Beijing has put forward specific measures for the core elements of industrial development such as computing power, data, algorithms, applications, and supervision, as well as the difficulties and blockages in key links, and the five dimensions have coordinated efforts.
·To a certain extent, how Beijing formulates and implements policies in the development of artificial intelligence has a national demonstration effect. “China is speeding up the cultivation of indigenous AI pioneers by strengthening state support,” and “it is not uncommon for the capital to take the lead in formulating policies for emerging industries.”
Recently, Beijing, Shanghai, Shenzhen, Chengdu and other places have successively issued a series of policy documents to encourage the development of artificial intelligence. Among them, Beijing, which has the highest policy density and attracts the most attention, announced three related documents in a row within 12 days: May 19 On May 30, issued the "Beijing General Artificial Intelligence Industry Innovation Partnership Plan" (hereinafter referred to as the "Partnership Program"); on May 30, issued "Several Measures for Beijing to Promote the Innovation and Development of General Artificial Intelligence" (hereinafter referred to as "Several Measures") ; On May 30, the "Implementation Plan for Beijing to Accelerate the Construction of a Globally Influential Artificial Intelligence Innovation Source (2023-2025)" (hereinafter referred to as the "Implementation Plan") was issued.
A careful review of these documents reveals that Beijing has proposed specific measures for the core elements of industrial development such as computing power, data, algorithms, applications, and supervision, as well as difficulties and blockages in key links, and the five dimensions have coordinated efforts.
"The "Implementation Plan" is a document at the strategic level. The "Partnership Plan" is a special document on innovation partners and value alliances, which belongs to the strategic level. The "Several Measures" propose a number of operational measures. These documents , there are strategies, strategies, and operational implementations, which complement each other and capture the trends and priorities of the times." Gong Yeming, Dean of the School of Artificial Intelligence Management (AIM) and Director of the Global Business Intelligence Center (BIC) of Emlyon Business School, told The Paper ( analyzed.
According to IDC's "2022-2023 China Artificial Intelligence Computing Power Development Evaluation Report", Beijing ranks first in the list of China's artificial intelligence cities in 2022. As of November 2022, Beijing leads the country in 17 AI-related fields. According to the "China Artificial Intelligence Large-scale Model Map Research Report" released by the Institute of Scientific and Technical Information of China on May 28 this year, at least 79 large-scale models with a scale of more than 1 billion parameters have been released in China, and Beijing accounts for 38, ranking first. .
To some extent, how Beijing formulates and implements policies in the development of artificial intelligence has a national demonstration effect. TechCrunch, an American technology media, recently commented on Beijing’s measures, saying that “China is accelerating the cultivation of local artificial intelligence pioneers through strengthening state support,” and “it is not uncommon for the capital to take the lead in formulating policies for emerging industries.”
Changes in the ranking of TOP10 artificial intelligence cities in China in the past five years. Source: IDC
Coordinate the supply of computing resources and strengthen domestic chips
Computing power is the basic element for training large language models (LLM), and the development of large language models also brings huge challenges to computing power. The recent surge in the market value of Nvidia, the "leader" of AI chips, can reflect the market's expectations for the importance of computing power after entering the ChatGPT era. "Currently, computing power is an element that needs to be dealt with and improved urgently in the innovation of general artificial intelligence." Gong Yeming said.
In terms of computing power, the first article of the "Several Measures" proposes to "enhance the ability to coordinate and supply computing power resources". On the one hand, organize commercial computing power to meet urgent needs: implement the computing power partnership plan, strengthen cooperation with cloud vendors, and provide diversified high-quality inclusive computing power; on the other hand, promote the construction of new computing power infrastructure: accelerate the promotion of Haidian District, Chaoyang District construction of Beijing Artificial Intelligence Public Computing Power Center, Beijing Digital Economy Computing Power Center and other projects to form large-scale advanced computing power supply capabilities; in addition, build a unified multi-cloud computing power scheduling platform to achieve unified management and unified operation, and improve the integrated dispatching capability of computing power in the area around Beijing.
Among them, the "supply of large-scale advanced computing power" is worth noting. Lu Yanxia, research director of IDC China, told The Paper that the development of next-generation AI requires the support of large-scale advanced computing power, especially the pre-training of large models and the research and development of generative AI. .
At present, more than 30 cities across the country are building or proposing to build smart computing centers. According to the forecast of Zheshang Securities, during the "14th Five-Year Plan" period, the investment in the intelligent computing center can drive the growth of the core industry of artificial intelligence by about 2.9-3.4 times. In the next few years, the compound annual growth rate of China's intelligent computing power scale will exceed 50%. Computing power replaces basic computing power as the main component of computing power structure, and intelligent computing power becomes the driving force for growth.
However, Lu Yanxia reminded that what needs to be paid attention to is that after building these computing power, one must consider the future utilization rate and not become a sunk cost. In addition, in the next few years, the most important thing to pay attention to is how to improve computing energy efficiency. With the development of data centers to today's scale and volume, energy consumption will become an important factor restricting development in the future.
In terms of hardware carrying computing power—chips, Gong Yeming pointed out that high-end artificial intelligence chips and related equipment are expensive and difficult to obtain, which is a major shortcoming in the development of general artificial intelligence.
