The context, components, capabilities, and consequences of China’s strategy to lead the world in AI
Jeffrey Ding* Governance of AI Program, Future of Humanity Institute, University of Oxford March 2018
*Address correspondence to jeffrey.ding@magd.ox.ac.uk
For comments and input, I thank: Miles Brundage, Allan Dafoe, Eric Drexler, Sophie-Charlotte Fisher, Carrick Flynn, Ben Garfinkel, Jimmy Goodrich, Chelsea Guo, Elsa Kania, Jade Leung, Jared Milfred, Luke Muehlhauser, Brian Tse, Helen Toner, and Baobao Zhang. A special thanks goes to Danit Gal for feedback that inspired me to rewrite significant portions of the report, Roxanne Heston for her generous help in improving the structure of the report, and Laura Pomarius for formatting and production assistance.
EXECUTIVE SUMMARY
Marked by the State Council’s release of a national strategy for AI development in July 2017, China’s pursuit of AI has, arguably, been “the story” of the past year. Deciphering this story requires an understanding of the messy combination of two subjects, China and AI, both of which are already difficult enough to comprehend on their own. Toward that end, I outline the key features of China’s strategy to lead the world in AI and attempt to address a few misconceptions about China’s AI dream. Building off of the excellent reporting and analysis of others on China’s AI development, this report also draws on my translations of Chinese texts on AI policy, a compilation of metrics on China’s AI capabilities vis-à-vis other countries, and conversations with those who have consulted with Chinese companies and institutions involved in shaping the AI scene.
The report is organized in four parts: (1) Context - I place China’s AI strategy in the context of its past science and technology plans, as well as other countries’ AI plans; (2) Components - I relate the key features of China’s AI strategy to the drivers of AI development (e.g. data, talented scientists); (3) Capabilities - I assess China’s current AI capabilities by constructing a novel index to measure a country’s AI potential; (4) Consequences - I highlight the potential implications of China’s AI dream for issues of AI safety, national security, economic development, and social governance. In each of these four parts, I dispel a common misconception about China’s approach to AI (Table 1). Then, using the deconstruction of these myths as a starting point, I derive my own findings. What follows is a summary of the key findings in each section.
Table1:DemystifyingChina’sAIDream
Myths
1. The State Council’s AI plan was the starting point of China’s AI planning
2. China’s approach to AI is defined by its top-down and monolithic nature
Reality
The plan both formalizes and definitively signals a national-level focus on AI, but local governments and companies were already engaging in subnational planning on AI. Additionally, crucial elements of the State Council’s AI plan are rooted in past science and technology plans.
While the central government plays an important guiding role, bureaucratic agencies, private companies, academic labs, and subnational governments are all pursuing their own interests to stake out their claims to China’s AI dream.
3. China is winning the AI arms race China may define “winning” differently than the U.S., and, according to my AI Potential Index (AIPI), China’s AI capabilities are about half of those of America.
4. There is little to no discussion of issues of AI ethics and safety in China
Substantive discussions about AI safety and ethics are emerging in China. A new book authored by Tencent’s Research Institute contains chapters that are relatively proactive in calling for stronger awareness of AI safety issues. No consensus exists on the endpoints of AI development.
The State Council’s AI plan is not the be-all and end-all of China’s AI strategy. In the “Context” section, this report benchmarks both the plan and China’s overall AI approach with regard to China’s current AI capabilities and the positions of other countries on AI. Analyzing China’s AI development in these two contexts gives the following conclusions:
• In addition to increased policy support for AI development in the past two years, the State Council plan’s targets for the growth of the AI industry confirm China’s high expectations for AI development. The 2020 benchmark for the core AI industry’s gross output (RMB 150 billion) would represent a tenfold increase of the AI industry in the next three years.1
• The plan clearly outlines China’s ambition to lead the world in AI. Additionally, Chinese AI experts and decision-makers are keenly aware of the AI strategies and capabilities of other countries, including the United States, the EU, Japan, and the United Kingdom. There is evidence that China focuses on U.S. AI strategy, in particular, as a reference point for its own approach. One reasonable hypothesis is that China views AI strategy as a bilateral competition to some extent; another is that the U.S. can credibly shape China’s approach in some respects.
In the second section on “Components,” I link the key features of China’s AI strategy – those consistent with other science and technology plans as well as those that differ – to four factors that drive the overall development of AI: (1) hardware in the form of chips for training and executing AI algorithms, (2) data as an input for AI algorithms, (3) research and algorithm development, and (4) the commercial AI ecosystem. Structuring the analysis by driver helps unpack how different features of China’s AI strategy are put in practice in the following ways:
• There are important similarities and differences between China’s approach to AI development and its past efforts to guide scientific and technological innovation in other areas. Key consistencies include: a strong degree of state support and intervention, transfer of both technology and talent, and investment in long-term, whole-of-society measures. Significant differences are rooted in two factors: AI’s “omni-use” potential means the breadth of actors involved is much wider than for other technologies, and as a result, internationally-facing, private tech giants and vigorous startups are leading players in driving innovation in AI.
• China is adopting a “catch-up” approach in the hardware necessary to train and execute AI algorithms. It has supported “national champions” with substantial funding, encouraged domestic companies to acquire chip technology through overseas deals, and made long-term bets on supercomputing facilities. Importantly, established tech companies like Baidu and startups like Cambricorn are designing chips specifically for use by AI algorithms.
• Access to large quantities of data is an important driver for AI systems. China’s data protectionism favors Chinese AI companies in accessing data from China’s large domestic market but it also detracts from cross-border pooling of data. Also, the common view that China’s AI development will benefit from relatively lax privacy protections on user data may no longer hold true with the promulgation of a new data protection law.
• China is actively recruiting and cultivating talented researchers to develop AI algorithms, another
1 This calculation takes the State Council’s targets and compares them to a iiMedia Research Group report’s estimate of the scale of China’s industry in 2017. I assume that the iiMedia Research Group’s estimate is close to what the State Council views as ground truth, but the State Council may be working off of a different estimate for the current gross output of the core AI industry (iiMedia, 2017).
essential factor in AI innovations. The State Council’s AI plan outlines a two-pronged “gathering” and “training” approach. National-level and local-level “talent programs” are gathering AI researchers to work in China, while China’s tech giants have set up their own overseas AI institutes to recruit foreign talent. The training plank takes a long-term view to growing AI talent through constructing an AI academic discipline and creating pilot AI institutes.
• Lastly, the Chinese government is starting to take a more active role in funding AI ventures, helping grow the fourth driver of AI development – the commercial AI ecosystem. Disbursing funds through “government guidance funds” (GGF) set up by local governments and state-owned companies, the government has invested more than USD 1 billion on domestic startups, with much of the investment shifting toward healthcare and AI as the priority areas in the last two years. At the same time, the central government is exploring methods, including through the establishment of party committees and “special management shares,” to exert more influence over large tech companies.
Next, the “Capabilities”section assesses the current state of China’s AI capabilities across the four drivers of AI development by constructing an AI Potential Index (AIPI), which approximates countries’ overall AI capabilities. For each driver, I find proxy measures that benchmark China’s capabilities as a proportion of the global total. Thus, a country’s AIPI, scored from 0 to 100, represents its share of the world’s total AI capabilities.
• China’s AIPI score is 17, which is about half of the U.S.’s AIPI score of 33. China trails the U.S. in every driver except for access to data. One could argue that China’s lead in data would outweigh its deficits in other drivers. The AIPI is useful for testing these arguments. I find that the relative importance of the data driver would have to be over four times that of each of the other three drivers for China’s AIPI score to equal that of the United States.
