缩，晶体管密度、计算性能和可靠性不断提升，计算机在科学工程计算、企业数据处理中得到了广泛应用。

  IT2.0又称网络计算时代（1980-2020），以“人”为中心。互联网将人使用的终端与后台的数据中心连接，互联网应用通过智能终端与人进行交互。以亚马逊等为代表的互联网公司提出了云计算的思想，将后台的算力封装成一个公共服务租借给第三方用户，形成了云计算与大数据产业。

  IT3.0又称智能计算时代，始于2020年，与IT2.0相比增加了“物”的概念，即物理世界的各种端侧设备，被数字化、网络化和智能化，实现“人-机-物”三元融合。智能计算时代，除了互联网以外，还有数据基础设施，支撑各类终端通过端边云实现万物互联，终端、物端、边缘、云都嵌入AI，提供与ChatGPT类似的大模型智能服务，最终实现有计算的地方就有AI智能。智能计算带来了巨量的数据、人工智能算法的突破和对算力的爆发性需求。

  2 智能计算发展简介

  智能计算包括人工智能技术与它的计算载体，大致历经了四个阶段，分别为通用计算装置、逻辑推理专家系统、深度学习计算系统、大模型计算系统。

  智能计算的起点是通用自动计算装置（1946年）。艾伦·图灵（Alan Turing）和冯·诺依曼（John von Neumann）等科学家，一开始都希望能够模拟人脑处理知识的过程，发明像人脑一样思考的机器，虽未能实现，但却解决了计算的自动化问题。通用自动计算装置的出现，也推动了1956年人工智能（AI）概念的诞生，此后所有人工智能技术的发展都是建立在新一代计算设备与更强的计算能力之上的。

  智能计算发展的第二阶段是逻辑推理专家系统（1990年）。E.A.费根鲍姆（Edward Albert Feigenbaum）等符号智能学派的科学家以逻辑和推理能力自动化为主要目标，提出了能够将知识符号进行逻辑推理的专家系统。人的先验知识以知识符号的形式进入计算机，使计算机能够在特定领域辅助人类进行一定的逻辑判断和决策，但专家系统严重依赖于手工生成的知识库或规则库。这类专家系统的典型代表是日本的五代机和我国863计划支持的306智能计算机主题，日本在逻辑专家系统中采取专用计算平台和Prolog这样的知识推理语言完成应用级推理任务；我国采取了与日本不同的技术路线，以通用计算平台为基础，将智能任务变成人工智能算法，将硬件和系统软件都接入通用计算平台，并催生了曙光、汉王、科大讯飞等一批骨干企业。

  符号计算系统的局限性在于其爆炸的计算时空复杂度，即符号计算系统只能解决线性增长问题，对于高维复杂空间问题是无法求解的，从而限制了能够处理问题的大小。同时因为符号计算系统是基于知识规则建立的，我们又无法对所有的常识用穷举法来进行枚举，它的应用范围就受到了很大的限制。随着第二次AI寒冬的到来，第一代智能计算机逐渐退出历史舞台。

  直到2014年左右，智能计算进阶到第三阶段——深度学习计算系统。以杰弗里·辛顿（Geoffrey Hinton）等为代表的连接智能学派，以学习能力自动化为目标，发明了深度学习等新AI算法。通过深度神经元网络的自动学习，大幅提升了模型统计归纳的能力，在模式识别①等应用效果上取得了巨大突破，某些场景的识别精度甚至超越了人类。以人脸识别为例，整个神经网络的训练过程相当于一个网络参数调整的过程，将大量的经过标注的人脸图片数据输入神经网络，然后进行网络间参数调整，让神经网络输出的结果的概率无限逼近真实结果。神经网络输出真实情况的概率越大，参数就越大，从而将知识和规则编码到网络参数中，这样只要数据足够多，就可以对各种大量的常识进行学习，通用性得到极大的提升。连接智能的应用更加广泛，包括语音识别、人脸识别、自动驾驶等。在计算载体方面，中国科学院计算技术研究所2013年提出了国际首个深度学习处理器架构，国际知名的硬件厂商英伟达（NVIDIA）持续发布了多款性能领先的通用GPU芯片，都是深度学习计算系统的典型代表。