For chips, the "Implementation Plan" emphasizes the word "domestic". The work goal proposed in the document is: "The market share of basic software and hardware products such as domestic artificial intelligence chips and deep learning frameworks has increased significantly, and computing power chips have basically achieved independent control. The proportion of domestic hardware has increased significantly, and it is fully compatible with domestic deep learning frameworks. In terms of main tasks, the document proposes: Actively guide large-scale model R&D enterprises to apply domestic artificial intelligence chips, accelerate the improvement of the localization rate of artificial intelligence computing power supply; strengthen the deployment and application of domestic chips, and promote independent controllable software and hardware computing power ecology Construction; establish a full-stack localized artificial intelligence innovation consortium, develop a full-stack localized generative large model, and gradually form an independent and controllable artificial intelligence technology system and industrial ecology.
Regarding the specific research and development direction of the chip, the "Implementation Plan" pointed out: to meet the needs of artificial intelligence cloud distributed training, carry out the development of general-purpose high-computing training chips; for the low power consumption requirements of edge-end application scenarios, develop multi-modal intelligent sensor chips , Independent intelligent decision-making execution chips, high-energy-efficiency edge-end heterogeneous smart chips; for innovative chip architectures, explore innovative architecture routes such as reconfigurable, storage-computing integration, brain-like computing, and Chiplets.
Zheshang Securities interpreted that it is expected that the construction of computing power centers in Beijing is expected to speed up in the future, domestic chips will receive priority support, and computing power hardware suppliers of the Chinese Department of Science and Technology and Huawei are expected to benefit deeply.
Construction of high-quality pre-training Chinese corpus
"The cost of computing power is indeed one of the main challenges in the current promotion of large-scale models and generative AI. But computing power is not the only challenge. Data resources and high-end talents are challenges. A series of supporting measures such as opening up data and cultivating industrial talents are needed." Lu Yanxia said.
In terms of data, the core idea of the Beijing series of documents is to build a safe and compliant open basic training data set and build a high-quality pre-trained Chinese corpus.
Data is an important bottleneck in the development of large-scale Chinese language models. Wang Hao, head of the Text Generation Algorithm Team of the Cognitive Computing and Natural Language Center of the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute, once told The Paper, "Compared with English data, the degree of open source of Chinese data is relatively low, resulting in a relatively large scale of Chinese data sets. In addition, English, as the mainstream language of scientific research, has been widely used in academia and industry, and has accumulated a large amount of high-quality corpus data, which provides a great advantage for the research of English natural language processing."
In terms of building high-quality data sets, the "Several Measures" proposed three specific implementation directions. The first is to organize relevant institutions to integrate and clean Chinese pre-training data to form a safe and compliant open basic training data set. Explore and promote the authorized operation of public data areas, and promote the docking of public data and market-oriented data platforms; second, plan and build data training bases, improve the scale and quality of data annotation databases, and explore commercial scenario cooperation based on data contributions and model applications; third, construction A crowdsourcing service platform for fine-grained annotation of datasets, encouraging professionals to participate in annotation, and researching the incentive mechanism of the platform.
The "Implementation Plan" emphasizes the need to strengthen the open sharing of public data. "Focus on areas such as city brain, smart government affairs, and smart people's livelihood services, dynamically update the public data open plan, and improve the construction of various public data areas such as finance, transportation, and space."
Zhou Xuejing, a senior researcher of Shanghai Cyber Security Industry Innovation Research, told The Paper that the current sources of training data for large models can be divided into two categories, one is network data, and the other is synthetic data.
Sources of network data include: collecting personal information in the physical world to form network data, directly collecting data on the Internet, and data transactions. Generally speaking, network data is expensive to collect, clean, and process in terms of time and money.
The processing of synthetic data is more efficient, cheap and reusable. Taking face data as an example, if the face data provided by a natural person is set to 1, then the basic face data (features or expressions) can be adjusted through functions such as synthesis and editing, and 10 or 100 people can be realized. Face data greatly reduces the cost and difficulty of obtaining training data. Synthetic data also requires personal information protection. According to the "Internet Information Service Deep Synthesis Management Regulations", before using the biometric information editing function, the edited individual should be informed according to law and their separate consent obtained.
Web data versus synthetic data.
Developing a basic theoretical system for general artificial intelligence
Thanks to the strong model generalization ability, low dependence on long-tail data and the improvement of the efficiency of downstream models, the large model is considered to have the embryonic form of "general intelligence".
The "Implementation Plan" shows that Beijing plans to develop a basic theoretical framework system for the new generation of artificial intelligence, focusing on the mathematical mechanism of artificial intelligence, big data intelligence, multimodal intelligence, decision-making intelligence, brain-like intelligence, scientific intelligence, embodied intelligence, etc. Develop a research layout.
"Several Measures" clearly proposes to explore new paths of general artificial intelligence such as general intelligence, embodied intelligence and brain-like intelligence. Support value and causality-driven general agent research, create a unified theoretical framework system, rating standards, and testing platform, develop operating systems and programming languages, and promote the application of the underlying technical architecture of general agents. Promote the research and application of embodied intelligence systems, and make breakthroughs in the perception, cognition, and decision-making technologies of robots under complex conditions such as open environments, generalized scenarios, and continuous tasks. Support the exploration of brain-inspired intelligence, study core technologies such as brain neuron connection patterns, coding mechanisms, and information processing, and inspire new artificial neural network model modeling and training methods.
Among them, the "embodied intelligence" mentioned in the document refers to intelligent bodies that have a body and support physical interaction, such as home service robots and unmanned vehicles. Recently, this concept has been popularized by Nvidia founder and CEO Huang Renxun. In mid-May, he said in his speech at the 2023 ITF World Congress, "The next wave of artificial intelligence will be a new type of artificial intelligence called embodied artificial intelligence. Artificial intelligence, or intelligent systems that understand, reason, and interact with the physical world."
