1、AUTHORSWm.C.HannasHuey-Meei ChangDaniel H.ChouBrian FleegerChinas Advanced AI ResearchMONITORING CHINAS PATHS TO“GENERAL”ARTIFICIAL INTELLIGENCE JULY 2022智智能能科科技技Center for Security and Emerging Technology2Established in January 2019,the Center for Security and Emerging Technology(CSET)at Georgetown

2、s Walsh School of Foreign Service is a research organization fo-cused on studying the security impacts of emerging tech-nologies,supporting academic work in security and tech-nology studies,and delivering nonpartisan analysis to the policy community.CSET aims to prepare a generation of policymakers,

3、analysts,and diplomats to address the chal-lenges and opportunities of emerging technologies.CSET focuses on the effects of progress in artificial intelligence,advanced computing,and biotechnology.CSET.GEORGETOWN.EDU|CSETGEORGETOWN.EDUChinas Advanced AI ResearchJULY 2022AUTHORSWm.C.HannasHuey-Meei C

4、hangDaniel H.ChouBrian FleegerMONITORING CHINAS PATHS TO“GENERAL”ARTIFICIAL INTELLIGENCE PRINT AND ELECTRONIC DISTRIBUTION RIGHTS 2022 by the Center for Security and Emerging Technology.This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.To view a copy

5、of this license,visit:https:/creativecommons.org/licenses/by-nc/4.0/.Cover illustration:Intelligence science and technology in Chinese.DOCUMENT IDENTIFIER doi:10.51593/20210064Center for Security and Emerging TechnologyiEXECUTIVE SUMMARYINTRODUCTION1|GENERAL AI:DEFINITIONS AND PATHWAYSA.RECTIFYING N

6、AMES:“AGI”OR“GENERAL AI”B.FIVE THEORETICAL PATHS TO GENERAL AI2|CHINA ADVANCED AIA SEED LISTC.CHINAS MAINSTREAM APPROACHES TO ADVANCED AID.BRAIN-INSPIRED AI,CONNECTOMICS,WHOLE-BRAIN EMULATIONE.BRAIN-COMPUTER INTERFACES AND NEUROMORPHIC CHIPS3|MONITORING FOR SAFETY AND SECURITYF.WHY BUILD A CHINA AI“

7、WATCHBOARD?”G.CREATING A DATA MODEL OF THE PROBLEMH.WATCHBOARDING AND MONITORING4|THE SYSTEM IN OPERATIONI.EARLY EXAMPLES OF THE PROJECTS UTILITYJ.CHINESE EXPERT VIEWS ON“GENERAL AI”RECOMMENDATIONSAPPENDIX:INDICATORS AND KEYWORD LISTSENDNOTESIIIV172129434753ContentsCenter for Security and Emerging T

8、echnologyivCenter for Security and Emerging Technologyiiihis paper seeks to determine on the basis of publicly available information(“open sources”)who in China is taking what steps toward general artificial intelligence,as shown by overt expres-sions and other common measures.While typically concei

9、ved as“artifi-cial general intelligence”or AGI,this paper rejects that ambiguous term,along with its usual association with human-level machine intelligence,in favor of an approach that recognizes diverse pathways to broadly capable AI that functions autonomously in novel circumstances.Accordingly,t

10、he paper examines what paths to general AI are avail-able in principle,as a prelude to describing work underway in China to realize that capability.Three broad areas of Chinese research are identi-fied as potentially germane:machine learning,brain-inspired AI research,and brain-computer interfaces(B

11、CI).Data on the persons,institutions,and research making up this ecosystem is given as a foundation for down-stream studies,*and as a starting point for a China-focused indications and warning watchboard.The paper recounts the methodology used to build a database and prototype watchboard that enable

12、 analysts to capture and potentially forecast China AI-related events.Data supporting this pilot is conditioned to accept accretions from follow-on research,done locally or with outside participants,on Chinese artificial intelligence,AIs political uses,and other emerging Chinese technologies.The pro

13、ject aims to become or inspire a general foreign technology monitoring platform.Executive Summary T*Hannas and Chang,et al.,“Chinas Cognitive AI Research,”CSET(in preparation).Center for Security and Emerging TechnologyivCenter for Security and Emerging TechnologyvIntroductionn 2017,Chinas State Cou

14、ncil released a comprehensive“New Generation AI Development Plan”aimed at making China the worlds leading AI power by 2030.1 While aspirational,Chinas Communist Party,national government,universities,research labs,and technology companies over the subsequent half decade have demon-strated an unwaver

15、ing commitment to promote AI not only in applications,where China has a solid track record,2 but alsoas per the planin cut-ting-edge research that historically had not been Chinas forte.3 The plan has a complementary goal of achieving a“first mover advan-tage”(先發優勢).In AIs case,this implies an ever-

16、widening gap between the front-runnerChinaand less capable nations,owing to AIs(project-ed)ability to generate better versions of itself.If China succeeds in either of these aims,the impact on the United States and other liberal democracies ability to compete may be severely tested.The risk extends

17、beyond AI to technology in general due to AIs role as a technology enabler.Accordingly,this paper explores Chinas efforts to make headway in three main sub-fields named in the 2017 plan:traditional(computational)AI research,“connectomics”or brain-inspired AI studies,and brain-com-puter interfaces.Th

18、is report identifies institutions and persons that make up this ecosystem and provides descriptions of work done at some 30 orga-nizations to convey a sense of the AI plans actual scope and direction.China views this research as moving toward general artificial intelligencesoftware that can function

19、 in novel circumstances with greater autonomy and effectivenesswhich may,or may not,resemble human intelligence.ICenter for Security and Emerging TechnologyviGiven the near absence of data-driven research on this topicin Chinese language sources especiallyby observers in the United States and Europe

20、,and its potential import for competing nations,this paper also introduces an open-source pilot project,with extant examples,to monitor Chinas progress from narrow to broadly capable AI,using this studys data as a foundation.It concludes with a pro-posal to track Chinas AI growth as a target of conc

21、ern and as a starting point for a science and technology(S&T)monitoring watchboard,which the United States currently lacks.A study of Chinas technology development inevitably generates questions about state versus private management,foreign sources of inspiration,and impact on Chinese military power

22、.4 The authors address these topics in other studies to which the reader is referred.The short answers are:state involvement in Chinese AI is ubiquitous;5 virtually all Chinese technology programs benefit from foreign know-how;6 and Chinese planners are aware of AIs warfighting role.7 Research on ne

23、xt-generation AI with broad applicationsunlike todays AI which is mostly limited to niche areasleads to speculation that the outcome will be so-called“artificial general intelligence”(AGI),or by some accounts“artificial superintelligence,”(ASI)i.e.,AI that matches or exceeds the intelligent behavior

24、 of humans in many or most respects.This paper does not discount that possibility nor treat it as especially probable,but proposes that advances in AI may be unpredict-able or unrecognizable,taking forms that diverge from what laypersons regard as“intelligence”based on human experiences.While not ne