  智能计算发展的第四阶段是大模型计算系统（2020年）。在人工智能大模型技术的推动下，智能计算迈向新的高度。2020年，AI从“小模型+判别式”转向“大模型+生成式”，从传统的人脸识别、目标检测、文本分类，升级到如今的文本生成、3D数字人生成、图像生成、语音生成、视频生成。大语言模型在对话系统领域的一个典型应用是OpenAI公司的ChatGPT，它采用预训练基座大语言模型GPT-3，引入3000亿单词的训练语料，相当于互联网上所有英语文字的总和。其基本原理是：通过给它一个输入，让它预测下一个单词来训练模型，通过大量训练提升预测精确度，最终达到向它询问一个问题，大模型产生一个答案，与人即时对话。在基座大模型的基础上，再给它一些提示词进行有监督的指令微调，通过人类的<指令，回复>对逐渐让模型学会如何与人进行多轮对话；最后，通过人为设计和自动生成的奖励函数来进行强化学习迭代，逐步实现大模型与人类价值观的对齐。

  大模型的特点是以“大”取胜，其中有三层含义，（1）参数大，GPT-3就有1700亿个参数；（2）训练数据大，ChatGPT大约用了3000亿个单词，570GB训练数据；（3）算力需求大，GPT-3大约用了上万块V100 GPU进行训练。为满足大模型对智能算力爆炸式增加的需求，国内外都在大规模建设耗资巨大的新型智算中心，英伟达公司也推出了采用256个H100芯片，150TB海量GPU内存等构成的大模型智能计算系统。

  大模型的出现带来了三个变革。

  一是技术上的规模定律（Scaling Law），即很多AI模型的精度在参数规模超过某个阈值后模型能力快速提升，其原因在科学界还不是非常清楚，有很大的争议。AI模型的性能与模型参数规模、数据集大小、算力总量三个变量成“对数线性关系”，因此可以通过增大模型的规模来不断提高模型的性能。目前最前沿的大模型GPT-4参数量已经达到了万亿到十万亿量级，并且仍在不断增长中；

  二是产业上算力需求爆炸式增长，千亿参数规模大模型的训练通常需要在数千乃至数万GPU卡上训练2-3个月时间，急剧增加的算力需求带动相关算力企业超高速发展，英伟达的市值接近两万亿美元，对于芯片企业以前从来没有发生过；

  三是社会上冲击劳动力市场，北京大学国家发展研究院与智联招聘联合发布的《AI大模型对我国劳动力市场潜在影响研究》报告指出，受影响最大的20个职业中财会、销售、文书位于前列，需要与人打交道并提供服务的体力劳动型工作，如人力资源、行政、后勤等反而相对更安全。

  人工智能的技术前沿将朝着以下四个方向发展。

  第一个前沿方向为多模态大模型。从人类视角出发，人类智能是天然多模态的，人拥有眼、耳、鼻、舌、身、嘴(语言)，从AI视角出发，视觉，听觉等也都可以建模为token②的序列，可采取与大语言模型相同的方法进行学习，并进一步与语言中的语义进行对齐，实现多模态对齐的智能能力。

  第二个前沿方向为视频生成大模型。OpenAI于2024年2月15日发布文生视频模型SORA，将视频生成时长从几秒钟大幅提升到一分钟，且在分辨率、画面真实度、时序一致性等方面都有显著提升。SORA的最大意义是它具备了世界模型的基本特征，即人类观察世界并进一步预测世界的能力。世界模型是建立在理解世界的基本物理常识（如，水往低处流等）之上，然后观察并预测下一秒将要发生什么事件。虽然SORA要成为世界模型仍然存在很多问题，但可以认为SORA学会了画面想象力和分钟级未来预测能力，这是世界模型的基础特征。

  第三个前沿方向为具身智能。具身智能指有身体并支持与物理世界进行交互的智能体，如机器人、无人车等，通过多模态大模型处理多种传感数据输入，由大模型生成运动指令对智能体进行驱动，替代传统基于规则或者数学公式的运动驱动方式，实现虚拟和现实的深度融合。因此，具有具身智能的机器人，可以聚集人工智能的三大流派：以神经网络为代表的连接主义，以知识工程为代表的符号主义和控制论相关的行为主义，三大流派可以同时作用在一个智能体，这预期会带来新的技术突破。