• Several caveats are important to note. The index is meant to be a first-cut estimate of the AI landscape, so the specific numbers are not as important as their relative sizes and differences. The methodology will need to be refined as this index is limited by proxy measures for which quantifiable, reliable data was collected for both the U.S. and China.
Finally, I examine the potential implications of China’s AI dream for issues of AI safety and ethics, national security, economic development, and social governance. I emphasize that Chinese thinking on these issues is becoming more diversified and substantive. Though it is too early for firm conclusions about the long-term trajectory of China’s AI development, it is useful to highlight the key areas of debate in each of these issues:
• One group of Chinese actors is increasingly engaged with issues of AI safety and ethics. A new book authored by Tencent’s Research Institute includes a chapter in which the authors discuss the Asilomar AI Principles in detail and call for “strong regulations” and “controlling spells” for AI.2 A wide range of Chinese AI researchers are also involved with translating the IEEE’s Ethically Aligned Design report, as part of the Global Initiative for Ethical Considerations in Artificial Intelligence
2 These terms are from my translations of the book, which are available upon request (Tencent Research Institute et al., 2017).
and Autonomous Systems. However, other Chinese AI leaders dismiss calls for regulation and philosophizing.
• Since military applications of AI could provide a decisive strategic advantage in international security, the degree to which China’s approach to military AI represents a revolution in military affairs is an important question to study. The level of civil-military integration will be a critical factor in keeping track of this question.
• Economic benefit is the primary, immediate driving force behind China’s development of AI. Per multiple reports, of all economies’ worldwide, China’s has the most to gain from AI technologies, since AI systems could enable China to improve its productivity levels and meet GDP targets.3 Initial figures are promising - new Chinese AI companies and investment in the years 2014-2016 surpassed the number of companies and amount of investment in all the years prior4 - but they should be tempered by the potential for speculative over-investment to cause boom-bust cycles.
• China’s adoption of AI technologies could also have implications for its mode of social governance. Per the State Council’s AI plan, AI will play an “irreplaceable” role in maintaining social stability, an aim reflected in local-level integrations of AI across a broad range of public services, including judicial services, medical care, and public security.5 Two key areas to watch are growing concerns about privacy and the willingness of private companies to participate in various social credit systems.
3 PwC, 2017; McKinsey Global Institute, 2017
4 Li, 2017
5 State Council, 2017a; for a full English translation of the “New Generation Artificial Intelligence Development Plan” by a group of experienced Chinese linguists with deep backgrounds on the subject matter, see this document from the New America foundation: https://www.newamerica.org/ cybersecurity-initiative/blog/chinas-plan-lead-ai-purpose-prospects-and-problems/.
INTRODUCTION
In his report to the 19th Party Congress in October 2017, Chinese President Xi Jinping reiterated his dream for China to become a “science and technology superpower.”6 Development of AI has become an integral part of China’s strategy to realize this goal. One turning point in China’s view of AI was the March 2016 victory by Google DeepMind’s AlphaGo over Lee Sedol,7 who is widely considered to be the greatest Go player of the past decade.8 Two professors who consulted on the State Council’s AI plan referred to AlphaGo’s mastery of the ancient Chinese strategy game as a “Sputnik moment,” prompting immediate reconsideration among government officials of China’s AI strategy. 9
A year later, the State Council issued the “New Generation AI Development Plan” in July 2017, formalizing existing investments in AI and unambiguously signaling China’s prioritization of AI development. The plan’s specific benchmarks for AI and AI-related industries10 – including by 2030 a gross output of RMB 1 trillion (U.S. $150.8 billion) for the core AI industry and RMB 10 trillion (1.5 trillion) for related industries –demonstrated China’s aspiration to lead the world in the field. While the plan serves as an important milestone in China’s AI development, it is still only one piece of China’s overall AI strategy. To explain the full picture, this report places the State Council plan and China’s broader approach to AI in the context of China’s past science and technology plans, as well as the AI strategies of other countries. It then analyzes China’s approaches to the growth of different drivers of AI development, and assesses the status and implications of China’s growing AI capabilities.
6 Economic Daily [jingji ribao], 2017
7 Borowiec, 2017
8 Deepmind, 2017
9 Mozur, 2017
10 The line between core AI and AI-related industries is fuzzy. In some AI plans, the Chinese government delineates core AI technologies from other related technology types like smart vehicles, smart wearable devices, and smart robots, among others. Under this conceptualization, core AI would include companies innovating in an industry-agnostic part of the AI architecture whereas AI-related industries would include parts of the AI pipeline focused on applications in specific industries. I clarify the term “gross output” in the introduction.
I. CONTEXT
China’s AI development plan did not begin with this State Council document in July; rather, the plan both formalizes and definitively signals a focus on AI—one that was already broadly known.11 For instance, a month before the State Council’s report, the government of the Chinese city of Tianjin had announced a USD 5 billion fund to support the AI industry.12 In this section, the report compares the plan and China’s overall AI approach with regard to China’s current AI capabilities, as well as the positions of other countries on AI.
A. China’s AI expectations vs. current scale of AI industry
The State Council’s plan represents the culmination of increased policy support for AI development. The Chinese government has significantly ramped up its AI plans in the past few years (Table 2). AI now appears among a select number of explicit government priorities in key, long-term plans related to science and technology13, and has backing from substantive funding measures – two key elements not present in past government support for AI.
Released in 2016, the “13th Five-Year Plan for Developing National Strategic and Emerging Industries” (2016-2020) identified AI development as 6th among 69 major tasks for the central government to pursue. The “Internet Plus” initiative, established in 2015, is tightly linked to AI development, as evidenced by the NDRC announcement of an “‘Internet Plus’ and AI Three-Year Implementation Plan” targeting the creation of an AI market that is hundreds of billions of RMB in size. Moreover, the NDRC, the Ministry of Industry and Information Technology, and the Ministry of Finance jointly released the “Robotics Industry Development Plan (2016-2020)” in April 2016.14 In 2017, Chinese Premier Li Keqiang incorporated the term “artificial intelligence” into the government’s work report15 for the first time, a development the news department of the State Council covered.16 Moreover, Chinese President Xi Jinping mentioned AI as a way to increase economic productivity in his opening speech of the 19th Party Congress.17
AI-related plans are increasingly tied to substantive funding mechanisms. Notably, in February 2017, the “Artificial Intelligence 2.0” plan received megaproject designation, which comes with substantial funding,
11 Mozur, 2017
12 Ibid
13 In 2016, the Five Year Plan for Developing National Strategic and Emerging Industries (2016-2020) highlighted development of AI as one of 69 major tasks for the central government to pursue; in 2017, “Artificial Intelligence 2.0,” a comprehensive effort to boost investment in AI education and development, was adopted as one of 16 Megaprojects in the Five Year Plan for National Science and Technology Innovation.
14 He, 2017
15 The government work report is an annual report on economic growth given by the Chinese premier to the National People’s Congress, China’s top legislative body. It summarizes the government’s efforts last year and outlines crucial tasks for the year ahead. Since the report normally sets the target GDP growth rate for the next year, its content is carefully scrutinized, making the work report an important signalling mechanism.