25、cessarily the fabled holy grail of human-level machine intelligence,the outcome of Chinas advanced AI research will be as portentous.Finally,describing advanced AI involves dissecting terms that interest AI pro-fessionals vitally and bear on how developments in this field are perceived.These section

26、s(1.A and B)can be skipped by non-specialistsbut at the cost of foregoing useful context.The authors thank Dr.Dewey Murdick for helping frame the study,Helen Toner for serving as its“red teamer,”Dr.Igor Mikolic-Torriera for his patience and encour-agement,Dr.Catherine Aiken for her detailed review,L

27、ynne Weil for marketing and production oversight,and Tessa Baker for administrative support.Thanks are also due Michael Miklaucic and John Chen for their reviews and substantive critiques,Maura McCarthy for her editorial review,Alex Friedland for copyediting support,and Ali Bours for layout and desi

28、gn.Center for Security and Emerging Technology1efining“AI”is problematic.In some ways it is a name in search of content.This section explores Chinese vs.U.S.and European views on AIs essence for a focal point from which to project its future trajectories.A.RECTIFYING NAMES:“AGI”OR“GENERAL AI”8 AI ha

29、s permeated global society and become a household word.Its successes notwithstanding,AI lacks a shared definition that captures what its makers are trying to create.9 To the extent that the focus is on narrow tasks,definitions may be irrelevantthe AI executes its code,albeit through processes not wh

30、olly transparent,and a particular result is obtained.However,as scientists seek more capable AIs that function in broad domains without specific guidance,a better understanding of the goalintelligenceis needed.Defining“intelligence”was problematic even before computers.10 Humans do not exhibit one t

31、ype of intelligence but demonstrate cognitive skills across a spectrum of tasks,such as visual-spatial processing,quan-titative reasoning,language use,and working memorynot to mention social and emotional intelligence.The discovery that measures of different mental abilities tend to correlate across

32、 tasks1 1led people to conceptual-ize human cognition in terms of a single“general intelligence factor,”a term that is valid in a restricted sense but misleading,12 and that has led to abuses.13 Moreover,measuring intelligence is not the same as understanding it.While it was possible for centuries t

33、o contemplate intelligence within a single biological domain,the advent of computing required an expanded General AI:Definitions and Pathways1DCenter for Security and Emerging Technology2definition to cover the problem-solving behavior of our electronic proxies.Inevita-bly,the sense that humansas a

34、seat of intellect“own”intelligence,coupled with a tradition of psychological research,led to an overly human-centric view of what“intelligence”is.This anthropomorphic outlook obscured the need for a foundation-al inquiry into the“intelligence”part of artificial intelligence,to the detriment of the d

35、iscipline and the ability to assess alternative AI programsthe discovery of which is a large part of any global monitoring regime.14 This complaint is widespread among specialists,not to mention analysts charged with informing and executing national policy,who struggle for a common defini-tion of th

36、e terms of discourse.15 Three leading artificial general intelligence experts explain how their niche has been impacted by terminological ambiguity:16“The debate on the essence of intelligence has been going on for decades,but there is still little sign of consensus.”17“The lack of a clear,universal

37、ly accepted definition is not unique to AGI.For instance,AI also has many different meanings within the AI research community,with no clear consensus on the definition.Intelligence is also a fairly vague concept.”18“For a long time,many different parties and factions in AI,adherent to more than one

38、ideology,have been trying to build AI without understanding intel-ligence.And their habits of thought have become ingrained in the field,and even transmitted to parts of the general public.”19 If“intelligence”and“AI”lack neutral(inclusive)definitions,“artificial general intelligence”is particularly

39、susceptible to association with its human progenitor.AGI typically is defined as“the hypothetical ability of an intelligent agent to understand or learn any intellectual task that a human can,”or“the capacity of an engineered system to display the same rough sort of general intelligence as humans,”o

40、r“the representation of generalized human cognitive abilities in software.”20 Proposed measures to test AGI such as the Turing Test,Coffee Test,and Robot College Student Test similarly set their standard by comparison with a human model.21This by itself is reasonable if human cognition is used as an

41、 extant example of a hypothesized general problem solving ability sought by software developers.The difficulty with casting this generalized goal in human terms is that the concept can be,and is,muddied by other human-like characteristics such as affect,conscious-ness,and morality,22 causing AIs dev

42、elopers to aim past what is needed,lose sight Center for Security and Emerging Technology3of alternatives or,potentially,ignore the research of countries like China less wed-ded to anthropomorphic paradigms.23 Equating AGI with human cognition distorts the former and impoverishes the latter.24 But t

43、he two are tied together so closely in lay understanding and even in specialized literature that the term AGI,in the authors view,has lost its operative value.25 Complicating matters,some approaches to AGI insist the AI be“embodied”human-like(have feedback mechanisms and reflect on its own outputs)t

44、o realize a general problem solving ability.26 If definitions matter within a language,they matter between languages.“Artificial general intelligence”is a recognized translation of the Chinese 通用人工智能(tngyng rngng zhnng),also rendered as“general(通用)artificial intelligence(人工智能).”Although both English

45、 terms are used in China,“general artificial intelligence”is preferred,as evidenced by its use in institutional nomenclaturee.g.,the 北京通用人工智能研究院(Beijing Institute for General Artificial Intelligence,BIGAI)27and,of course,by the Chinese word order itself.Herein lies the solution to the conundrum.By p

46、utting the adjective“general”before“artificial intelligence,”China,possibly without thinking about it,solved in one shot the two problems previously identified.It eliminates the problematic as-sertion of“general intelligence”in its original context and human intelligence per se as a necessary goal f

47、or artificial emulation.What remains is a generalized form of artificial intelligence,difficult to achieve in its own right,but with no necessary connection to the mixed attributes attached to“AGI”outside China.Use of the term“通用人工智能”in the hundreds of Chinese documents examined supports equating it

48、 with a general search for more broadly capable artificial intel-ligence,which may be comparable(by some measure)to human-level intelligence,or less capable,or exceed it in some or all domains.While the adjectives may be transposed(“人工通用智能”),it is far less commona neologism inspired probably by the

49、western“AGI,”which this papers authors hope does not take root.Use of the term“AGI”in this studyin cases where it cannot be avoidedis in the sense described here,namely“general AI”sans the anthropomorphisms and hyperbole.Changing the term of reference to“general AI”comports with trends in AIs de-vel

50、opment from software with narrow to wide applications.An example is OpenAIs large language model GPT-3,28 used in customer service,therapy,disinformation,and creative literature.Accordingly,this paper uses“general AI”as a synonym for“advanced AI.”This formulation acknowledges brain-inspired AI resea

51、rch and artifi-cially enhanced human intelligence as parts of AIs future but not necessarily the whole of it.Center for Security and Emerging Technology4B.FIVE THEORETICAL PATHS TO GENERAL AI29 Having stated what this study is not about,a working definition of“intelligence”is needed.Three definition

52、s taken from the literature follow(with comments from this reports authors in parentheses):“Intelligence measures an agents ability to achieve goals in a wide range of environments.”30(Narrow AI is an oxymoron.)“Intelligence is the capacity of a system to adapt to its environment while operating wit