  第四个前沿方向是AI4R(AI for Research)成为科学发现与技术发明的主要范式。当前科学发现主要依赖于实验和人脑智慧，由人类进行大胆猜想、小心求证，信息技术无论是计算和数据，都只是起到一些辅助和验证的作用。相较于人类，人工智能在记忆力、高维复杂、全视野、推理深度、猜想等方面具有较大优势，是否能以AI为主进行一些科学发现和技术发明，大幅提升人类科学发现的效率，比如主动发现物理学规律、预测蛋白质结构、设计高性能芯片、高效合成新药等。因为人工智能大模型具有全量数据，具备上帝视角，通过深度学习的能力，可以比人向前看更多步数，如能实现从推断(inference)到推理(reasoning)的跃升，人工智能模型就有潜力具备爱因斯坦一样的想象力和科学猜想能力，极大提升人类科学发现的效率，打破人类的认知边界。这才是真正的颠覆所在。

  最后，通用人工智能③（Artificial General Intelligence，简称AGI）是一个极具挑战的话题，极具争论性。曾经有一个哲学家和一个神经科学家打赌：25年后（即2023年）科研人员是否能够揭示大脑如何实现意识？当时关于意识有两个流派，一个叫集成信息理论，一个叫全局网络工作空间理论，前者认为意识是由大脑中特定类型神经元连接形成的“结构”，后者指出意识是当信息通过互连网络传播到大脑区域时产生的。2023年，人们通过六个独立实验室进行了对抗性实验，结果与两种理论均不完全匹配，哲学家赢了，神经科学家输了。通过这一场赌约，可以看出人们总是希望人工智能能够了解人类的认知和大脑的奥秘。从物理学的视角看，物理学是对宏观世界有了透彻理解后，从量子物理起步开启了对微观世界的理解。智能世界与物理世界一样，都是具有巨大复杂度的研究对象，AI大模型仍然是通过数据驱动等研究宏观世界的方法，提高机器的智能水平，对智能宏观世界理解并不够，直接到神经系统微观世界寻找答案是困难的。人工智能自诞生以来，一直承载着人类关于智能与意识的种种梦想与幻想，也激励着人们不断探索。

  3 人工智能的安全风险

  人工智能的发展促进了当今世界科技进步的同时，也带来了很多安全风险，要从技术与法规两方面加以应对。

  首先是互联网虚假信息泛滥。这里列举若干场景：

  一是数字分身。AI Yoon是首个使用 DeepFake 技术合成的官方“候选人”，这个数字人以韩国国民力量党候选人尹锡悦（Yoon Suk-yeol）为原型，借助尹锡悦 20 小时的音频和视频片段、以及其专门为研究人员录制的 3000 多个句子，由当地一家 DeepFake 技术公司创建了虚拟形象 AI Yoon，并在网络上迅速走红。实际上 AI Yoon 表达的内容是由竞选团队撰写的，而不是候选人本人。

  二是伪造视频，尤其是伪造领导人视频引起国际争端，扰乱选举秩序，或引起突发舆情事件，如伪造尼克松宣布第一次登月失败，伪造乌克兰总统泽连斯基宣布“投降”的信息，这些行为导致新闻媒体行业的社会信任衰退。

  三是伪造新闻，主要通过虚假新闻自动生成牟取非法利益，使用ChatGPT生成热点新闻，赚取流量，截至2023年6月30日全球生成伪造新闻网站已达277个，严重扰乱社会秩序。

  四是换脸变声，用于诈骗。如由于AI语音模仿了企业高管的声音，一家香港国际企业因此被骗3500万美元。

  五是生成不雅图片，特别是针对公众人物。如影视明星的色情视频制作，造成不良社会影响。因此，迫切需要发展互联网虚假信息的伪造检测技术。

  其次，AI大模型面临严重可信问题。这些问题包括：（1）“一本正经胡说八道”的事实性错误；（2）以西方价值观叙事，输出政治偏见和错误言论；（3）易被诱导，输出错误知识和有害内容；（4）数据安全问题加重，大模型成为重要敏感数据的诱捕器，ChatGPT将用户输入纳入训练数据库，用于改善ChatGPT，美方能够利用大模型获得公开渠道覆盖不到的中文语料，掌握我们自己都可能不掌握的“中国知识”。因此，迫切需要发展大模型安全监管技术与自己的可信大模型。