It is worth noting that the "Several Measures" also proposed to encourage third-party non-profit organizations to build multi-modal and multi-dimensional basic model evaluation benchmarks and evaluation methods; to study artificial intelligence-assisted model evaluation algorithms, and to develop models that include versatility, efficiency, A multi-dimensional basic model evaluation tool set including intelligence and robustness; build an open service platform for large model evaluation, establish a fair and efficient adaptive evaluation system, and realize automatic adaptive evaluation of large models according to different goals and tasks.
Since the explosion of ChatGPT, domestic large-scale models have experienced a "hurricane" spring. Zhao Zhiyun, director of the China Institute of Scientific and Technological Information, said on May 28 that according to incomplete statistics, 79 large-scale models with a scale of more than 1 billion parameters have been released in China so far, and 14 provinces and cities/regions are developing large-scale models. Facing the situation of "a hundred flowers blooming", how to evaluate large models has been put on the policy agenda. The Paper learned from many people familiar with the matter that many key artificial intelligence laboratories in Beijing and Shanghai are currently completing the evaluation work in a concentrated manner.
Just on June 3, the National Key Laboratory of Cognitive Intelligence, the Chinese Academy of Sciences Artificial Intelligence Industry-University-Research Innovation Alliance and the Yangtze River Delta Artificial Intelligence Industry Chain Alliance jointly released the "General Cognitive Intelligence Large Model Evaluation System", aiming to form a set of coverage The multi-task objective evaluation system of large-scale model capabilities guides the healthy development of China's cognitive large-scale model technology and industry.
Focus on six scenes
With the support of large-scale models, AIGC (Artificial Intelligence Generated Content) applications such as Vincent diagrams and virtual digital humans will quickly enter the commercialization stage.
"Several Measures" proposes to promote the application of general artificial intelligence technology innovation scenarios, focusing on six major scenarios of government services, medical care, scientific research, finance, autonomous driving, and urban governance.
The "Implementation Plan" emphasizes that relying on Beijing's superior scene resources, it is necessary to accelerate the convergence of capital, technology, data, computing power, talents and other elements, create a batch of benchmarking demonstration application scenarios that can be replicated and promoted, and promote the artificial intelligence innovation chain. The industrial chain, the capital chain, and the talent chain are deeply integrated. In addition, support artificial intelligence to empower the construction of smart cities, support Haidian District to build City Brain 2.0, and promote the smooth implementation of projects such as Beijing High-level Autonomous Driving Demonstration Zone 3.0.
The Partner Program mentions some more specific scenarios. "Focusing on scenarios such as Window of the Capital's intelligent question-and-answer and online guides, relying on privatized deployment of computing power cluster resources, we will gradually carry out work such as training, fine-tuning, pruning, and distillation of proprietary models in the government service industry, empowering the '12345 Immediately implement the 'auxiliary scene." "Focus on the city's virtual digital human, digital medical care, e-commerce retail and other innovative and active data advantage fields, accelerate the commercialization of large-scale models, and accelerate text creation, human-computer interaction, education, audio-visual, etc. Scenario landing application."
The development path of China's artificial intelligence application scenarios. Source: IDC
"I think it is still time to develop dedicated artificial intelligence. Even if it is ChatGPT or GPT (generative pre-training Transformer model), if it is to be truly industrialized, it still needs to be polished for the scene." Lu Yanxia said.
IDC believes that in the next five years, as human-computer interaction, machine learning, computer vision, and speech recognition technologies reach a more mature stage, the development trend of artificial intelligence applications includes creative work that uses knowledge as the main production tool (such as writing , video, image and audio creation, software development, IP incubation, etc.) will achieve a greater degree of intelligence; industry enterprises will expand the integration of digital twin and artificial intelligence technology, promote the development of energy and power, manufacturing, construction and other industries, build Virtual factories, digital twin power grids, and digital twin cities strengthen the connection between the digital and the real world, optimize processes, realize global management, and make intelligent decisions.
Create an inclusive and prudent regulatory environment
In terms of supervision, both documents refer to "inclusive prudence".
Zhou Xuejing told The Paper that "tolerance and prudence" at the regulatory level is divided into two dimensions: "tolerance" and "prudence". Tolerance is to encourage the innovation and development of domestic general artificial intelligence at the current stage, and to give enterprises the opportunity to try and make mistakes by establishing pilot projects (demonstration areas, test areas), thereby accelerating the formation of a number of benchmark demonstration cases, and accelerating the development of the entire industry . Prudence, on the other hand, starts from the perspective of security and compliance, and considers the impact of network security, data security, and other new risks that arise in the development process on the country, society, and individuals. Therefore, timely regulatory intervention in the market is required. .
"Essentially, tolerance and prudence is a flexible, classified and hierarchical regulatory strategy, especially in the early stage of the development of emerging technologies such as large models and AIGC and new industries, which can better reduce the trial and error costs and compliance costs of enterprises, and encourage my country's artificial intelligence innovation and development." Zhou Xuejing said.