16 State Council, 2017b
17 Dwnews. 2017
Table 2: Recent AI Plans
Plan
13th Five Year Plan for Developing National Strategic and Emerging Industries (2016-2020) [“十 三五”国家战略性新兴 产业发展规划]
“Internet Plus” and AI ThreeYear Implementation Plan (2016-2018) [“互联网+”人 工智能三年行动实施方案]
Robotics Industry Development Plan (20162020) [机器人产业发展计划]
Description Key Elements Importance
A State Council policy document which specifies implementation measures for the 13th Five-Year Plan, focused on strategic industries
Jointly issued by the National Development and Reform Commission (NDRC)b, the MoST, MIIT, and the Cyberspace Administration of China
Plan to develop robotics industry released by the NDRC, the MIIT, and the Ministry of Finance (MOF)
Highlighted development of AI as 6th among 69 major tasksa for the central government to pursue; Identified five agencies responsible for developing central government policies in AI in the next five years
Established a goal to grow the scale of the AI industry’s market size to the “hundreds of billions” (RMB)
Links AI to the current Five Year Plan through this guiding plan
“Artificial Intelligence 2.0” [人
工智能2.0]
Proposal by Chinese Academy of Engineering added to a list of 15 “SciTech Innovation 2030 –Megaprojects”c
Three-Year Action Plan for Promoting Development of a New Generation Artificial Intelligence Industry (20182020)
MIIT action plan for implementing tasks related to State Council’s AI Plan and “Made in China 2025”
Set specific targets for advancing the robotics industry; the second of two development plans containing a focus on AI released by central agencies with a policy planning mandate
Megaprojects were proposed and finalized in 2016 with the release of the “13th Five-Year Plan for National Science and Technology Innovation” but AI was added in Feb. 2017
Sets out specific benchmarks for 2020 in a range of AI products and services, including smart, inter-connected cars, and intelligent service robots
Connects AI development to highly touted “Internet Plus” policy which aims to catapult China to becoming a digital powerhouse
Sets goal of manufacturing 100,000 industrial robots annually by 2020, making China the world’s leading robot-maker
Demonstrates how AI was elevated to the level of a megaproject only recently
Shows government’s strong guiding role in developing the AI industry (convened top 30 companies to develop indicators)
a Priorities listed above AI development, ordered from first to fifth: constructing internet network infrastructure, including rural broadband projects; improving radio and television networks; promoting “Internet Plus”; implementing big data development projects; and strengthening information and communications technology industries (State Council, 2016).
b The NDRC is the Chinese government’s central economic planning ministry. It has significant powers in allocating investment funds and approving major projects and has been dubbed China’s “mini State Council” and “number one ministry.” In recent years as President Xi’s administration has stressed a “decisive role for market forces, the NDRC has tried to reposition itself as a macroeconomic coordinator that is more relevant to a market-driven Chinese economy (Martin, 2014).
c For context, the original 15 S&T Innovation Megaprojects (2030) were announced in July of 2016, so AI was added on 7 months later to make 16 total projects. Focus areas for the other 15 include quantum communication, national cyberspace security, and neuroscience. These megaprojects are not new policy innovations. The “National Medium- and Long-Term Plan for the Development of Science and Technology (2006-2020)” also established 16 S&T Innovation Megaprojects to end in 2020. If past megaprojects are any precedent, the AI megaproject will likely involve a combination of significant grant money and various other policy levers (R&D tax credits, investment in talent pipeline, promotion of technical standards). The “mega-project” approach has been criticized by US-based scientists who are involved with the Chinese Academy Science for diverting resources from supporting investigator-driven projects. Another cynical take on megaprojects is that they are merely a repackaging of existing MOST programs and national programs administered by other agencies (Cao, Suttmeier, and Simon, 2006; Springut, Schlaikjer, and Chen, 2011).
alongside fifteen other technologies deemed crucial to China’s science and technology innovation.18 Additionally, the Fund for Industrial Restructuring and Upgrading allocated RMB 2.78 billion (USD 404.3 million) to projects in smart manufacturing in 2016 alone, and the 2017 Central Basic Infrastructure Budget allocated a combined RMB 5.28 billion (USD 614 million) to infrastructure for “Internet Plus” and “key projects in emerging industries” in 2017.19 Two other trends are notable. First, the history of China’s government support for AI-related development demonstrates a consistent emphasis on robotics and indigenous innovation, an indication that smart manufacturing will continue to be a priority. Second, bureaucratic agencies have begun to compete for authority over AI policy, a trend highlighted by the fact that the State Council has tasked 15 offices with implementing their AI plan.20
Analyzing the State Council plan’s targets for the growth of China’s AI industry in context of the current scale of its AI industry confirms China’s high expectations for AI development. The plan outlines an ambitious threestage process toward achieving China’s dream of leading the world in AI:21
1) By 2020, China’s AI industry will be “in line” with the most advanced countries, with a core AI industry gross output exceeding RMB 150 billion (USD 22.5 billion) and AI-related industry gross output exceeding RMB 1 trillion (USD 150.8 billion).22
2) By 2025, China aims to reach a “world-leading” level in some AI fields, with a core AI industry gross output exceeding RMB 400 billion (USD 60.3 billion) and AI-related industry gross output exceeding RMB 5 trillion (USD 754.0 billion).
3) By 2030, China seeks to become the world’s “primary” AI innovation center, with a core AI industry gross output exceeding RMB 1 trillion (USD 150.8 billion) and AI-related gross output exceeding RMB 10 trillion (USD 1.5 trillion).
Conceptually, these benchmarks map neatly onto three strategic phases of AI development: (1) catching up to the most advanced AI powers, (2) becoming one of the world leaders in AI, and (3) achieving primacy in AI innovation.
Unpacking the context behind the target numbers helps illustrate the degree of aspiration behind China’s AI push. According to iiMedia Research Group’s “2017 Special Report on China’s Artificial Intelligence Industry,”23
18 New Intellectual Report [xinzhiyuan baodao], 2017
19 He, 2017
20 China Economic Net [zhongguo jingjiwang], 2017
21 These are my translations of the report. Emphasis mine.
22 Some English-language reports on these benchmarks have translated them as “industry scale” or “market size” indicators, but the more precise translation is “gross output,” a measure of the production side of specific industries. An industry’s gross output is the sum of sales to final users in the economy (GDP) and sales to other industries (intermediate inputs). For a frame of reference, the estimated gross output of China’s robotics industry in 2017 was U.S.$6.8 billion. I cover the distinction between core AI and AI-related in the next two paragraphs.
23 iiMedia, 2017
China’s AI industry had a gross output of RMB 10 billion in 2016, and is expected to grow to around RMB 15 billion in 2017. Thus, the 2020 benchmark for the core AI industry’s gross output (RMB 150 billion) would represent a tenfold increase of the AI industry in the next three years.24 China’s ambitions in AI can also be understood in the context of the global AI industry. Per a report by McKinsey Global Institute, forecasts of the global market size for AI in 2025 range from USD 644 million to USD 126 billion.25 If these projections refer to core AI industries, China’s 2025 benchmark for a USD 60.3 billion, world-leading core AI industry corresponds with the high end of market forecasts for AI.
To be clear, the line between core AI and AI-related industries is fuzzy, so how China’s State Council interprets the difference between the two is important to analyze. The slipperiness of what exactly constitutes AI is a problem that plagues analysis of AI strategy. The flip side of this slipperiness is AI’s “omni-use” potential (i.e. its similarity to electricity), which I investigate later by comparing it to other technologies. Perhaps the most credible distinction in the Chinese context can be found in the “Internet Plus” and AI Three-Year Implementation Plan issued by the NDRC. This plan outlines nine major technology areas, listing “core AI technologies” along with eight other technology types. These “core AI technologies” include basic research in fields such as deep learning, the development of basic software and hardware such as chips and sensors, and applied research in areas like computer vision and cybersecurity.26
Notably, these core AI technologies are differentiated from the other eight technology types, which include smart vehicles, smart wearable devices, and smart robots, among others. The implementation plan’s definition of “core AI” fits with that of CB Insights, a leading market research firm, which defines “core AI companies” as those focused on general-purpose AI applicable across a variety of industries.27 Under this conceptualization, core AI would include companies innovating in a specific, industry-agnostic part of the AI architecture, whereas AIrelated companies would include parts of the AI pipeline focused on applications in specific industries.