  除了技术手段外，人工智能安全保障需要相关立法工作。2021年科技部发布《新一代人工智能伦理规范》，2022年8月，全国信息安全标准化技术委员会发布《信息安全技术机器学习算法安全评估规范》，2022-2023年，中央网信办先后发布《互联网信息服务算法推荐管理规定》《互联网信息服务深度合成管理规定》《生成式人工智能服务管理办法》等。欧美国家也先后出台法规，2018年5月25日，欧盟出台《通用数据保护条例》，2022年10月4日，美国发布《人工智能权利法案蓝图》，2024年3月13日，欧洲议会通过了欧盟《人工智能法案》。

  我国应加快推进《人工智能法》出台，构建人工智能治理体系，确保人工智能的发展和应用遵循人类共同价值观，促进人机和谐友好；创造有利于人工智能技术研究、开发、应用的政策环境；建立合理披露机制和审计评估机制，理解人工智能机制原理和决策过程；明确人工智能系统的安全责任和问责机制，可追溯责任主体并补救；推动形成公平合理、开放包容的国际人工智能治理规则。

  4 中国智能计算发展困境

  人工智能技术与智能计算产业处于中美科技竞争的焦点，我国在过去几年虽然取得了很大的成绩，但依然面临诸多发展困境，特别是由美国的科技打压政策带来的困难。

  困境一为美国在AI核心能力上长期处于领先地位，中国处于跟踪模式。中国在AI高端人才数量、AI基础算法创新、AI底座大模型能力（大语言模型、文生图模型、文生视频模型）、底座大模型训练数据、底座大模型训练算力等，都与美国存在一定的差距，并且这种差距还将持续很长一段时间。

  困境二为高端算力产品禁售，高端芯片工艺长期被卡。A100，H100，B200等高端智算芯片对华禁售。华为、龙芯、寒武纪、曙光、海光等企业都进入实体清单，它们芯片制造的先进工艺④受限，国内可满足规模量产的工艺节点落后国际先进水平2-3代，核心算力芯片的性能落后国际先进水平2-3代。

  困境三为国内智能计算生态孱弱，AI开发框架渗透率不足。英伟达CUDA⑤(Compute Unified Device Architecture, 通用计算设备架构)生态完备，已形成了事实上的垄断。国内生态孱弱，具体表现在：一是研发人员不足，英伟达CUDA生态有近2万人开发，是国内所有智能芯片公司人员总和的20倍；二是开发工具不足，CUDA有550个SDK(Software Development Kit, 软件开发工具包)，是国内相关企业的上百倍；三是资金投入不足，英伟达每年投入50亿美元，是国内相关公司的几十倍；四是AI开发框架TensorFlow占据工业类市场，PyTorch占据研究类市场，百度飞桨等国产AI开发框架的开发人员只有国外框架的1/10。更为严重的是国内企业之间山头林立，无法形成合力，从智能应用、开发框架、系统软件、智能芯片，虽然每层都有相关产品，但各层之间没有深度适配，无法形成一个有竞争力的技术体系。

  困境四为AI应用于行业时成本、门槛居高不下。当前我国AI应用主要集中在互联网行业和一些国防领域。AI技术推广应用于各行各业时，特别是从互联网行业迁移到非互联网行业，需要进行大量的定制工作，迁移难度大，单次使用成本高。最后，我国在AI领域的人才数量与实际需求相比也明显不足。

  5 中国如何发展智能计算的道路选择

  人工智能发展的道路选择对我国至关重要，关系到发展的可持续性与最终的国际竞争格局。当前人工智能的使用成本十分高昂，微软Copilot套件要支付每月10美元的使用费用，ChatGPT每天消耗50万千瓦时的电力，英伟达B200芯片价格高达3万美元以上。总体来说，我国应发展用得起、安全可信的人工智能技术，消除我国信息贫困人口、并造福“一带一路”国家；低门槛地赋能各行各业，让我国的优势产业保持竞争力，让相对落后的产业能够大幅地缩小差距。

  选择一：统一技术体系走闭源封闭，还是开源开放的道路？

  支撑智能计算产业的是一个相互紧耦合的技术体系，即由一系列技术标准和知识产权将材料、器件、工艺、芯片、整机、系统软件、应用软件等密切联系在一起的技术整体。我国发展智能计算技术体系存在三条道路：