53、h insufficient knowledge and resources.”31(Intelligence is gen-erative and transferable.)“This definition sees intelligence as efficient cross-domain optimization.”32 (A system that depends continuously on big data and overspecification is not intelligent.)These definitions,mutually compatible,refle

54、ct an understanding of AIs trajec-tory:narrow to wide,and allow a study of Chinas AI development without bias(advanced AI requires this or that computational approach),or framing the matter unproductively(who gets to“AGI”first).The core question can now be addressed:what paths is China taking to en-

55、hance the cognitive power of agents under its jurisdiction or control?33 Goertzel and Pennachin,in their foundational study of AGIs conceptual underpinnings,distin-guish four types of approaches to AGI,each with extant programmatic examples,with and without links to human cognition:34 1.approaches t

56、hat attempt to model biological brains2.approaches explicitly guided by the human mind and brain3.approaches inspired by the human mind much more than the brain4.approaches that depend little on known science about human intelligenceThis taxonomy is complemented by another schema,derived from a revi

57、ew of Chinese brain-inspired AI(BI-AI)programs,which qualifies“inspiration”as fol-lows:35 Inspiration by design.Accurate mathematical descriptions of real brain pro-cesses are run on computers to reproduce the behavior.De facto inspiration.Cognition is simulated by algorithms.The“inspiration”owes le

58、ss to explicit modeling and more to equating ML successes with macro-level brain processes.Center for Security and Emerging Technology5 Inspiration by default.Since“intelligence”derives originally from biological brains,efforts to emulate intelligence on artificial platforms are said to be brain-ins

59、pired.Taking Goertzel and Pennachins schema into account,the literature suggests there are(at least)five principled ways to build“advanced”AI:1.Attempt to understand intelligence with cues from human behavior and create machine algorithms that emulate it.36 This has been the majority view-point,asso

60、ciated with traditional ML/deep learning.37 2.Reverse-engineer a human brain on the assumption that what emerges is intelligence.38 This“neuromorphic”or brain-imitative approach derives function from structure and is the province of“brain-inspired AI”and“con-nectomics.”39 3.Force the emergence of in

61、telligence,in theory,by running algorithms fast enough to“recreate the same amount of cumulative optimization power that the relevant processes of natural selection instantiated throughout our evolutionary past.”40 4.Expand the definition of intelligence.41 As we argue above,there is no rea-son to v

62、iew intelligence as uniquely human.42 Any“de novo”AI substantially able to achieve wide goals would qualify.5.Finally,use brain-computer interfaces to position both elements,human and machine,to achieve(or overachieve)human goals.Embedded nanoscale chips and high-throughput cognitive“offloading”(par

63、tial brain emulation)are hypothetical approaches.These categories are prototypes culled from the AI literature to illustrate possibil-ities.In practice,they merge and their boundaries are disputed.Yudkowsky,for ex-ample,treats whole-brain emulation as outside the AI family,43 while Hansen views it a

64、s the paramount approach to AGI.44 Baum sees WBEs as“computational entities with general intelligence”and hence within the AI pale,45 and so on.These paths tee up the study of Chinas approaches to advanced AI,while un-derscoring the fact that no one knows for certain how advanced AI will be realized

65、;where and when it may happen;whether it will be smooth or discontinuous;or even what“it”is,as there are no precedents,no metrics to measure progress,and no clear break between AI and what meets(someones)threshold for“advanced”AI.46 Center for Security and Emerging Technology6Accordingly,this paper

66、examines three areas that cover some known possibili-ties,namely,so-called“mainstream”(主流)computational approaches to advanced AI;brain-inspired(類腦)AI research,including whole-brain emulation(“connecto-mics”);and brain-computer interfaces(腦機接口)that potentially result in cognitive augmentation.This d

67、ivision of paths to advanced intelligence is supported by most Chinese AI scientists,is grounded in state enactments dating from 2016,and sets the stage for effective mapping and monitoring(Sections 3 and 4 below).47 However,it is possible that this papers authors have overlooked some transformative

68、 develop-ment happening in a provincial lab or hiding in plain sight.Center for Security and Emerging Technology7he following is a record of Chinas top institutions engaged in one or more of the three typological areas of research identified as precursors to“general AI,”namely,mainstream“computation

69、al”approaches,brain-inspired approaches,and brain-computer interfac-es.The accounting is not exhaustivethat task is laid out in Section 3.Criteria for organizations to be included are professional status,accom-plishments,published research,and declarations of intent known to the authors from Chinese

70、 internet sources judged to be reliable.Descriptions of research conducted at each are provided as an introduction to Chi-nas multiple approaches.48 C.CHINAS MAINSTREAM APPROACHES TO ADVANCED AI49 Compute-intensive“big data”approaches to advanced(broadly capa-ble)AI constitute both in China and glob

71、ally the main focus of resources and attention.Here follow 10 such examples:1.Baidu Research(百度研究院),technical arm of the Chinese search engine giant,is part of the companys Baidu AI Technology Platform,further divided into a cognitive computing laboratory(認知計算實驗室),deep learning and big data labs,and

72、 others for biocomputing,robotics,safety,and quantum computing.50 It is directed by Wang Haifeng(王海峰),Baidus CTO.Baidu Vice President Wu Tian(吳甜)also serves as deputy director of Chinas National Engineering Laboratory for Deep Learning Technology and Applications(深度學習技術及應用國家工程實Chinas Advanced AIA Se

73、ed List2TCenter for Security and Emerging Technology8驗室),51 which,under Baidus leadership,announced in 2017 its intent to“build a research platform for general AI,”52 complementing the NELs work on com-puter vision,biometric ID,and human-machine interfaces.53 In July 2021,Baidu partnered with Pengch

74、eng Lab(鵬城實驗室,C.6 below)to launch“ERNIE 3.0 Titan,”54 a pre-trained large language model with 260 billion parameters55 that achieved state-of-the-art scores on 60 natural language processing(NLP)tasks,demonstrating its ability to“generalize across various downstream tasks with a small quantity of la

75、beled data.”56 2.Alibabas DAMO Academy(阿里巴巴達摩院),a branch of the e-commerce company,was founded in 2017 with facilities in Beijing,Hangzhou,and five sites abroad.57 Its director is Zhang Jianfeng(張建鋒),president of Alibaba Cloud Intelligence(阿里云智能).The academy hosts 16 laboratories researching eve-ryi

76、ng from machine intelligence(機器智能)to quantum.58 In June 2021,DAMO announced a trillion-parameter AI model with“nascent cognitive and creative abilities,and a goal of becoming the worlds leading AGI(通用性的人工智能)model.”59 Five months later,it released a 10 trillion-parameter Multi-Modality to Multi-Modal

77、ity Multitask Mega-transformer(M6)artificial intelligence system,touted as the worlds largest AI model.60 Zhou Jingren(周靖人),head of DAMOs data analytics and intelligence lab,explained that the goal is to improve M6 to“a level close to human beings”on the way to building general AI.61 A 2022 doc-umen

78、t“Top Ten Technology Trends of the DAMO Academy”(達摩院十大科技趨勢)regards“The ultra-large-scale pre-trained model as a breakthrough exploration from weak AI to general AI.”62 3.Tencent AI Lab(騰訊人工智能實驗室)is the research arm of Tencent(騰訊),the Shenzhen-based online gaming company and developer of the messagin

79、g app WeChat.63 Founded in 2016,the lab researches computer vision,speech recog-nition,NLP,and ML.64 It was under Zhang Tongs(張潼)tutelage from 2017 to 2019;65 speech recognition expert Yu Dong(俞棟)became vice director in May 2017.66 In 2019,Tencents AI agent Juewu(絕悟)won recognition for defeating top

80、 international gamers,67 while papers describing it were presented at the AAAI and NeurIPS conferences.68 Tencents goal of working toward“general AI”has been iterated in multiple venues.In July 2021,Juewus development team ac-knowledged the software is“a seed(種子)for general artificial intelligence.”