How to create an inclusive and prudent regulatory environment? The "Several Measures" mainly includes four points:
First, continue to promote the innovation of regulatory policies and regulatory processes, encourage innovation subjects to adopt safe and reliable software, tools, computing and data resources, and carry out independent innovation, promotion and application, and international cooperation of basic technologies such as artificial intelligence algorithms and frameworks; , establish a normalized service and guidance mechanism, carry out normalized contact services for artificial intelligence-related Internet information services with public opinion attributes or social mobilization capabilities, guide innovation subjects to introduce technical tools for security testing, declare security assessments in accordance with regulations, and perform algorithm filing Third, strengthen network service security protection and personal data protection; fourth, continue to strengthen scientific and technological ethics governance.
In fact, the rapid development of generative artificial intelligence has triggered regulatory actions in many countries and regions around the world. Under the impetus of technology giants such as OpenAI and Google, the US government and Congress have launched discussions. The European Union, which is ahead in terms of regulation, has almost completed the legislative process for its "Artificial Intelligence Act".
The EU white paper "The Development of Artificial Intelligence for Excellence and Trust" provides an idea worthy of reference, that is, to carry out a "risk level-based" regulatory path. In principle, mandatory regulations are only applicable to high-risk artificial intelligence activities to ensure Regulatory intervention is applicable and targeted. In addition, multi-party participation in regulatory governance is implemented, and certification bodies designated by member states conduct independent review and evaluation of artificial intelligence systems.
At the implementation level, Zhou Xuejing analyzed that from the perspective of the current artificial intelligence supervision departments, my country presents the current situation of multi-management supervision. Relevant departments include the State Administration for Market Regulation, the National Internet Information Office, the Ministry of Industry and Information Technology, and the Ministry of Science and Technology. At the level of laws and regulations, on the one hand, it is to regulate the obligations and responsibilities of network operators in the use of artificial intelligence technology through specialized and comprehensive legislation; specification.
Talent Highland and Government Investment Fund
In addition to the five dimensions of computing power, data, algorithms, applications, and supervision, the Beijing series of documents also proposed specific measures in terms of talent and innovation.
The "Implementation Plan" has a clear goal: to build a group of artificial intelligence scientific research institutions with world-class influence, to introduce and cultivate world-class innovative talent teams, and to make new breakthroughs in international talent introduction. The number of high-level scholars exceeds 10,000, and the domestic proportion remains at the leading position.
The main tasks include promoting the construction of a highland of talents in the field of artificial intelligence. "Explore and implement the policy of supporting overseas talents to come to Beijing and focus on introducing and cultivating a group of top talents and young talents with world influence." Level scientists, industrial and engineering talents."
As one of the most important research institutions in the field of artificial intelligence in China, Beijing Zhiyuan Artificial Intelligence Research Institute has recently received unprecedented attention. Microsoft President Brad Smith (Brad Smith) said in April when talking about Microsoft and ChatGPT's competitors, "We see three institutions that are absolutely at the forefront, one is OpenAI and Microsoft, the second is Google, The third is Beijing Zhiyuan Artificial Intelligence Research Institute.”
Beijing Zhiyuan Artificial Intelligence Research Institute was established in December 2018. It positions itself as a "new research and development institution" and proposes a set of "Zhiyuan Model", which is to establish a scientific research management mechanism that combines free exploration and goal orientation, and selects from the perspective of small peers. Support the free exploration of Zhiyuan scholars, and promote the implementation of major scientific research tasks such as the "Enlightenment" model with the scientific research organization model of "concentrating efforts to do big things". In 2021, Zhiyuan Research Institute released the world's largest and most powerful intelligent model "Enlightenment 2.0", with a parameter volume of 1.75 trillion (ten times that of GPT-3.5). On May 28 this year, Beijing Zhiyuan Artificial Intelligence Research Institute launched the general segmentation model SegGPT (Segment Everything In Context) at the parallel forum of the 2023 Zhongguancun Forum.
In addition to world-class research institutions, the innovation environment of enterprises is also particularly important. The "Implementation Plan" puts forward goals: the research investment of leading artificial intelligence enterprises will continue to increase, the number of start-up enterprises will continue to increase, the total number of enterprises will maintain the leading position in China, and 5-10 new unicorn enterprises will be cultivated.
The main tasks include giving play to the guiding role of government investment funds and supporting long-term capital and patient capital in early-stage hard technology investments in artificial intelligence chips, frameworks, and core algorithms. Continue to do a good job in the cultivation of artificial intelligence enterprises listed on the market.
The "Implementation Plan" pointed out that it is necessary to strengthen the gradient cultivation of artificial intelligence enterprises. "Intensify the cultivation of innovative small and medium-sized enterprises, and bring companies that have the potential to become unicorns into the cultivation system in advance." "Promote a group of internationally renowned research institutions, multinational companies, and domestic leading companies to build innovative business entities in Beijing." " Implement the two-level enterprise service package and service stewardship mechanism in urban areas, and for innovative companies with potential in the field of artificial intelligence, the standards for inclusion in service packages can be appropriately relaxed to increase service coverage.”
Similar to Beijing's policy, the recently released "Shenzhen City Accelerates the Promotion of High-quality Development of Artificial Intelligence and High-Level Application Action Plan (2023-2024)" proposes that Shenzhen will play the role of government investment guidance funds, coordinate and integrate fund resources, and form a scale of 1,000 100 million yuan artificial intelligence fund group. Almost at the same time, "Several Policies and Measures for Strengthening Support for the Development of Private Investment in Shanghai" was released, which will use the "fund investment" model to gather efforts to attract "three leading industries", artificial intelligence is one of the "three leading industries" one.
View Original
The content is for reference only, not a solicitation or offer. No investment, tax, or legal advice provided. See Disclaimer for more risks disclosure.