  一是追赶兼容美国主导的A体系。我国大多数互联网企业走的是GPGPU/CUDA兼容道路，很多芯片领域的创业企业在生态构建上也是尽量与CUDA兼容，这条道路较为现实。由于在算力方面美国对我国工艺和芯片带宽的限制，在算法方面国内生态林立很难形成统一，生态成熟度严重受限，在数据方面中文高质量数据匮乏，这些因素会使得追赶者与领先者的差距很难缩小，一些时候还会进一步拉大。　　

  二是构建专用封闭的B体系。在军事、气象、司法等专用领域构建企业封闭生态，基于国产成熟工艺生产芯片，相对于底座大模型更加关注特定领域垂直类大模型，训练大模型更多采用领域专有高质量数据等。这条道路易于形成完整可控的技术体系与生态，我国一些大型骨干企业走的是这条道路，它的缺点是封闭，无法凝聚国内大多数力量，也很难实现全球化。　　

  三是全球共建开源开放的C体系。用开源打破生态垄断，降低企业拥有核心技术的门槛，让每个企业都能低成本地做自己的芯片，形成智能芯片的汪洋大海，满足无处不在的智能需求。用开放形成统一的技术体系，我国企业与全球化力量联合起来共建基于国际标准的统一智能计算软件栈。形成企业竞争前共享机制，共享高质量数据库，共享开源通用底座大模型。对于全球开源生态，我国企业在互联网时代收益良多，我国更多的是使用者，是参与者，在智能时代我国企业在RISC-V⑥+AI开源技术体系上应更多地成为主力贡献者，成为全球化开放共享的主导力量。

  选择二：拼算法模型，还是拼新型基础设施？　　

  人工智能技术要赋能各行各业，具有典型的长尾效应⑦。我国80%的中小微企业，需要的是低门槛、低价格的智能服务。因此，我国智能计算产业必须建立在新的数据空间基础设施之上，其中关键是我国应率先实现智能要素即数据、算力、算法的全面基础设施化。这项工作可比肩二十世纪初美国信息高速公路计划（即信息基础设施建设）对互联网产业的历史作用。　　

  信息社会最核心的生产力是网络空间(Cyberspace)。网络空间的演进过程是：从机器一元连接构成的计算空间，演进到人机信息二元连接构成的信息空间，再演进到人机物数据三元连接构成的数据空间。从数据空间看，人工智能的本质是数据的百炼成钢，大模型就是对互联网全量数据进行深度加工后的产物。在数字化时代，在互联网上传输的是信息流，是算力对数据进行粗加工后的结构化抽象；在智能时代，在互联网上传输的是智能流，是算力对数据进行深度加工与精炼后的模型化抽象。智能计算的一个核心特征就是用数值计算、数据分析、人工智能等算法，在算力池中加工海量数据件，得到智能模型，再嵌入到信息世界、物理世界的各个过程中。　　

  我国政府已经前瞻性地提前布局了新型基础设施，在世界各国竞争中抢占了先机。

  首先，数据已成为国家战略信息资源。数据具有资源要素与价值加工两重属性，数据的资源要素属性包括生产、获取、传输、汇聚、流通、交易、权属、资产、安全等各个环节，我国应继续加大力度建设国家数据枢纽与数据流通基础设施。　　

  其次，AI大模型就是数据空间的一类算法基础设施。以通用大模型为基座，构建大模型研发与应用的基础设施，支撑广大企业研发领域专用大模型，服务于机器人、无人驾驶、可穿戴设备、智能家居、智能安防等行业，覆盖长尾应用。　　

  最后，全国一体化算力网建设在推动算力的基础设施化上发挥了先导作用。算力基础设施化的中国方案，应在大幅度降低算力使用成本和使用门槛的同时，为最广范围覆盖人群提供高通量、高品质的智能服务。算力基础设施的中国方案需要具备“两低一高”，即在供给侧，大幅度降低算力器件、算力设备、网络连接、数据获取、算法模型调用、电力消耗、运营维护、开发部署的总成本，让广大中小企业都消费得起高品质的算力服务，有积极性开发算力网应用；在消费侧，大幅度降低广大用户的算力使用门槛，面向大众的公共服务必须做到易获取、易使用，像水电一样即开即用，像编写网页一样轻松定制算力服务，开发算力网应用。在服务效率侧，中国的算力服务要实现低熵高通量，其中高通量是指在实现高并发⑧度服务的同时，端到端服务的响应时间可满足率高；低熵是指在高并发负载中出现资源无序竞争的情况下，保障系统通量不急剧下降。保障“算得多”对中国尤其重要。　　