81、69 Tencent Vice President Yao Xing(姚星)claimed the company is“working together to tackle the ultimate goal of AGI.”70 And a July 2021 news release from Tencents AI lab stated that“AI game research will be a key step for Tencent to overcome the ultimate AI research problemGeneral Artificial Intelligen

82、ce(AGI).”71 Center for Security and Emerging Technology94.Huawei(華為),the Shenzhen-based telecommunications equipment company,is heavily invested in AI,as demonstrated by its Ascend(昇騰)AI chips,Atlas AI computing platform used in“Safe Cities”applications,and MindSpore(昇思)AI computing framework.72 The

83、 company signaled its formal entry into general AI development in May 2021,when it signed an agreement with the CAS Institute of Automation(CASIA,below)to build a“general artificial intelligence”platform in Wuhan.73 In fact,Huawei engineers contemplated hybrid approaches to AGI as early as 201774 an

84、d continue to explore its computational costs and work-arounds.75 The companys latest AGI precursor is Pangu (盤古),76 a 200-billion parameter pre-trained language model developed with Peking University and Pengcheng Laboratory.77 An indication of where Huawei may be headed is seen in its five-year AI

85、 investment plan,premised on achieving“general AI”by 2030the projected date when China achieves world AI dominance78and self-boot-strapping“artificial superintelligence”(超級人工智能)thereafter.79 5.JD Research Institute(京東AI研究院)was established in 2017 as a division of the Beijing-based e-commerce company

86、 JD.com.80 The institute has three labs for ML,computer vision,and NLP aimed at achieving“human-like cognitive abilities”in language and speech.81 This last goal is approximated by ViDA-MAN,“a dig-ital-human agent for multi-modal interaction,which offers realtime audio-visual responses to instant sp

87、eech inquiries”and sub-second latency.82 JD Research Institutes leaders are a distinguished lot:Director of AI Research Zhou Bowen(周伯文)was chief scientist of IBMs Watson Group;He Xiaodong(何曉冬),JDs VP of technology,is a 15-year Microsoft veteran;Institute VP Mei Tao(梅濤)is another Microsoft alumnus.JD

88、 sponsors a second research armthe JD Explore Academy(京東探索研究院)set up in 2020 for trusted AI(可信人工智能),super deep learn-ing,and quantum ML,all targeted at“disruptive”(顛覆式)innovation at the basic theoretical level.83 The deep learning model reportedly works from insufficient data and is capable of knowl

89、edge distillation and transfer learning.84 6.Pengcheng Lab(鵬城實驗室)was stood up in March 2018 as a Provincial Re-search Lab by the Guangdong and Shenzhen governments.The state-supported institution lists 59 universities,corporations,and research institutes as“strategic collaborators”beyond the Baidu a

90、nd Huawei links discussed above.85 Its out-reach opportunities were further expanded in January 2019 with the establish-ment of a Pengcheng Lab International AI Development Center(鵬城實驗室人工智能國際發展中心).86 Originally a network IT research facility,AI has assumed an increasingly large part of the labs port

91、folio.Pengcheng Lab has focused on da-ta-intensive large AI models,where its ultra-fast Cloud Brain II(云腦2),developed Center for Security and Emerging Technology10with Huawei to support AI computing,87 was used to create Pangu.88 Although its director,Gao Wen(高文),89 believes todays“narrow”AI will de

92、velop into“general AI,”90 he is less confident traditional AI can achieve that prospect and is moving toward a BI-AI paradigm.91 Pengcheng Labs partnership with NEL-BITA(see D.9 below)in June 2019 is evidence of that transition.92 7.Horizon Robotics(地平線機器人)was founded in 2015 in Beijings Haidian Dis

93、trict by Yu Kai(余凱),an ML expert and former director of Baidus IDL deep learning lab.93 Chen Liming(陳黎明)became its CEO in 2021.The company spe-cializes in AI chips for smart vehicles and computer vision.In 2018,Horizon es-tablished a“General AI Lab”in Californias Silicon Valley,the only major Chines

94、e company to research“general AI”there explicitly in name and in fact.94 The lab is headed by Xu Wei(徐偉),a highly regarded deep learning expert,who left IDL because of“present AIs extremely numerous shortcomings”to produce machines“with learning abilities like humans.”95 Xus goal to“form a small and

95、 sophisti-cated team focused on general artificial intelligence research”is reflected in job postings that seek“research scientists with strong background in fields of artificial intelligence.96 Our mission is to build artificial general intelligence(AGI).We will focus on developing novel algorithms

96、 and technologies to allow machines to learn new knowledge and skills as efficient as humans.”97 8.Beijing Institute for General Artificial Intelligence(北京通用人工智能研究院,BI-GAI),under Ministry of Science and Technology and Beijing municipal auspices,aims to create a“grand unified theory”(大一統理論)of intelli

97、gence,a“general intelligent agent”(通用智能體),and is assembling the capacity to pursue these goals seriously.98 The institute is led by returned UCLA professor and renowned AI scientist Zhu Songchun(朱松純),in concert with PKUs Institute for Artificial Intelligence and Tsinghuas own(future)AGI institute.99

98、 Plans are for a staff of one thousand research-ers drawn from China and“all over the world.”100 The authors of this report predict it will lead to clones,first in Shanghai then the provinces.101 Beyond its scale,BIGAI is important for being the first state AI facility to bear the“AGI”name,102 forma

99、lizing Chinas acceptance of a paradigm shift that we have been at pains to document,103 from big data-dependent“narrow AI”to broadly capable AI that transfers learned patterns to new and unforeseen problems.104 Its leadership views AGI as“the focus of international AI competition”in the coming decad

100、e.1059.CAS Institute of Automation(中科院自動化研究所),established in Beijing in 1956,straddles“traditional”and“brain-inspired”AI research,exemplifying the shift from compute-heavy AI to broader,more practical models.CASIA is host to several AI luminaries,including Director Xu Bo(徐波),Zeng Yi(曾毅),who runs its

101、 Center for Security and Emerging Technology1 1Research Center for Brain-inspired Intelligence(類腦智能研究中心),Jiang Tianzi(蔣田仔)head of its Brainnetome Center(腦網絡組研究中心),and Tan Tieniu(譚鐵牛),the computer vision expert and deputy chief of the PRCs liaison office in Hong Kong.Its multi-modal pre-trained model