B. China’s AI ambitions vs. other countries’ AI strategies
Some Chinese AI experts and decision-makers are keenly aware of the AI strategies and capabilities of other countries, in particular the United States, the EU, Japan, and the United Kingdom. In a chapter titled “Top-level Plans,” scholars from Tencent’s Research Institute and the China Academy of Information and Communications Technology, a research institute under the Ministry of Industry and Information Technology (MIIT), laid out their view of the current international strategic landscape for AI development as follows:28
• ‘Defend the lead’ America — a comprehensive, strategic layout: “In sum, the United States is,
24 The State Council may be working off of a different estimate for the current gross output of the core AI industry. This calculation assumes that the iiMedia Research Group’s estimate is close to what the State Council views as ground truth.
25 McKinsey Global Institute, 2017b
26 He, 2017
27 CB Insights Research, 2017
28 This arrangement is from my translations of a book published by Tencent on AI strategy (Tencent Research Institute et al., 2017).
at this point, the country that has introduced the most strategies and policy reports on artificial intelligence strategies. The United States is undoubtedly the forerunner in the field of artificial intelligence research and its every move necessarily affects the fate of all of humanity.”
• Ambitious EU — ‘Human Brain’ and ‘SPARC’ Projects: “In 2013, the European Union proposed a 10-year Human Brain Project, currently the most important human brain research project in the world.”
• Robot superpower Japan — ‘New Industrial Revolution’: “For the past 30 years, Japan has been called the “robot superpower” and has the world’s largest number of robot users, robotics equipment, and service manufacturers.”
• Unwilling to fall behind Britain — facing the fourth industrial revolution challenge: “The UK considers itself to be a global leader in ethical standards for robotics and AI systems. At the same time, this leadership in this area could extend to the field of artificial intelligence regulation.”
As for the authors’ assessment of China’s own position within this landscape, they titled China’s section as “China, from ‘running after’ to ‘setting the pace’,” and wrote the following, “In terms of AI, China followed the United States and Canada in releasing a national AI strategy. In the wave of AI industry, our country should go from system follower and move towards being a leader, actively seizing the strategic high ground.” 29 This concept of “setting the pace” is essential to understanding China’s high ambitions for its AI development.
Second, there is evidence that China is particularly attuned to U.S. AI strategy, and sees it as a reference point for its own approach. Many key junctures in China’s AI development are related to significant AIrelated pronouncements that are linked to the United States. For instance, after the Department of Defense’s announcement of the “Third Offset” strategy in 2014 – which Chinese defense analysts and policymakers followed closely – the Chinese military establishment responded by revising its modernization approach to increase investments into AI technologies.30 China also reacted to other significant developments in U.S. AI policy. In October 2016, the Obama administration released the first of three reports on AI, which also corresponded with a large spike in Baidu searches for AI. Some analysts have noted similarities between the State Council’s AI plan and these three reports, suggesting that the drafters of China’s AI plan were closely familiar with the previous U.S. administration’s policy statements.31 In 2016, the biggest spike in Baidu searches (the Chinese equivalent of Google searches) for “artificial intelligence” [人工智能] occurred right after AlphaGo’s victory, per a report by the Wuzhen Institute.32 Chinese leaders and scholars also paid significant attention to AlphaGo’s victory. After AlphaGo’s win in February 2016, high-level seminars and symposiums were conducted on the implications. One such event was “A Summary of the Workshop on the Game between AlphaGo and Lee Sedol and the Intelligentization of Military Command and Decision-Making” (围棋人机大战与军事指挥决 策智能化研讨会观点综述), which took place in April 2016 and included PLA thinkers from the Academy of
29 Ibid
30 Wood, 2017
31 Allen and Kania, 2017
32 Wuzhen Institute, 2017
Military Science and the Central Military Commission.33
There are multiple ways to interpret what lessons Chinese decision-makers took away from these critical junctures. While AlphaGo shocked the entire world, its victory over Lee Sedol appeared to have particularly affected China, where the game was invented.34 Perhaps concerned that AlphaGo’s mastery of Go would sting the country’s national pride, the Chinese government banned outlets from covering its May 2016 match with the Chinese player Ke Jie, the world’s number one player at the time.35 For China, AlphaGo may have demonstrated that advances in AI are linked to national prestige and the perceived status of great powers. Additionally, the types of high-level seminars conducted after AlphaGo indicate that some Chinese policymakers interpreted AlphaGo’s victory as having significant implications for military affairs. Per testimony before the U.S.-China Economic and Security Review Commission by Elsa Kania, the PLA “anticipates the advent of artificial intelligence will fundamentally alter the character of warfare, ultimate resulting in a transformation from today’s ‘informatized’ ways of warfare to future ‘intelligentized’ warfare.”36 Another reasonable hypothesis is that China’s reaction to major American AI-related developments, including the way in which the State Council’s plan was drafted, is partly inspired by U.S. strategy.37 Under this interpretation, the U.S. government may have some degree of influence in shaping a potential template for China’s AI planning.
33 China Military Science Editorial Department [zhongguo junshi kexue bianjibu], 2016 cited in Kania, 2017
34 Hern, 2017
35 Ibid
36 Kania, 2017
37 This may fall prey to a “mirror-imaging” bias, the assumption that another actor will react to and interpret events in the same way as oneself (Inkster, 2016).
II. COMPONENTS
A. Key consistencies and differences with other science and technology plans
Much of China’s approach to AI is old in the sense that it is consistent with past science and technology plans. While there are also some critical new factors, the features that stay consistent are important to highlight because they can be mined for empirical examples. Chinese government support for AI development, emphasis on indigenous innovation, and prioritization of frontier technologies traces back to February 2006, when the State Council issued their “National Medium- and Long-Term Plan (MLP) for the Development of Science and Technology (2006-2020).”At the time, the MLP was Beijing’s most ambitious science and technology plan to date. It allocated long-term funding for science research, estimated at RMB 500 billion (USD 75 billion), and launched sixteen national megaprojects for developing vanguard science and technology, including programs for integrated circuit manufacturing and large advanced nuclear reactors.38 Indeed, the designation of “Artificial Intelligence 2.0” as a megaproject follows the framework set by the MLP. The plan also contained an explicit target to strengthen indigenous innovation. China’s “Made in China 2025” initiative, released in May 2015, further emphasized the need for indigenous innovation to reduce the country’s dependence on other countries for high-end manufacturing.
In the drafting of the MLP, infighting among Chinese scientists and bureaucrats became so serious that it leaked out into the public sphere, an underappreciated aspect of Chinese science and technology policy that also applies in the AI context. In the early 2000s, Premier Wen Jiabao brought together the Chinese Academy of Sciences (CAS) and the Ministry of Science and Technology (MoST) to draft this MLP: in total, 2000 bureaucrats, researchers, and business managers were involved in the drafting process.39 As the bureaucrats at MoST and MIIT gradually shifted the direction of the MLP toward megaprojects, Chinese scientists bristled at the degree of control given to bureaucrats over scientific inquiry. In fall of 2004, a group of prominent Chinese scientists, from both inside and outside of China, published a collection of essays in a special issue of Nature that criticized the draft MLP plan.