Beijing has issued 3 policy documents in 12 days. What is the demonstration effect on accelerating the development of AI in the whole country?
Source: The Paper
Reporter Shao Wen
Beijing has put forward specific measures for the core elements of industrial development such as computing power, data, algorithms, applications, and supervision, as well as the difficulties and blockages in key links, and the five dimensions have coordinated efforts.
·To a certain extent, how Beijing formulates and implements policies in the development of artificial intelligence has a national demonstration effect. “China is speeding up the cultivation of indigenous AI pioneers by strengthening state support,” and “it is not uncommon for the capital to take the lead in formulating policies for emerging industries.”
Recently, Beijing, Shanghai, Shenzhen, Chengdu and other places have successively issued a series of policy documents to encourage the development of artificial intelligence. Among them, Beijing, which has the highest policy density and attracts the most attention, announced three related documents in a row within 12 days: May 19 On May 30, issued the "Beijing General Artificial Intelligence Industry Innovation Partnership Plan" (hereinafter referred to as the "Partnership Program"); on May 30, issued "Several Measures for Beijing to Promote the Innovation and Development of General Artificial Intelligence" (hereinafter referred to as "Several Measures") ; On May 30, the "Implementation Plan for Beijing to Accelerate the Construction of a Globally Influential Artificial Intelligence Innovation Source (2023-2025)" (hereinafter referred to as the "Implementation Plan") was issued.
A careful review of these documents reveals that Beijing has proposed specific measures for the core elements of industrial development such as computing power, data, algorithms, applications, and supervision, as well as difficulties and blockages in key links, and the five dimensions have coordinated efforts.
"The "Implementation Plan" is a document at the strategic level. The "Partnership Plan" is a special document on innovation partners and value alliances, which belongs to the strategic level. The "Several Measures" propose a number of operational measures. These documents , there are strategies, strategies, and operational implementations, which complement each other and capture the trends and priorities of the times." Gong Yeming, Dean of the School of Artificial Intelligence Management (AIM) and Director of the Global Business Intelligence Center (BIC) of Emlyon Business School, told The Paper ( analyzed.
According to IDC's "2022-2023 China Artificial Intelligence Computing Power Development Evaluation Report", Beijing ranks first in the list of China's artificial intelligence cities in 2022. As of November 2022, Beijing leads the country in 17 AI-related fields. According to the "China Artificial Intelligence Large-scale Model Map Research Report" released by the Institute of Scientific and Technical Information of China on May 28 this year, at least 79 large-scale models with a scale of more than 1 billion parameters have been released in China, and Beijing accounts for 38, ranking first. .
To some extent, how Beijing formulates and implements policies in the development of artificial intelligence has a national demonstration effect. TechCrunch, an American technology media, recently commented on Beijing’s measures, saying that “China is accelerating the cultivation of local artificial intelligence pioneers through strengthening state support,” and “it is not uncommon for the capital to take the lead in formulating policies for emerging industries.”
Coordinate the supply of computing resources and strengthen domestic chips
Computing power is the basic element for training large language models (LLM), and the development of large language models also brings huge challenges to computing power. The recent surge in the market value of Nvidia, the "leader" of AI chips, can reflect the market's expectations for the importance of computing power after entering the ChatGPT era. "Currently, computing power is an element that needs to be dealt with and improved urgently in the innovation of general artificial intelligence." Gong Yeming said.
In terms of computing power, the first article of the "Several Measures" proposes to "enhance the ability to coordinate and supply computing power resources". On the one hand, organize commercial computing power to meet urgent needs: implement the computing power partnership plan, strengthen cooperation with cloud vendors, and provide diversified high-quality inclusive computing power; on the other hand, promote the construction of new computing power infrastructure: accelerate the promotion of Haidian District, Chaoyang District construction of Beijing Artificial Intelligence Public Computing Power Center, Beijing Digital Economy Computing Power Center and other projects to form large-scale advanced computing power supply capabilities; in addition, build a unified multi-cloud computing power scheduling platform to achieve unified management and unified operation, and improve the integrated dispatching capability of computing power in the area around Beijing.
Among them, the "supply of large-scale advanced computing power" is worth noting. Lu Yanxia, research director of IDC China, told The Paper that the development of next-generation AI requires the support of large-scale advanced computing power, especially the pre-training of large models and the research and development of generative AI. .
At present, more than 30 cities across the country are building or proposing to build smart computing centers. According to the forecast of Zheshang Securities, during the "14th Five-Year Plan" period, the investment in the intelligent computing center can drive the growth of the core industry of artificial intelligence by about 2.9-3.4 times. In the next few years, the compound annual growth rate of China's intelligent computing power scale will exceed 50%. Computing power replaces basic computing power as the main component of computing power structure, and intelligent computing power becomes the driving force for growth.
However, Lu Yanxia reminded that what needs to be paid attention to is that after building these computing power, one must consider the future utilization rate and not become a sunk cost. In addition, in the next few years, the most important thing to pay attention to is how to improve computing energy efficiency. With the development of data centers to today's scale and volume, energy consumption will become an important factor restricting development in the future.
In terms of hardware carrying computing power—chips, Gong Yeming pointed out that high-end artificial intelligence chips and related equipment are expensive and difficult to obtain, which is a major shortcoming in the development of general artificial intelligence.