102、 Zidong Taichu(紫東太初)with 100 billion parameters is seen as“an important first step from perceptual intelligence to general intelligence.”106 In late 2021,CASIA took the lead in a Multimodal Artificial Intelligence Industry Alliance meant to focus academic,laboratory,and industrial efforts on the eme

103、rging paradigm.107 Meanwhile,the institute research-es brain anatomical and functional connectivity at the micro-,meso-,and mac-ro-scales and the application of these findings to advanced AI models.108 10.Bohai University(渤海大學)is AGI advocate Liu Kais(劉凱)institute of affiliation and a platform for T

104、emple University AI maven Pei Wang(王培).109 Professor Wang is a world-class AGI innovator,second in name recognition only to Ben Goertzel,whose multiple publications decisively shaped our thinking.1 10 Wangs non-axiomatic model of intelligence offers a uniform foundation for AI and human thought and

105、sup-port for a credible AGI modelNARS.1 1 1 Liu Kai chairs regular sessions of the China Artificial General Intelligence Annual Conference(中國通用人工智能年會),an exten-sion of the global event that Goertzel and Wang tend to dominate;the China ses-sions are followed by AGI workshops.1 12 The 2021 convention

106、featured keynote talks on number sense,computer vision,and methods for building self models.1 13 Liu,who taught at Wuhans Huazhong Normal University(華中師范大學)until 2017,teaches machine education(機器教育),computational psychiatry,and brain-like systems in the universitys College of Education Science,an un

107、common affiliation that makes perfect sense in an AGI context.1 14 BOX 1Major Chinese entities pursuing general AI via traditional(computational)approaches Baidu Research Alibabas DAMO Academy Tencent AI Lab Huawei JD Research Institute Pengcheng Lab Horizon Robotics Beijing Institute for General Ar

108、tificial Intelligence CAS Institute of Automation Bohai UniversityCenter for Security and Emerging Technology12D.BRAIN-INSPIRED AI,CONNECTOMICS,WHOLE-BRAIN EMULATIONChinese critics of mainstream“big datasmall task”computational approaches to general AI point to biological brains as models of inspira

109、tion or,in extreme cases,detailed emulation.The extent of such Chinese efforts to“merge”artificial and biological intelligence can be discerned in the following examples:1.The Beijing Academy of Artificial Intelligence(北京智源人工智能研究院),also called the Zhiyuan Research Institute(智源研究院),was established in

110、 2018 under Huang Tiejun(黃鐵軍),vice-dean of PKUs Institute for Artificial Intelligence (人工智能研究院),1 15 to integrate neuroscience,cognitive science,and information science,on its way to building“strong artificial intelligence”(強人工智能)and“super-brain”(超腦)intelligent systems.1 16 Chinas Ministry of Scienc

111、e and Tech-nology described the goal as“transformative and disruptive breakthroughs.”1 17 Xu Bo(徐波),Tang Jie(唐杰),and Liu Jia(劉嘉),all strong proponents of AGI,have leadership roles.1 18 Zhang Hongjiang(張宏江),former CTO of Microsoft Research Asia is also on its board.Recent accomplishments are the Shen

112、ji(神機)series of simulation platforms and a“Bio-intelligence Open Source Platform”(生物智能開源開放平臺)with five components.1 19 BAAI is the home of Wudao(悟道)2.0,a multi-modal AI model comparable to GPT-3,meant to“enable machines to think like humans and move toward general AI.”120 2.Beijing Normal University

113、s(BNU)State Key Laboratory of Cognitive Neu-roscience and Learning(認知神經科學與學習國家重點實驗室)was established in 2005 to perform micro-and mesoscopic connectomics research,with a recent focus on reward processing and long-term episodic memory.121 BNU is one of three Chinese universities,with Peking and Tsingh

114、ua,allied with MITs McGov-ern Institute for Brain Research.122 The same three universities and a fourth Beijing facilitythe CAS Institute of Psychology(中科院心理研究所)form the Chinese end of a“Transregional Collaborative Research Centre on Crossmodal Learning”run by Zhang Jianwei(張建偉)and colleagues at Uni

115、versitt Hamburg,aimed at describing“the neural,cognitive,and computational mechanisms of crossmodal learning.”123 Its goals are improved deep learning and use of brain-computer interfaces to“accelerate AI.”124 The CAS Institute of Psychology is home to Zuo Xi-nian(左西年),lead author of“An open science

116、 resource for establishing reliability and reproducibility in functional connectomics.”125 Zuos doctorate is from BNU.1263.The Chinese Institute for Brain Research(北京腦科學與類腦研究中心,CIBR),also in Beijing,dates from March 2018.CIBR was“strategically deployed”by the Center for Security and Emerging Technol

117、ogy13citys S&T commission as a cooperative framework for Beijing-area universities,the PLA Academy of Military Science(軍事科學院),and others.127 Its mission is“co-ordinating research institutes and managing research programs under the guid-ance of the China Brain Initiative and Beijing Brain Initiative,

118、and making Beijing the world epicenter for neuroscience and brain-inspired computation.”128 Besides medical research,CIBR studies brain-inspired AI,optical imaging,and brain-computer interaction(腦機交互作用).129 The institute is co-directed by Rao Yi(饒毅),president of Capital Medical University in Beijing

119、,former dean of sciences at Peking University,and founding director of the PKU-IDG/McGovern Institute for Brain Research,and Luo Minmin(羅敏敏),an investigator at Beijings National Institute of Biological Sciences and a professor at Tsinghua University.Pu Muming and the Allen Institutes Christof Koch a

120、re on its advisory board.130 4.The State Key Laboratory of Brain&Cognitive Science(腦與認知科學國家重點實驗室),another Beijing-based institute,was established in 2005 as a part of CASs Institute of Biophysics(生物物理研究所).The lab does multi-disciplinary research on the“cognitive basic unit,”learning and decision-mak

121、ing,and neural mechanisms of information processing in drosophila and non-human primates,131 supported by an“ultra-high field MRI”platform used for brain imaging.132 It is headed by Chen Lin(陳霖),who specializes in visual cognition and brain imaging.Its academic committee reads like a whos who in Chi

122、na BI-AI,including BAAIs Xu Bo;CIBRs Luo Minmin;Guo Aike(郭愛克),also at CASs Institute of Neurosci-ence(CAS-ION,functional brain mapping,mind-body problem);Zhang Xu(張旭),director of the Chinese Academy of Science and Technologys“Neuroscience Direction Forecasting and Technology Roadmap project;”and CAS

123、IAs Tan Tieniu(譚鐵牛).133 The lab is creating a platform to track visual cognition from the genetic to behavioral level.134 5.CASs Center for Excellence in Brain Science and Intelligence Technology(腦科學與智能技術卓越創新中心,CEBSIT),established in 2014,is an umbrella orga-nization for 39 research institutions dis