There is some evidence that similar infighting has already begun over AI policy. The “‘Internet Plus’ and AI Three-Year Implementation Plan” gives four agencies – the NDRC, the MoST, the MIIT, and the Cyberspace Administration of China – the mandate to advance the AI industry. In contrast, the State Council’s “New Generation AI Development Plan” called for the establishment of an AI Plan Implementation Office under the authority of MoST. None of the other bureaucratic entities involved with the “‘Internet Plus’ and AI ThreeYear Implementation Plan” received mention in the State Council’s new plan, a notable exclusion given how comprehensive the document is in other respects. One researcher at the Council on Foreign Relations posited that this was an instance of bureaucrats at MoST asserting their claim on high-tech developments, undercutting
38 Bitzinger and Raska, 2015
39 McGregor, 2010
the authority of other ministries or academic efforts.40 In December 2017, MIIT issued its own three-year action plan to implement tasks related to the State Council’s plan and “Made in China 2025.” When the AI Implementation Office was officially created four months later, the official number of agencies involved had risen to 15 offices. Two offices, MoST and NDRC, were named in the announcement, ensuring that bureaucratic infighting over China’s AI path will not cease anytime soon.41 Although the central government plays an important guiding role, bureaucratic agencies, private companies, academic labs, and subnational governments are all pursuing their own interests to stake out their claims to China’s AI dream.
Lastly, there are important similarities and differences between China’s approach to AI development and its past efforts to spur innovation in strategic, emerging technologies. Take the example of biotechnology. The model of ramping up state support and intervention is similar to AI. First, there was a modest “climbing program,” which was initiated in the 1980s and lasted about eight years before the government made biotech more of a priority.42 Second, the Chinese government set up an independent entity, China National Center for Biotechnology Development, to coordinate the development of biotech, and important central planning documents begin to focus on the technology, in particular the State Council’s National Biotechnology Development Policy Outline in 1988, which established thirty national key laboratories.43 Third, the government signaled that biotech was a national-level priority and committed substantial funding toward its development. For example, the 863 program, China’s main vehicle for science and technology funding at the time, designated biotech as one of seven critical areas, and allocated around 1.5 billion RMB toward its development over the years 1986 to 2000.44
Other consistencies between China’s biotech strategy and AI approach include: international transfer of both technology and talent, as well as investment in whole-of-society and long-term measures. In the domain of tech transfer, Chinese firms in the pharmaceutical, biotech, and healthcare industries reached a record amount of $3.9 billion in overseas acquisitions in 2016.45 Talent programs have also attracted overseas Chinese working at the cutting edge of bioscience. “Deng Xiaoping sent many Chinese students and scholars out of China to developed countries 30 to 40 years ago, and now it is time for them to come back,” stated George Fu Gao, who received his doctorate at Oxford, of the Chinese Academy of Science’s Institute of Microbiology.46 There are also signs that China’s long-term investments in biotechnology are bearing fruit over thirty years later, as evidenced by recent advancements on cloning techniques and research on viral epidemics.47 However, China contributes relatively little to fundamental research: only 2.5 percent of the new molecules discovered from 2007 to 2015 came from China, compared to 56.3 percent from America.48
40 Council on Foreign Relations, 2017
41 China Economic Net [zhongguo jingjiwang], 2017
42 Huang and Wang, 2003
43 Ibid 44 Ibid
Bloomberg, 2016
46 Ball, 2018
47 Ibid
48 Bloomberg, 2017
Significant differences between China’s AI policy and biotech policy are rooted in two factors: AI’s “omni-use” potential means the breadth of actors involved is much wider than for other technologies; internationally-facing, private tech giants and vigorous startups are leading players in driving innovation in AI. Even the influence of the largest biotech companies pales in comparison to the power and sheer size of China’s tech giants. Consider the case of China’s genomic giant BGI, which has produced a number of major breakthroughs in genome sequencing. Its initial public offering raised $81 million, which is around 1/300 the size of Alibaba’s IPO.49 One could argue that it is unfair to take Alibaba’s IPO as the proxy for its influence in shaping AI development, because the IPO encompasses all of Alibaba’s business portfolio. That Alibaba announced in October 2017 an investment of $15 billion in AI-related R&D, with foci on quantum computing and human-machine interaction, serves as an effective rebuttal to that argument.50 Finally, core AI technologies are more fundamental than biotechnologies. That is, innovations in AI algorithms can revolutionize BGI’s genome sequencing, whereas the relationship does not operate in reverse. So while there are many similarities among China’s AI strategy and its aims for other strategic, emerging technologies, the immense power of tech companies and AI technology itself mark out key differences.
B. Channels from these key features to drivers of AI development
A comprehensive assessment of the components of China’s AI strategy requires an understanding of the broad range of drivers related to the AI development, including: (1) hardware in the form of chips and supercomputing facilities, (2) data as an input for AI algorithms, (3) research and algorithm development, and (4) the commercial AI ecosystem. Analyzing China’s current landscape for each of these drivers clarifies crucial features of its strategy to become a world leader in AI (Table 3).
With respect to hardware for AI algorithms, China’s promotion of national champions, encouragement of
Table3:KeyfeaturesofChina’sAIStrategy
Promote national champions, encourage overseas deals, build supercomputers
Share data between gov. and companies, protectionist toward cross-border flows Increasing privacy concerns toward AI applications
Research and Algorithms
Support for basic research, gathering and training talent
Commercial AI Sector Set up government guidance funds, picking winners
Tech giants establish overseas institutes to recruit AI talent
More actors involved (startups, local governments, agencies, etc.) due to omni-use capabilities
Tech giants and unicorn startups invest in AI chips
overseas acquisitions to facilitate technology transfer, and investment in supercomputers51 are all consistent with past approaches to spurring innovation in strategic technologies. First, it promulgated a national semiconductor policy in June 2014 that prioritized support for “national champions,” such as Tsinghua Unigroup.52 The policy launched a national Integrated Circuit Fund, which has raised more than $20 billion so far,53 with a goal to raise USD 138 billion total in funds to seed semiconductor investors throughout the country.54 In October 2017, China’s MoST announced a project to invest in chips that run artificial neural networks; as one of 13 “transformative” technology projects with a delivery date of 2021, the AI chip project specifically references Nvidia’s M40 chip as a benchmark, aiming to beat the M40’s performance and energy efficiency by 20 times.55
Second, the Chinese government has encouraged domestic companies that enjoy political support to sign deals with international firms to facilitate access to high-quality chip technology.56 A January 2017 report by the U.S. President’s Council of Advisors on Science and Technology on the semiconductor industry noted that Chinese firms have been increasingly active in the acquisition space and that China places conditions on access to its market in order to incentivize technology transfer.57 Recently, China’s two-pronged strategy has faced increased international scrutiny. After the U.S. government banned Intel and other chip-makers from selling China high-powered Xeon chips,58 the Committee on Foreign Investment in the United States (CFIUS) has subjected China’s investments in U.S. chip-makers to harsher scrutiny. In September 2017, the White House blocked a state-backed Chinese investment fund from acquiring a US semiconductor company, marking only the fourth time in history that an American president had blocked a corporate acquisition on national security grounds.59
A similar story has played out in Europe. In his 2017 State of the European Union Speech, Jean-Claude Juncker, president of the EU Commission, rolled out a new framework for screening foreign direct investments into the European Union. The framework identified critical technologies including, “artificial intelligence, robotics, semiconductors, technologies with potential dual-use applications, cybersecurity, space or nuclear technology.”60 While Juncker’s speech did not explicitly call out Chinese investments, analysts interpreted his warnings about investments from “state-owned companies” as an implicit reference to China’s economic activities.61
Third, China has made long-term bets on building supercomputing facilities. A few top-line figures indicate that China has made significant advances in the hardware necessary to power these potential breakthroughs. For
51 To date, it is uncertain whether supercomputers can spur AI-related progress. Chinese supercomputers, based on their own chips, have only been used for scientific projects and the chips have not been sold on the commercial market. I thank Jimmy Goodrich for this point.