For chips, the "Implementation Plan" emphasizes the word "domestic". The work goal proposed in the document is: "The market share of basic software and hardware products such as domestic artificial intelligence chips and deep learning frameworks has increased significantly, and computing power chips have basically achieved independent control. The proportion of domestic hardware has increased significantly, and it is fully compatible with domestic deep learning frameworks. In terms of main tasks, the document proposes: Actively guide large-scale model R&D enterprises to apply domestic artificial intelligence chips, accelerate the improvement of the localization rate of artificial intelligence computing power supply; strengthen the deployment and application of domestic chips, and promote independent controllable software and hardware computing power ecology Construction; establish a full-stack localized artificial intelligence innovation consortium, develop a full-stack localized generative large model, and gradually form an independent and controllable artificial intelligence technology system and industrial ecology.
Regarding the specific research and development direction of the chip, the "Implementation Plan" pointed out: to meet the needs of artificial intelligence cloud distributed training, carry out the development of general-purpose high-computing training chips; for the low power consumption requirements of edge-end application scenarios, develop multi-modal intelligent sensor chips , Independent intelligent decision-making execution chips, high-energy-efficiency edge-end heterogeneous smart chips; for innovative chip architectures, explore innovative architecture routes such as reconfigurable, storage-computing integration, brain-like computing, and Chiplets.
Zheshang Securities interpreted that it is expected that the construction of computing power centers in Beijing is expected to speed up in the future, domestic chips will receive priority support, and computing power hardware suppliers of the Chinese Department of Science and Technology and Huawei are expected to benefit deeply.
Construction of high-quality pre-training Chinese corpus
"The cost of computing power is indeed one of the main challenges in the current promotion of large-scale models and generative AI. But computing power is not the only challenge. Data resources and high-end talents are challenges. A series of supporting measures such as opening up data and cultivating industrial talents are needed." Lu Yanxia said.
In terms of data, the core idea of the Beijing series of documents is to build a safe and compliant open basic training data set and build a high-quality pre-trained Chinese corpus.
Data is an important bottleneck in the development of large-scale Chinese language models. Wang Hao, head of the Text Generation Algorithm Team of the Cognitive Computing and Natural Language Center of the Guangdong-Hong Kong-Macao Greater Bay Area Digital Economy Research Institute, once told The Paper, "Compared with English data, the degree of open source of Chinese data is relatively low, resulting in a relatively large scale of Chinese data sets. In addition, English, as the mainstream language of scientific research, has been widely used in academia and industry, and has accumulated a large amount of high-quality corpus data, which provides a great advantage for the research of English natural language processing."
In terms of building high-quality data sets, the "Several Measures" proposed three specific implementation directions. The first is to organize relevant institutions to integrate and clean Chinese pre-training data to form a safe and compliant open basic training data set. Explore and promote the authorized operation of public data areas, and promote the docking of public data and market-oriented data platforms; second, plan and build data training bases, improve the scale and quality of data annotation databases, and explore commercial scenario cooperation based on data contributions and model applications; third, construction A crowdsourcing service platform for fine-grained annotation of datasets, encouraging professionals to participate in annotation, and researching the incentive mechanism of the platform.
The "Implementation Plan" emphasizes the need to strengthen the open sharing of public data. "Focus on areas such as city brain, smart government affairs, and smart people's livelihood services, dynamically update the public data open plan, and improve the construction of various public data areas such as finance, transportation, and space."
Zhou Xuejing, a senior researcher of Shanghai Cyber Security Industry Innovation Research, told The Paper that the current sources of training data for large models can be divided into two categories, one is network data, and the other is synthetic data.
Sources of network data include: collecting personal information in the physical world to form network data, directly collecting data on the Internet, and data transactions. Generally speaking, network data is expensive to collect, clean, and process in terms of time and money.
The processing of synthetic data is more efficient, cheap and reusable. Taking face data as an example, if the face data provided by a natural person is set to 1, then the basic face data (features or expressions) can be adjusted through functions such as synthesis and editing, and 10 or 100 people can be realized. Face data greatly reduces the cost and difficulty of obtaining training data. Synthetic data also requires personal information protection. According to the "Internet Information Service Deep Synthesis Management Regulations", before using the biometric information editing function, the edited individual should be informed according to law and their separate consent obtained.
Developing a basic theoretical system for general artificial intelligence
Thanks to the strong model generalization ability, low dependence on long-tail data and the improvement of the efficiency of downstream models, the large model is considered to have the embryonic form of "general intelligence".
The "Implementation Plan" shows that Beijing plans to develop a basic theoretical framework system for the new generation of artificial intelligence, focusing on the mathematical mechanism of artificial intelligence, big data intelligence, multimodal intelligence, decision-making intelligence, brain-like intelligence, scientific intelligence, embodied intelligence, etc. Develop a research layout.
"Several Measures" clearly proposes to explore new paths of general artificial intelligence such as general intelligence, embodied intelligence and brain-like intelligence. Support value and causality-driven general agent research, create a unified theoretical framework system, rating standards, and testing platform, develop operating systems and programming languages, and promote the application of the underlying technical architecture of general agents. Promote the research and application of embodied intelligence systems, and make breakthroughs in the perception, cognition, and decision-making technologies of robots under complex conditions such as open environments, generalized scenarios, and continuous tasks. Support the exploration of brain-inspired intelligence, study core technologies such as brain neuron connection patterns, coding mechanisms, and information processing, and inspire new artificial neural network model modeling and training methods.