124、tributed in Beijing,Shanghai,and 13 other locations.135 The CAS Institute of Neuroscience(神經科學研究所,ION)and CASIA are its main“supporting units.”136 CEBSIT is directed by Pu Muming(蒲慕明,AKA Muming Poo),who also manages CAS-ION.137 Its vice directors are Xu Bo and Du Jiulin(杜久林),a neuroscientist and ION

125、s vice director.Tan Tieniu of CASIA is its chief scientist.CEBSIT researches whole brain connectomics(全腦聯結組),multi-sensory mode perception,computational models for semantic comprehen-sion,and neuron-inspired computing chips.”138 In 2018,the center established a“G60 Brain Intelligence Innovation Park

126、”(G60腦智科創基地)under Shanghai Center for Security and Emerging Technology14municipal auspices139 with a USD$1.5 billion budget for BI-AI.140 The facility uses cloned monkeys to eliminate variables between specimens.141 Pu Muming has been running connectome projects here since the turn of the century.14

127、26.Fudan Universitys Institute of Science and Technology for Brain-Inspired Intel-ligence(復旦大學類腦智能科學與技術研究院,ISTBI),launched in 2015 in Shang-hai,hosts“centers”for cognitive neuroscience,computational biology,big data biomedical science,biomedical imaging,neural and intelligent engineering,and brain-i

128、nspired chips.143 Run by Feng Jianfeng(馮建峰),144 it boasts the worlds larg-est brain science database,with access to the U.S.Human Connectome Project,the UKs Biobank,ISTBIs own 10 terabyte holdings,and the“largest number of magnetic resonance imaging devices in Asia”paired with AI algorithms to scree

129、n the images.145 ISTBIs Zhangjiang International Brain Imaging Center(張江國際腦影像中心)reportedly“is building the worlds most advanced and Asias largest ultra-high-end scientific research magnetic resonance system.”The center claims to have built“internationally leading intelligent algorithms and spatio-te

130、mporal data analysis and processing software”and“the worlds largest full-dimensional brain database and algorithm center.”146 7.Shanghai Jiao Tong Universitys Center for Brain-like Computing and Machine Intelligence(上海交通大學仿腦計算與機器智能研究中心,BCMI)was founded in 2002.Its mission,described in 2013,is“to und

131、erstand the mechanism of intelli-gent information processing and cognitive process in the brain”with research in computer vision,NLP,cognitive computing,BCIs,and electroencephalography(EEG)signal processing.147 A 2021 skill search by the center listed scene percep-tion and understanding,commonsense

132、learning,multimodal interaction learning,selective attention,and visual causal reasoningall indicative of AIs evolution at BCMI and in China.148 The centers Lu Baoliang(呂寶糧,E.6 below)and Zhang Yaqian(張亞倩)have published on“affective BCI”that“can recognize and mod-ulate human emotion”and its role in g

133、eneral AIa technology that has raised ethical concerns in some circles.149 The center is complemented by SJTUs Artificial Intelligence Research Institute,Machine Cognitive Computing Research Center (人工智能研究院,機器認知計算研究中心).150 8.The Shanghai Center for Brain Science and Brain-inspired Intelligence(上海腦科學

134、與類腦研究中心,BSBII)or“Shanghai Brain/AI Center”was set up by CAS in 2018,151 the same year its Beijing-based namesake,CIBR(see above),was established.BSBII is a coordinating center for Yangtze Delta BI-AI research and is linked with CEBSIT,the Fudan and Shanghai Jiao Tong University institutes,the Center

135、 for Security and Emerging Technology15HUST-Suzhou Institute for Brainsmatics,and other regional players.152 Organiza-tionally,it is part of the Zhangjiang Labs Institute of Brain-Intelligence Technology(張江實驗室腦與智能科技研究院,BIT),stood up a year earlier in Pudongs Zhangji-ang Science City and run by Zhang

136、 Xu(張旭).153 BSBII itself is part of Pu Mumings local empire;Feng Jianfeng and Du Jiulin are vice directors.154 Both BIT and the BSBII center collaborate with iFlytek and other Chinese AI companies.155 BSBII operates BCI and“brain atlas big data”platforms.Research includes macro-and mesoscopic connec

137、tomics,BI-AI,and brain-inspired computing devices.156 9.The National Engineering Laboratory for Brain-inspired Intelligence Technol-ogy and Application(類腦智能技術及應用國家工程實驗室,NEL-BITA)was estab-lished in Hefeis High-tech Zone under the University of Science and Technologys(中國科學技術大學)auspices in May 2017.15

138、7 A member of the AI Industry Technol-ogy Innovation Strategic Alliance,158 its research priorities are brain cognition and neural computing,brain-inspired multimodal sensing,brain-inspired chips,quan-tum AI,and brain-inspired intelligent robots.159 NEL-BITAs director is Wu Feng (吳楓),former chief re

139、searcher at Microsoft Research Asia and assistant dean of USTC.Zha Zhengjun(查正軍)is its executive director;Sun Xiaoyan(孫曉艷),an-other MSRA alumna,is deputy director.Although Pu Muming is on its board,none of the other 23 board members or its top research staff are among those persons mentioned above.I

140、n September 2017,the lab spun off a commercial venture to integrate its AI technologies in computer vision,“small sample learning”(小樣本學習),and cross-media multimodal analysis with industry.160 10.The HUST-Suzhou Institute for Brainsmatics(華中科技大學蘇州腦空間信息研究院)was established in 2016 at Huazhong Universit

141、y of Science and Technology(in Wuhan).Its director is Li Pengcheng(李鵬程)of HUSTs National Laboratory for Optoelectronics(武漢光電國家實驗室).161 Deputy directors are Gong Hui(龔輝)and Li Anan(李安安).Notable researchers are Ye Chaohui(葉朝輝)and Luo Qingming(駱清銘),both CAS academicians.Ye was director of CASs Wuhan br

142、anch and has held high posts at Wuhan institutes.162 Luo founded Brainsmatics and is president of Hainan University.163 The institute uses micro-optical sectioning tomography to model a high-resolution mammalian brain.164 Based on“structural and functional imaging of neuron types,neural circuits and

143、 networks,neural-gli-al interfaces,vascular networks,etc.with high temporal-spatial resolution and specific spatial locations,”Brainsmatics is working to decipher brain function and“promote brain-inspired artificial intelligence by extracting cross-level and multi-scale temporal-spatial characterist

144、ics”of brain connectivity.165 Center for Security and Emerging Technology16E.BRAIN-COMPUTER INTERFACES AND NEUROMORPHIC CHIPS66 Brain-computer interfaces use AI to improve their operation,while opening a path to cognitive enhancement.Neuromorphic chips that imitate brain structure prom-ise faster pr

145、ocessing speeds for algorithms that support general intelligence and are being adapted in China for use in BCIs.Examples of both are provided.1.Tsinghua University BCI Lab(清華大學腦機接口研究組)was established in 2004 and is one of several interlocking labs and institutes operated by the university,in-cluding