52 Orr and Thomas, 2014
53 McKinsey Global Institute, 2017a
54 Weinland, 2017
55 Simonite, 2017
56 Ray et al. 2016
57 President’s Council of Advisors on Science and Technology, 2017
58 Tomson, 2015
59 Donnan, 2017
60 Fischer, 2017
61 John, 2017
instance, China surpassed the U.S. to have the most supercomputing facilities in the world at 167, compared to 164 in the U.S.,62 and China’s Sunway TaihuLight, which uses Chinese-designed processors, became the world’s fastest system in June 2016.63 Since much of the link to AI is speculative, I do not include these metrics in my index of AI capabilities, but some have argued that China’s long-term commitments to supercomputing facilities, along with its funding for quantum computing, may have real applications for AI.64
What is new in the hardware driver is that Chinese tech giants and unicorn startups are competitive with some of the world’s leading companies in designing AI chips. For instance, Chinese company Cambricon, a statebacked startup valued at $1 billion, has developed chips that are six times faster than the standard GPUs for deep learning applications and use a fraction of the power consumption.65 Moreover, equipped with a new “neural processing unit,” Huawei has arguably overtaken Apple in mobile AI chips.66
The Chinese government’s policies on the second driver, access to data, reveal two other critical aspects of its broader AI strategy: its leverage over big tech companies and its tendency toward protectionism. In October of 2016, some of China’s largest tech companies agreed to share data with government authorities to improve consumer trust online.67 The NDRC stated that the agreement was part of a broader project to create a national “social credit system,” which some privacy advocates have argued is designed for mass surveillance.68 As AI-fueled tech companies like the BAT companies become more and more powerful, the Chinese government has pushed for more influence over these big tech giants, even discussing the possibility of internet regulators taking 1% stakes in the companies.69 Dubbed “special management shares,” these small stakes would give Chinese government officials positions on company boards and the right to monitor content on the company’s online platforms.70
China’s sharing of data stops at the water’s edge. This fits with a larger trend of what some deem China’s techno-nationalism, an approach that aggressively protects domestic companies from foreign competitors.71 Even if the Chinese companies that rise from this approach do not compete internationally - though many have successfully expanded to Asian and African countries - they still thrive by serving China’s huge market. Data security concerns have motivated China’s efforts to ensure valuable data stays under the control of Chinese tech companies. In this vein, China has pushed for national standards in AI-related industries, such as cloud computing, industrial software, and big data, that differ from international standards, a move that may favor Chinese companies over foreign companies in the domestic market. According to a Mercator Institute report,
62 McKinsey Global Institute, 2017a
63 Vincent, 2016
64 Costello, 2017
65 Giles, 2017
66 Vincent, 2017
67 Clover and Ju, 2016
68 Ibid
69 Yuan, 2017
70 Zhong and Wee, 2017
71 For a history of this term, See: Feigenbaum, 2017
Chinese standards for smart manufacturing, cloud computing, industrial software, and big data differ significantly from the international standards in those domains.72 Data protectionism, such as the 2017 cybersecurity law that prevents foreign firms from storing data collected on Chinese customers outside of China, could disincentivize cross-border data pooling and the development of common standards for data sharing.73
One unique aspect of China’s AI development in the data driver is the emergence of a major debate over data privacy protections.74 Companies, different levels of government, and even the general public have been active participants in this debate, which pits those advocating for greater data privacy protections against those pushing for data liberalization to benefit AI technologies.75 In a chapter titled “Top-level Plans,” Tencent and CAICT researchers attribute the success of Silicon Valley to the existence of strong institutions such as copyright and tort law, and they argue that data liberalization is a form of institution building that could spur further innovation. They write, “If there is no government data liberalization policy, many AI applications will become ‘water without a source, a tree without roots.’ It can be said that the issue of data liberalization is a pain point in the development of AI in China and needs to be elaborated upon in a more comprehensive and in-depth manner in the strategy.” Recently, in January 2018, advocates for data privacy celebrated when the Chinese government released a new national standard on the protection of personal information, which contains more comprehensive and onerous requirements than even the European Union’s General Data Protection Regulation, per analysis by CSIS senior fellow Samm Sacks.76 This vigorous and unresolved debate over data privacy combats common misperceptions of China’s relatively lax privacy protections and is an important one to follow as China advances in AI.
In order to incentivize top quality-AI research and development, the State Council’s AI plan dedicates a section to accelerating the training and gathering of high-end AI talent.77 In the “gathering” section, the report calls for recruiting top international scientists through a variety of “Thousand Talents” plans. China’s Ten Thousand Talents program, launched in 2007 with substantial financial backing,78 has enticed talented scholars in AIrelated fields to work in China. Andrew Chi-Chih Yao, a Turing Award79 winner who renounced US citizenship, is now researching “AI theory development.” Additionally, Tim Byrnes, an Australian physicist is aiming to develop a quantum computer at NYU Shanghai, and Zhang Liang-jie, a former research staff member at IBM Watson, will investigate AI and virtual reality as chief scientist at enterprise software group Kingdee in Shenzhen.80 Lastly, Zenglin Xu, a former research associate at Purdue University, who now leads the statistical
72 Wübbeke et al, 2016
73 The Economist, 2017
74 I thank Danit Gal for pointing me toward this discussion.
75 Sacks, 2018
76 Ibid
77 State Council, 2017a
78 There are three main “Thousand Talents” Programs: 1. The “Long-Term Thousand Talents Program” awards grants of RMB 3 million (USD 452,000) to work full-time in China, as well as a one million RMB (USD 151,000) allowance; 2. The “Short-Term Thousand Talents Program requires hired employees to work in China for at least two months per academic year - talents will receive a RMB 500,000 (USD 75,000) allowance; 3. The Thousand Talents Program for Distinguished Young Scholars provides RMB 1-3 million (USD 151,000-USD 452,000) in research funding from the central government, an additional RMB 700,000 (USD 105,000) in research funding from the Chinese Academy of Sciences, as well as a RMB 600,000 (USD 90,376) allowance. Per the Chinese Academy of Sciences: http://english.ucas.ac.cn/index.php/join/job-vacancy/2040-the-long-term-thousand-talents-program
79 The Turing Award is often referred to as “the Nobel Prize of Computing.”
80 Lucas, and Feng, 2017
machine intelligence and learning lab at the University of Electronic Science and Technology of China, moved back to China through a portion of the Ten Thousand Talents program dedicated to attracting young academics.81
China’s talent programs have a mixed track record. From 2009 to 2011, the Thousand Talents program may have attracted the largest influx of high quality talent within a limited timeframe in all of China’s history, per data released by the Chinese Academy of Personnel Science.82 In those three years, 1510 scientists were selected as talent program awardees at the national level, out of an application pool of 6200.83 However, multiple empirical studies and interviews with recruiters for the talent programs reveal that these programs have not managed to attract the “best and brightest” Chinese scientists to return.84 A multitude of factors play a role, including: a research culture focused on instant results, lack of connection with domestic Chinese networks to advance, and problems with educational opportunities for their children. Nonetheless, as China works to reform its research culture and ramps up its efforts to encourage researchers to work in China, particularly those of Chinese descent, it could expand China’s pool of AI experts, as China’s scientific diaspora numbers over 400,000 scientists and other scholars.85
Talent transfer also occurs through commercial avenues: an investor who specializes in AI identified the strategy of hiring talented AI scientists to work in China - where salaries are now comparable to those in America, ranging from 70-150% of average pay for U.S. AI scientists - as a “shortcut” to accelerate AI development.86 In order to recruit foreign talent, the BAT companies have established their own overseas AI institutes.87 Worldleading AI talents have returned to China for work: Andrew Ng, former head of Google Brain, worked at Baidu for three years, and Qi Lu, former executive vice president of Microsoft, now serves as Baidu’s Chief Operating Officer. Headhunters working for China’s city governments and technology companies regularly visit international scholars and engineers in universities, companies, and startups and attempt to convince them to work in China.88 These different channels for talent transfer reveal an important point about China’s AI strategy - it is not a monolithic, completely top-down approach; many actors are maximizing their own interests and responding to broad signals from the central government.