Among them, the "embodied intelligence" mentioned in the document refers to intelligent bodies that have a body and support physical interaction, such as home service robots and unmanned vehicles. Recently, this concept has been popularized by Nvidia founder and CEO Huang Renxun. In mid-May, he said in his speech at the 2023 ITF World Congress, "The next wave of artificial intelligence will be a new type of artificial intelligence called embodied artificial intelligence. Artificial intelligence, or intelligent systems that understand, reason, and interact with the physical world."
It is worth noting that the "Several Measures" also proposed to encourage third-party non-profit organizations to build multi-modal and multi-dimensional basic model evaluation benchmarks and evaluation methods; to study artificial intelligence-assisted model evaluation algorithms, and to develop models that include versatility, efficiency, A multi-dimensional basic model evaluation tool set including intelligence and robustness; build an open service platform for large model evaluation, establish a fair and efficient adaptive evaluation system, and realize automatic adaptive evaluation of large models according to different goals and tasks.
Since the explosion of ChatGPT, domestic large-scale models have experienced a "hurricane" spring. Zhao Zhiyun, director of the China Institute of Scientific and Technological Information, said on May 28 that according to incomplete statistics, 79 large-scale models with a scale of more than 1 billion parameters have been released in China so far, and 14 provinces and cities/regions are developing large-scale models. Facing the situation of "a hundred flowers blooming", how to evaluate large models has been put on the policy agenda. The Paper learned from many people familiar with the matter that many key artificial intelligence laboratories in Beijing and Shanghai are currently completing the evaluation work in a concentrated manner.
Just on June 3, the National Key Laboratory of Cognitive Intelligence, the Chinese Academy of Sciences Artificial Intelligence Industry-University-Research Innovation Alliance and the Yangtze River Delta Artificial Intelligence Industry Chain Alliance jointly released the "General Cognitive Intelligence Large Model Evaluation System", aiming to form a set of coverage The multi-task objective evaluation system of large-scale model capabilities guides the healthy development of China's cognitive large-scale model technology and industry.
Focus on six scenes
With the support of large-scale models, AIGC (Artificial Intelligence Generated Content) applications such as Vincent diagrams and virtual digital humans will quickly enter the commercialization stage.
"Several Measures" proposes to promote the application of general artificial intelligence technology innovation scenarios, focusing on six major scenarios of government services, medical care, scientific research, finance, autonomous driving, and urban governance.
The "Implementation Plan" emphasizes that relying on Beijing's superior scene resources, it is necessary to accelerate the convergence of capital, technology, data, computing power, talents and other elements, create a batch of benchmarking demonstration application scenarios that can be replicated and promoted, and promote the artificial intelligence innovation chain. The industrial chain, the capital chain, and the talent chain are deeply integrated. In addition, support artificial intelligence to empower the construction of smart cities, support Haidian District to build City Brain 2.0, and promote the smooth implementation of projects such as Beijing High-level Autonomous Driving Demonstration Zone 3.0.
The Partner Program mentions some more specific scenarios. "Focusing on scenarios such as Window of the Capital's intelligent question-and-answer and online guides, relying on privatized deployment of computing power cluster resources, we will gradually carry out work such as training, fine-tuning, pruning, and distillation of proprietary models in the government service industry, empowering the '12345 Immediately implement the 'auxiliary scene." "Focus on the city's virtual digital human, digital medical care, e-commerce retail and other innovative and active data advantage fields, accelerate the commercialization of large-scale models, and accelerate text creation, human-computer interaction, education, audio-visual, etc. Scenario landing application."
"I think it is still time to develop dedicated artificial intelligence. Even if it is ChatGPT or GPT (generative pre-training Transformer model), if it is to be truly industrialized, it still needs to be polished for the scene." Lu Yanxia said.
IDC believes that in the next five years, as human-computer interaction, machine learning, computer vision, and speech recognition technologies reach a more mature stage, the development trend of artificial intelligence applications includes creative work that uses knowledge as the main production tool (such as writing , video, image and audio creation, software development, IP incubation, etc.) will achieve a greater degree of intelligence; industry enterprises will expand the integration of digital twin and artificial intelligence technology, promote the development of energy and power, manufacturing, construction and other industries, build Virtual factories, digital twin power grids, and digital twin cities strengthen the connection between the digital and the real world, optimize processes, realize global management, and make intelligent decisions.
Create an inclusive and prudent regulatory environment
In terms of supervision, both documents refer to "inclusive prudence".
Zhou Xuejing told The Paper that "tolerance and prudence" at the regulatory level is divided into two dimensions: "tolerance" and "prudence". Tolerance is to encourage the innovation and development of domestic general artificial intelligence at the current stage, and to give enterprises the opportunity to try and make mistakes by establishing pilot projects (demonstration areas, test areas), thereby accelerating the formation of a number of benchmark demonstration cases, and accelerating the development of the entire industry . Prudence, on the other hand, starts from the perspective of security and compliance, and considers the impact of network security, data security, and other new risks that arise in the development process on the country, society, and individuals. Therefore, timely regulatory intervention in the market is required. .
"Essentially, tolerance and prudence is a flexible, classified and hierarchical regulatory strategy, especially in the early stage of the development of emerging technologies such as large models and AIGC and new industries, which can better reduce the trial and error costs and compliance costs of enterprises, and encourage my country's artificial intelligence innovation and development." Zhou Xuejing said.