146、 also Tsinghua University Institute for Brain and Cognitive Sciences(清華大學腦與認知科學研究院,THUIBCS),run by Dai Qionghai(戴瓊海),167 and Tsinghua Laboratory of Brain and Intelligence(清華大學腦與智能實驗室),whose research includes computational neuroscience and BI-AI.168 Liu Jia(劉嘉,D.1 above),Zhu Jun(朱軍),and Gao Xiaorong(

147、高小榕)are prominent members.169 Gao runs the BCI Lab with colleague Gao Shangkai(高上凱)and is its leading figure.170 The lab studies the application of BCI to cognitive skill assessment,which besides med-BOX 2Major Chinese entities pursuing general AI via brain-inspired research Beijing Academy of Artif

148、icial Intelligence BNUs State Key Laboratory of Cognitive Neuroscience and Learning Chinese Institute for Brain Re-search State Key Laboratory of Brain&Cognitive Science CASs Center for Excellence in Brain Science and Intelligence Technology Fudan Universitys Institute of Science and Technology for

149、Brain-Inspired Intelligence Shanghai Jiao Tong Universitys Center for Brain-like Comput-ing and Machine Intelligence Shanghai Center for Brain Science and Brain-inspired Intelligence National Engineering Lab-oratory for Brain-inspired Intelligence Technology and Application HUST-Suzhou Institute for

150、 BrainsmaticsCenter for Security and Emerging Technology17icine can also be used in lie detection and“human-machine collaboration.”171 It is also reportedly pushing the frontier on high-throughput interfaces,including wireless BCI transmission.172 The lab was the first to implement non-invasive BCI

151、technology based on steady-state visual-evoked potential(SSVEP),used to study the relationship between physical stimuli and human cognition.1732.The Center for Brain-Inspired Computing Research(腦計算研究中心)was es-tablished by Tsinghua University in 2014 to study neural functional/computational theory,ma

152、chine learning,and chip architecture.174 The center draws on faculty from seven departments in“brain science,electronics engineering,microelectronics,computer science,automation,materials science,and precision instruments.”175 In 2019,a team led by Center Director Shi Luping(施路平)announced a brain-in

153、-spired computing chip called“Tianjic”(天機 Chip)and software tool chain able to“simultaneously support the neural network models of computer science and neuroscience,”176 thus providing a platform for AGI.177 The chip is partly analog and partly digital and meant to mimic the computational principles

154、 of biological brains.A year later they built a Turing-complete software model to bridge the di-vide between traditional“computer-science-based artificial neural networks”and neuroscience-driven AI models.178 Co-developer Pei Jing(裴京)and Shi Luping market the chip through Beijing Lingxi Technology C

155、o.(靈汐科技).179 3.NeuraMatrix(寧矩科技)was incubated by Tsinghua University in 2019 to build“active implantable systems interfacing with the human body and artificial devic-es.”180 As its name suggests,the project draws inspiration from the U.S.company Neuralink.181 China has decades of experience developi

156、ng non-invasive BCI sys-tems,but this is the countrys first effort to create an implantable device.182 More-over,unlike other Chinese BCI projects whose spoken goal is to alleviate disabil-ity,NeuraMatrix states openly its aim to augment the cognitive power of healthy persons“by effectively merging

157、human and artificial intelligence,”183 fulfilling the 2017“New Generation AI Development Plan”to effect a merger,misinterpreted by some as metaphor.184 The full-service package will include electrode materials,a neural interface chip,“infinite multi-point interface equipment,a signal acquisi-tion an

158、d analysis platform and system-level brain-computer interface platform.”185 NeuraMatrixs founders are Bai Shuo(白碩,CAS)and Zhang Milin(張沕琳,Tsing-huas Laboratory of Brain and Intelligence,E.1 above).186 4.Tianjin Universitys Brain Science and Brain-like Research Center(科學與類腦研究中心)was established in Sep

159、tember 2019.187 It is complemented by the uni-versitys Institute of Medical Engineering and Translational Medicine(醫學工程Center for Security and Emerging Technology18與轉化醫學研究院)and a Neural Engineering Center(天津神經工程中心)focused on brain cognition,medicine,and BCI.188 All three units are directed by Profes

160、sor Ming Dong(明東).189 In July 2019,prior to the centers establishment,the universi-ty unveiled its“Brain Talker”(腦語者)chip able to separate signal from noise with great accuracy.190 The chip will“replace traditional computer devices used in BCI”thanks to its portability,greater precision,and faster t

161、ransmission rates.191 This non-invasive system was developed jointly with the state-owned China Electron-ics Corp.(中國電子信息產業集團).192 Xu Minpeng(許敏鵬),assistant director of the Brain Science Center and project lead,acknowledges Chinas gap in invasive BCI but claims its non-invasive technology is“world c

162、lass.”A follow up chip re-portedly three times faster than competing systems is under development.193 5.Fudan Universitys Institute of Brain-inspired Circuits and Systems(類腦芯片與片上智能系統研究院)was set up in July 2017 to support the China Brain Project and Shanghais Zhangjiang National Lab(張江國家實驗室)with rese

163、arch on BI-AI chips and neuron signal acquisition.194 It is led by Min Hao(閔昊),an expert on non-volatile memory and wireless chip design.195 In 2021,the institute announced Chinas first wireless BCI circuit for transmitting information between a chip and nerve cells.Team leader Ye Dawei(葉大蔚)claims i

164、t“outperforms foreign versions on many levels”at half the cost.196 The Ministry of Educations Key Laboratory of Brain Functional Genomics(腦功能基因組學教育部重點實驗室)at Huadong Nor-mal University supported its development.197 The device is currently installed on the skulls of freely moving mice.Its technical as

165、pects are described in a paper by eight of the institutes researchers,198 who also claim affiliations with the State Key Lab of ASIC and Systems(專用集成電路與系統國家重點實驗室),which has been operating at Fudan University since 1995.199 6.Shanghai Jiao Tong Universitys Ruijin Hospital BCI and Neuromodulation Cent

166、er(瑞金醫院腦機接口及神經調控中心)was founded in 2020 with a goal of using BCI to address depression and other types of neuropsychological illness-esBCIs“affective”applications.200 The center is co-directed by Sun Bomin(孫伯民)and Lu Baoliang,the latter co-posted to the Center for Brain-like Computing and Machine Int

167、elligence(D.7 above).According to Professor Sun,the plan is to“implant chips into patients brains via a minimally invasive surgery”followed by electrical stimulation based on AI analysis.201 Meanwhile,Lu in December 2021 spun off a BCI enterprise in partnership with Chinese video game giant miHoYo (

168、米哈游)called Lngwiys(零唯一思,no English name),202 whichcuriously focuses on medical research(including BCI)and game development.203 MiHoYo,for its part,in March that year agreed to build a laboratory with Ruijin Hospitals Center for Security and Emerging Technology19Encephalopathy Center(腦病中心)run jointly

169、 by Lu and miHoYos Anti-Entropy Studio(逆熵工作室)to develop brain interface technology.204 7.Zhejiang Universitys Frontier Science Center for Brain and Brain-Machine Inte-gration(腦與腦機融合前沿科學中心),also called the Double Brain Center (雙腦中心),was established under the Ministry of Educations auspices in October