Finally, China is taking the long-view to growing AI talent. The State Council’s plan also calls for constructing an AI academic discipline, involving a comprehensive effort to establish AI majors, create AI institutes in pilot
81 Zenglin Xu’s curriculum vitae is available at: http://www.bigdata-research.org/people/faculty/5.html. Note that Xu has won travel grants for NIPS and ICJAI.
82 Zweig and Wang, 2013
83 Ibid
84 Cao, 2008
85 Schiermeier, 2014
86 Harbringer, 2017
87 Alibaba recently invested USD 15 billion into global R&D, including 7 overseas labs, with a priority on AI; Baidu now has two research labs in Silicon Valley; and Tencent has established a lab in Seattle.
88 South China Morning Post, 2017
institutions, and include “AI + X” hybrid professional training.89 This whole-of-society push is a trademark of China’s central-guided development, and it demonstrates that China is placing a long-term bet on AI.90 While the government encourages the flow of talent and technology into Chinese AI sector, it prevents foreign companies from establishing a foothold in critical, AI-related sectors and restricts the flow of data out of China. The door is half open: China seeks to benefit from the open flow of talent and technology, while preventing international companies from gaining a foothold in its AI industry.
In the last driver regarding the commercial AI ecosystem, the Chinese government actively picks winners in the AI space. For example, in November 2017, MoST designated four companies — Baidu, Alibaba, Tencent, and iFlyTek — to lead the development of national AI innovation platforms in self-driving cars, smart cities, computer vision for medical diagnosis, and voice intelligence, respectively.91 These national endorsements could give Baidu an advantage in working with car manufacturers and Tencent wider access to hospital data, but they may also dampen competition in these specific markets.
The Chinese government is beginning to play a larger role in funding AI ventures. Disbursing funds through “government guidance funds” (GGF) set up by local governments and state-owned companies, the government has invested more than USD 1 billion on domestic startups.92 Per statistics from Sun Hung Kai Financial, these GGFs are projected to eclipse China’s private VC funds in size: for the year 2016, GGFs set a total fundraising target of RMB 3.3 trillion (USD 500 billion) vs. a RMB 2.2 trillion (USD 330 billion) total raised by private funds.93 One report on GGFs noted that from 2015 to 2016, the direction of GGF investment shifted toward healthcare and AI as the main priority areas.94 There is some initial evidence that increased GGF attention to AI has met some initial success. Per a 2017 report, China has a higher percentage of AI companies that have received investments (69%) than the U.S. (51%).95 Additionally, the velocity of AI investment is relatively fast: from incorporation to receiving angel investment, the average time for Chinese companies is 9.73 months while it is 14.82 months for US companies.96 These funds may help the central government achieve two goals at once, helping speed up AI development while also incorporating tech companies within the party apparatus. In the past few years, more than 35 tech companies, including Baidu and Sina, have created company party committees, which evaluate the company’s operations to ensure the party’s objectives are being followed.97
In some respects, the success or failure of China’s AI commercial sector will be a test of China’s unique mode of public-private partnerships. For reference, funding schemes similar to China’s GGFs were instrumental in transforming Israel into a leading technological powerhouse. Advantages for these types of government vehicles
89 State Council, 2017a
90 Council on Foreign Relations, 2017
91 Jing and Dai, 2017
92 Yang, 2017
93
include policy support, ample resources, and in some cases, a guaranteed minimum return for investors.98 But the fact that there has not been a single successful exit for any of the 911 GGFs to date reflects the scheme’s myriad issues, such as geographical and industry sector restrictions on investment and complicated exit procedures.99
Lastly, in all the drivers of AI, China is investing in long-term, whole-of-society approaches to advancing AI technologies. Indeed, the directives laid out in the State Council’s AI plan not only apply across government departments but they also strongly guide the actions of universities, research institutes, and the private sector. In contrast, other governments— limited in its power over society and subject to sudden policy shifts depending on which political party is in power — tend to implement short-term, whole-of-government solutions. What follows is an evaluation of how these components of China’s AI development have influenced its actual capabilities along the range of four drivers , captured by a series of comparative, quantitative metrics.
III. CAPABILITIES
In the four following sections, this report explains the importance of each driver in detail, so it will draw out some broader points about the relationship among drivers here. First, though this report analyzes each of the drivers separately, connections between drivers cannot be ignored. For instance, hardware improvements (e.g. the development of GPUs) have enhanced the performance for AI algorithms, and innovation algorithms have, in turn, enabled more efficient use of larger amounts of hardware through parallelization (running a program on multiple processors).100 Second, the importance of each driver relative to the others is the subject of much debate. When AI experts were surveyed on the sensitivity of AI progress to various drivers, opinions varied widely and no consensus was reached on the relative importance of each input.101 Other analysts have pointed out that the relative weighting of each driver has and will change over time, subject to significant trends like open access to advanced algorithms or large datasets.102 In the last part of the “Capabilities” section, this report assesses how adjusting the relative weight of each driver could change assessments of China’s AI capabilities.
A. Evaluation of China’s current AI capacities by driver
i. Catch-up approach in hardware
Due to their high initial costs and long creation cycle, processor and chip development may be the most difficult component of China’s AI plan. Currently, AI hardware falls into two categories: (1) chips originally designed for other computing processes but used to train AI algorithms (e.g. CPUs and GPUs) and (2) chips designed specifically to execute machine learning and deep learning algorithms (e.g. Google’s TPUs and Microsoft’s FPGAs).103 While the manufacturing of chips these two categories are more immediately relevant for running AI algorithms, supercomputing facilities may become relevant for future AI development if researchers are better able to leverage the benefits of co-located, interconnected compute.104
In the first category of hardware, measures of the strength of China’s semiconductor industry reveal a potential bottleneck for AI development. General metrics for traditional semiconductor firms are important to consider since these firms are scaling up their own processors to handle AI software, as well as acquiring startups that are building AI chips. In the year 2015, China only had 4% of the global market share of semiconductor production, while the U.S. accounted for 50% of the global market share.105 This correlates well with total financing figures which show that total financing for China’s semiconductor industry was only 4.3% of the amount for its
100 Brundage, 2016
101 AI Impact,s 2016
102 Cronin, 2016
103 CPU stands for central processing unit and is used for general purposes processing; GPU is a graphics processing unit, which was originally designed to process images but happen to be very efficient at training machine learning algorithms. A tensor processing unit (TPU) is a type of application-specific integrated circuit (ASIC), specialized for AI applications. Field-programmable gate arrays (FPGA) chips are reconfigurable, programmable hardware that are relatively efficient for AI applications.