  选择三：AI+着重赋能虚拟经济，还是发力实体经济？　　

  “AI+”的成效是人工智能价值的试金石。次贷危机后，美国制造业增加值占GDP的比重从1950年的28%降低为2021年的11%，美国制造业在全行业就业人数占比从1979年的35%降低为2022年的8%，可见美国更倾向于回报率更高的虚拟经济，轻视投资成本高且经济回报率低的实体经济。中国倾向于实体经济与虚拟经济同步发展，更加重视发展装备制造、新能源汽车、光伏发电、锂电池、高铁、5G等实体经济。　　

  相应地美国AI主要应用于虚拟经济和IT基础工具，AI技术也是“脱实向虚”，自2007年以来硅谷不断炒作虚拟现实（Virtual Reality，VR）、元宇宙、区块链、Web3.0、深度学习、AI大模型等，是这个趋势的反映。　　

  我国的优势在实体经济，制造业全球产业门类最齐全，体系最完整，特点是场景多、私有数据多。我国应精选若干行业加大投入，形成可低门槛全行业推广的范式，如选择装备制造业作为延续优势代表性行业，选择医药业作为快速缩短差距的代表性行业。赋能实体经济的技术难点是AI算法与物理机理的融合。

  人工智能技术成功的关键是能否让一个行业或一个产品的成本大幅下降，从而将用户数与产业规模扩大10倍，产生类似于蒸汽机对于纺织业，智能手机对于互联网业的变革效果。

  我国应走出适合自己的人工智能赋能实体经济的高质量发展道路。

  注释：　　

  ①模式识别是指用计算的方法根据样本的特征将样本划分到一定的类别中去，是通过计算机用数学方法来研究模式的自动处理和判读，把环境与客体统称为“模式”。以图像处理与计算机视觉、语音语言信息处理、脑网络组、类脑智能等为主要研究方向。　　

  ②Token可翻译为词元，指自然语言处理过程中用来表示单词或短语的符号。token可以是单个字符,也可以是多个字符组成的序列。　　

  ③通用人工智能是指拥有与人类相当甚至超过人类智能的人工智能类型。通用人工智能不仅能像人类一样进行感知、理解、学习和推理等基础思维能力，还能在不同领域灵活应用、快速学习和创造性思考。通用人工智能的研究目标是寻求统一的理论框架来解释各种智能现象。　　

  ④芯片制造工艺指制造CPU或GPU的制程，即晶体管门电路的尺寸，单位为纳米，目前国际上实现量产的最先进工艺以台积电的3nm为代表。更先进的制造工艺可以使CPU与GPU内部集成更多的晶体管，使处理器具有更多的功能以及更高的性能，面积更小，成本更低等。　　

  ⑤CUDA是英伟达公司设计研发一种并行计算平台和编程模型，包含了CUDA指令集架构以及GPU内部的并行计算引擎。开发人员可以使用C语言来为CUDA架构编写程序，所编写出的程序可以在支持CUDA的处理器上以超高性能运行。　　

  ⑥RISC-V（发音为“risk-five”）是一个由美国加州大学伯克利分校发起的开放通用指令集架构，相比于其他付费指令集，RISC-V允许任何人免费地使用RISC-V指令集设计、制造和销售芯片和软件。　　

  ⑦长尾效应是指那些原来不受到重视的销量小但种类多的产品或服务由于总量巨大，累积起来的总收益超过主流产品的现象。在互联网领域，长尾效应尤为显著。　　

  ⑧高并发通常指通过设计保证系统能够同时并行处理很多请求。

【责任编辑：孟瑾】

About 高大伟 David Cowhig

After retirement translated, with wife Jessie, Liao Yiwu's 2019 "Bullets and Opium", and have been studying things 格物致知. Worked 25 years as a US State Department Foreign Service Officer including ten years at US Embassy Beijing and US Consulate General Chengdu and four years as a China Analyst in the Bureau of Intelligence and Research. Before State I translated Japanese and Chinese scientific and technical books and articles into English freelance for six years. Before that I taught English at Tunghai University in Taiwan for three years. And before that I worked two summers on Norwegian farms, milking cows and feeding chickens.
This entry was posted in Economy 经济, Law 法律, Science, Technology and Academic 科技学术 and tagged , , , , , , , , , , , , , , , , , , . Bookmark the permalink.

Leave a comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.