How to create an inclusive and prudent regulatory environment? The "Several Measures" mainly includes four points:
First, continue to promote the innovation of regulatory policies and regulatory processes, encourage innovation subjects to adopt safe and reliable software, tools, computing and data resources, and carry out independent innovation, promotion and application, and international cooperation of basic technologies such as artificial intelligence algorithms and frameworks; , establish a normalized service and guidance mechanism, carry out normalized contact services for artificial intelligence-related Internet information services with public opinion attributes or social mobilization capabilities, guide innovation subjects to introduce technical tools for security testing, declare security assessments in accordance with regulations, and perform algorithm filing Third, strengthen network service security protection and personal data protection; fourth, continue to strengthen scientific and technological ethics governance.
In fact, the rapid development of generative artificial intelligence has triggered regulatory actions in many countries and regions around the world. Under the impetus of technology giants such as OpenAI and Google, the US government and Congress have launched discussions. The European Union, which is ahead in terms of regulation, has almost completed the legislative process for its "Artificial Intelligence Act".
The EU white paper "The Development of Artificial Intelligence for Excellence and Trust" provides an idea worthy of reference, that is, to carry out a "risk level-based" regulatory path. In principle, mandatory regulations are only applicable to high-risk artificial intelligence activities to ensure Regulatory intervention is applicable and targeted. In addition, multi-party participation in regulatory governance is implemented, and certification bodies designated by member states conduct independent review and evaluation of artificial intelligence systems.
At the implementation level, Zhou Xuejing analyzed that from the perspective of the current artificial intelligence supervision departments, my country presents the current situation of multi-management supervision. Relevant departments include the State Administration for Market Regulation, the National Internet Information Office, the Ministry of Industry and Information Technology, and the Ministry of Science and Technology. At the level of laws and regulations, on the one hand, it is to regulate the obligations and responsibilities of network operators in the use of artificial intelligence technology through specialized and comprehensive legislation; specification.
Talent Highland and Government Investment Fund
In addition to the five dimensions of computing power, data, algorithms, applications, and supervision, the Beijing series of documents also proposed specific measures in terms of talent and innovation.
The "Implementation Plan" has a clear goal: to build a group of artificial intelligence scientific research institutions with world-class influence, to introduce and cultivate world-class innovative talent teams, and to make new breakthroughs in international talent introduction. The number of high-level scholars exceeds 10,000, and the domestic proportion remains at the leading position.
The main tasks include promoting the construction of a highland of talents in the field of artificial intelligence. "Explore and implement the policy of supporting overseas talents to come to Beijing and focus on introducing and cultivating a group of top talents and young talents with world influence." Level scientists, industrial and engineering talents."
As one of the most important research institutions in the field of artificial intelligence in China, Beijing Zhiyuan Artificial Intelligence Research Institute has recently received unprecedented attention. Microsoft President Brad Smith (Brad Smith) said in April when talking about Microsoft and ChatGPT's competitors, "We see three institutions that are absolutely at the forefront, one is OpenAI and Microsoft, the second is Google, The third is Beijing Zhiyuan Artificial Intelligence Research Institute.”
Beijing Zhiyuan Artificial Intelligence Research Institute was established in December 2018. It positions itself as a "new research and development institution" and proposes a set of "Zhiyuan Model", which is to establish a scientific research management mechanism that combines free exploration and goal orientation, and selects from the perspective of small peers. Support the free exploration of Zhiyuan scholars, and promote the implementation of major scientific research tasks such as the "Enlightenment" model with the scientific research organization model of "concentrating efforts to do big things". In 2021, Zhiyuan Research Institute released the world's largest and most powerful intelligent model "Enlightenment 2.0", with a parameter volume of 1.75 trillion (ten times that of GPT-3.5). On May 28 this year, Beijing Zhiyuan Artificial Intelligence Research Institute launched the general segmentation model SegGPT (Segment Everything In Context) at the parallel forum of the 2023 Zhongguancun Forum.
In addition to world-class research institutions, the innovation environment of enterprises is also particularly important. The "Implementation Plan" puts forward goals: the research investment of leading artificial intelligence enterprises will continue to increase, the number of start-up enterprises will continue to increase, the total number of enterprises will maintain the leading position in China, and 5-10 new unicorn enterprises will be cultivated.
The main tasks include giving play to the guiding role of government investment funds and supporting long-term capital and patient capital in early-stage hard technology investments in artificial intelligence chips, frameworks, and core algorithms. Continue to do a good job in the cultivation of artificial intelligence enterprises listed on the market.
The "Implementation Plan" pointed out that it is necessary to strengthen the gradient cultivation of artificial intelligence enterprises. "Intensify the cultivation of innovative small and medium-sized enterprises, and bring companies that have the potential to become unicorns into the cultivation system in advance." "Promote a group of internationally renowned research institutions, multinational companies, and domestic leading companies to build innovative business entities in Beijing." " Implement the two-level enterprise service package and service stewardship mechanism in urban areas, and for innovative companies with potential in the field of artificial intelligence, the standards for inclusion in service packages can be appropriately relaxed to increase service coverage.”
Similar to Beijing's policy, the recently released "Shenzhen City Accelerates the Promotion of High-quality Development of Artificial Intelligence and High-Level Application Action Plan (2023-2024)" proposes that Shenzhen will play the role of government investment guidance funds, coordinate and integrate fund resources, and form a scale of 1,000 100 million yuan artificial intelligence fund group. Almost at the same time, "Several Policies and Measures for Strengthening Support for the Development of Private Investment in Shanghai" was released, which will use the "fund investment" model to gather efforts to attract "three leading industries", artificial intelligence is one of the "three leading industries" one.