104 I thank Miles Brundage and Allan Dafoe for this point.
105 International Trade Administration, 2016
U.S. counterpart, per a IT Juzi and Tencent Research Institute report.106 China is particularly dependent on international companies for GPUs, which are the best option for training AI algorithms. Microsoft AI researcher, XD Huang labels GPUs “the real weapon,” saying that without GPUs, a Microsoft project that recognizes certain conversational speech as well as humans would have taken 4 years longer to complete.107 Out of the top 10 American chip-makers, 4 specialize in making GPUs; whereas from the top 10 Chinese chip-making companies, none specialize in GPUs.108
In the second category of hardware, chips like TPUs and some ASICs are designed specifically to rapidly execute neural networks.109 Of the top 10 Chinese chip-makers, 6 specialize in ASIC chips, which are not as flexible as other chips in this category, such as FPGAs which provide high, efficient performance as well as flexibility to change the underlying hardware to adjust to rapidly changing software.110 Both the U.S. and China have two chip-making companies which specialize in FPGA chips out of their top 10 chip-making companies; the two U.S companies received a total of 192.5 million in total financing, while the two Chinese companies received a total 34.4 million in total funding.111 As with many aspects of AI, chip innovation is constantly occurring. For instance, Google recently launched a second-generation of TPUs, which Alphazero used to learn chess, that are able to train AI algorithms more efficiently than GPUs and CPUs.112
China’s success in building supercomputers demonstrates its potential to catch-up to world leaders in AI hardware. One metric that demonstrates this finding is the share of the highest-performing supercomputers located in China, per the global Top500 list. In 2014, China’s share of the Top500 list consisted of 76 systems (15.2%), which was a distant second to the U.S. at 232 systems (46.4%).113 The June 2017 version of the Top500 list saw China nearly catch up to the U.S., with the former boasting 159 systems (31.8%) and the latter having 168 systems (33.6%).114 Further distinctions can be made with respect to this category of hardware.. It is possible that supercomputing facilities can become more applicable in future AI development on a very large scale. Nevertheless, as noted by Larry Smarr, a physicist at the University of California, China’s excellence in manufacturing traditional supercomputers may not matter as much if other countries develop new, more efficient supercomputers that are designed specifically for challenges like AI.115
In sum, China has relied on imports and acquisitions to boost the most immediately relevant aspects of AI hardware. As this strategy has come under more scrutiny by the U.S. and EU, China is promoting national
106 Li, 2017
107 Metz, 2017
108 Li, 2017
109 Boundaries between the uses of different chips are fuzzy. Some companies use GPUs to execute algorithms as well. The tendency is for AI companies to use GPUs to train algorithms, and use TPUs and FPGAs to execute algorithms.
110 Freund, 2017
111 The IT Juzi and Tencent Institute report does not specify the time range of these figures. Per author’s check of the figures, they appear to refer to total money raised in all funding rounds since the company’s launch.
112 Tung, 2017
113 HPCwire, 2014
114 Author’s calculations from www.top500.org
115 Markoff, 2016
champions in its domestic chip-making industry and making long-term bets on powerful supercomputing facilities. In some respects, China’s approach to building its domestic semiconductor industry is a microcosm for its overall approach to AI development. State-directed theft of intellectual property, targeted poaching of talent, and strong government guidance have all been part of China’s brute force approach to boosting its semiconductor industry.116 Yet despite this effort, China’s domestic production of integrated circuits (IC) accounts for less than 13% of the country’s demand, and its trade deficit in the global IC market has more than doubled since 2005.117 Thus, catching up in the domain of AI hardware may take a long time, if it happens at all.
ii. Closed critical mass of data
Data is another important driver for AI systems because machine learning is notoriously data-hungry. Access to large quantities of data has been cited as one of the advantages for China’s AI development.118 With relatively lax privacy protections, Chinese technology giants collect vast troves of data, and sharing among government agencies and companies is common. Chinese consumers, the source of much of this data, are early and eager tech adopters, as reflected by smartphone penetration rates across the country and industry forecasts which show that the mainland will account for over 50% of the global retail e-commerce market by 2018.119 Per a report by CCID Consulting, China is projected to possess 30% of the world’s data by 2030.120 President of the Chinese Academy of Sciences, Bai Chunli, estimated, “By 2020, China will hold 20% of the global data, which is expected to reach 44 trillion gigabytes.”121
China’s data protectionism is part of a broader trend toward digital protectionism in which China’s internet is a closed ecosystem: in this world, the Chinese government censored and blocked Facebook and Google, thereby enabling the rise of domestic platforms like Wechat and Weibo. One can see the advantages of data protectionism for AI development. If data is a scarce resource for AI development, China could establish exclusive control over this resource for its companies and research institutes. On the other hand, if more and more data is shared across platforms and countries, other actors could benefit from global data sharing while China remains closed off.
iii. Algorithm development is high-quality but still lacking in fundamental innovation
Research and algorithm development is a critical factor for the advancement of AI. Chinese researchers are able to quickly replicate the most advanced algorithms developed anywhere in the world. Drawing from a domestic pool of talent, which includes the most STEM graduates out of any country in the world,122 China has pumped out a large quantity of AI research, but still cannot match the leading countries in the most innovative
116 I thank Elsa Kania for this framing.
117 Ernst, 2016
118 The Economist, 2017b; The New York Times 2017
119 South China Morning Post, 2016
120 Kania, 2017b
121 Chinese Academy of Sciences, 2017
122 World Economic Forum, 2016
Table4:AAAIConferencePresentationsbyCountry
Source: Japan’s National Institute of Science and Technology Policy. National affiliation of each presentation determined by location of researchers’ organization. Each co-author was counted once, so each presentation of findings could have resulted in more than one count recorded in the table (Koshiba et al. 2016).
research and the most talented researchers. In 2014, China surpassed the U.S. in the volume of AI research, as evidenced by metrics on AI-related patent registration and articles on deep learning,123 which was noted in the Obama White House’s strategic plan for AI research.124 This is not a case of volume devoid of quality: data on presentations at the Association for the Advancement of Artificial Intelligence (AAAI) annual conference, widely recognized as a leading AI research conference, revealed that Chinese researchers accounted for over 20% of the findings presented, second only to those from the United States (Table 4).
However, China lags behind both the U.S. and UK in fundamental research, according to a McKinsey Global Institute report which found that U.S. and UK research is more influential by citation impact, as measured by the H-index.125 When asked to compare the U.S. and Chinese AI strengths, Yann LeCun, director of Facebook’s AI research, highlighted the importance of the top advanced AI research labs, which have been established in the U.S. (Google Brain, Facebook AI Research, OpenAI, and others).126 Currently, both Chinese academics and companies tend to research applications of pre-existing AI technology; whether these two groups begin to adopt the “moonshot” mindsets that inspire the creation of new AI technologies will be a critical question for China’s future AI research.127
The difference in fundamental AI research may also be partly due to a talent shortage. Despite the larger pool of STEM graduates, China has a talent pool of around 39,000 AI researchers, less than half of the size of the U.S. pool of over 78,000 researchers.128 The U.S. benefits from having a large number of world-leading universities for AI research, as well as a more mature AI commercial ecosystem. This leads to more AI experts who have led multiple full cycles of projects. Nearly 50% of the AI researchers in the U.S. have more than 10 years of work
123 He, 2017
124 Zhang, 2017
125 McKinsey Global Institute, 2017a
126 Sixth Tone, 2017
127 Tse and Wang, 2017
128 Li, 2017; Studies have defined AI expertise in different ways. While Tencent’s methodology focuses on employees at AI companies, others use job sites like Linkedin or authors of conference papers (Gagne et al., 2018).