170、 2018.Wu Zhaohui(吳朝暉)is director and Duan Shumin(段樹民)is its chief scien-tist.205 Duan doubles as dean of Zhejiangs Universitys School of Brain Science and Brain Medicine(腦科學與腦醫學學院),built in 2019 to research hybrid intelligence,BCI,and brain-inspired computing.206 There is also a Zhejiang Laboratory(

171、之江實驗室),stood up in 2017 under provincial Communist Party auspices to guide develop-ment of a“national strategic scientific and technology force”in AI areas.207 Zheji-ang Universitys research in BCI dates from 2012.It has also achieved a number of BI-AI“firsts,”including the worlds largest neuromorph

172、ic computer in 2020 called the“Darwin Mouse.”208 The device,built by Zhejiang Lab,contains 792“Darwin II”BI chips that can emulate around 120 million spiking neurons and 100 billion synapsesabout the same as a mouseon just 350500 watts.209 8.South China University of Technologys Center for Brain Com

173、puter Interfaces and Brain Information Processing(腦機接口與腦信息處理中心)was established in 2007 in Guangzhou to research brain-computer interaction and large-scale brain data.210 It is directed by Li Yuanqing(李遠清),21 1 who also runs the universi-tys School of Automation Science and Engineering(自動化科學與工程學院).Th

174、e center has five research teams whose topics collectively cover EEG and fMRI signal analysis,sparse signal representation,pattern recognition,ML,neural networks,big data processing,robotics,and,of course,BCI.Li acknowledges that in China BCI research“is still mainly aimed at normal people(正常人為主).”2

175、12 In March 2019,the center formed a joint venture with China AI giant iFlytek(科大訊飛)called South China Brain-computer Interface Technology(華南腦控智能科技,iHNNC),213 founded by Li and managed by CEO Xiao Jing(肖景).214 The compa-ny sees BCI as empowering“the fields of elderly care and disability,mental and s

176、piritual health,education,entertainment,security,military and other fields.”215 9.Tianqiao and Chrissy Chen Institute(陳天橋雒芊芊研究院,TCCI)was founded at Caltech in 2016 by online gaming pioneer Chen Tianqiao(陳天橋)and spouse Chen Qianqian(陳芊芊).In 2020 and 2021,the institute established two“Frontier Labs”(前

177、沿實驗室)for brain research,one at Huashan Hospital(華山醫院)in Shanghai,the other at the Shanghai Mental Health Center(上海精神衛生中心)whose vision statement reads:“To enrich peoples lives with brain-and mind-related Center for Security and Emerging Technology20technology.”216 The institute intends to extend deve

178、lopment of its invasive BCI products beyond therapeutics to cognitive augmentation217 and hopes eventually to address scientifically every important mind-body problem in the history of philosophy,from the cognitive basis for beliefs through free will,emotion,AGI,and cognitive uploads aimed at immort

179、ality.218 Tao Hu(陶虎),who runs the Huashan facility,in late 2021 founded his own BCI company NeuroXess(腦虎科技)to build invasive BCI that can“twin”(孿生)human and artificial intelligence.219 Tao is joined by billionaire Peng Lei (彭蕾,Lucy Peng),one of Alibabas co-founders.220 10.CAS Institute of Automation

180、(中科院自動化研究所,CASIA)besides traditional and BI-AI research is also invested in BCI development,and its importance to Chinas Brain-AI program warrants this double listing.CASIAs work on coding and decoding of visual neural information,see as“the core technology of brain-computer inter-faces,”221 is rega

181、rded as“an important stepping stone in the work to create better brain-machine interfaces.”222 CASIA researcher Yu Shan(余山),who is also deputy director of Chinas State Key Laboratory of Pattern Recognition(模式識別國家重點實驗室副主任)with specialties in brain information processing,BI-AI,and BCI,223 sees BCI“ult

182、imately enhancing and expanding brain functions,”which puts CASIA into the augmentation camp.224 A 2021 report titled“BAAI AI Frontiers”listed among CASIAs achievements a robotic system that can“accurately implant flexible electrodes into the cerebral cortex of animals under the guidance of microsco

183、pic images,”thus laying a path toward invasive BCIs.225 BOX 3Major Chinese entities pursuing general AI via brain-computerinterfaces or neuromorphic chips Tsinghua University BCI Lab Tsinghua Universitys Center for Brain-Inspired Computing Research NeuraMatrix Tianjin Universitys Brain Science and B

184、rain-like Research Center Fudan Universitys Institute of Brain-inspired Circuits and Systems Shanghai Jiao Tong Universitys Ruijin Hospital BCI and Neuromodulation Center Zhejiang Universitys Frontier Science Center for Brain and Brain-Machine Integration South China University of Technologys Center

185、 for Brain Computer Interfaces and Brain Information Processing Tianqiao and Chrissy Chen Institute CAS Institute of AutomationCenter for Security and Emerging Technology21Monitoring for Safety and Security3his section makes a case for monitoring Chinas AI development as a bellwether for AI risk,und

186、erstood in terms of global safety and U.S.national security(both 安全 in Chinese).226 The authors efforts to build a relational database to support discovery and implement a scal-able indications and warnings(I&W)watchboard are described.F.WHY BUILD A CHINA AI“WATCHBOARD?”In terms of global AI safety,

187、China is one of many countriesthe United States includedable to create advanced forms of artificial intelligence that pose unknown and potentially catastrophic risks.From a national security standpoint,China is no more or less likely than a half dozen other states to inflict harm on the United State

188、s and its allies through advanced AI when motivation is considered.So why focus on China?There are multiple reasons why U.S.policymakers may want to sup-port a rigorous monitoring regime able to track Chinas advances toward general AI:China intends to lead the world in AI and achieve a“first mover”a

189、dvantage.Chinas“New Generation AI Development Plan,”released in 2017,and other official proclamations declare Chinas intent to lead the world in artificial intelligence by 2030 and achieve a first mover advantage(先發優勢)through general AI.227 These goals,if realized,have serious impli-cations for U.S.

190、security,since early success at building advanced AI portends not only commercial and strategic advantages but potentially an asymptotic accretion of cognitive resources(a so-called“intelligence TCenter for Security and Emerging Technology22explosion”)widely viewed as an existential risk and game-ch

191、anger vis-vis competing nations.228 Data-based monitoring will guard against hype and overreaction to imag-ined threats.Fair-minded patriots note these proclamations,take China at its word,imbibe fear-mongering hype from pundits misinformed by the same sourcesand sup-port a crash program that leads

192、to a dangerous and unnecessary“AI arms race.”This scenario is not without precedent(the 1950-60s alleged U.S.“missile gap”with the Soviet Union)and,in the AI realm,has nearly happened at least once already.229 Accurate and timely accounting of Chinas progress toward advanced AI will reduce the likel

193、ihood of miscalculation.Chinas ownership and export of general AI will exacerbate existing politi-cal abuses.Risks from weaponized or flawed AI is not the only cause for concern.Equally worrisome is Chinas use of AI for political oppression and its ability to export this technology through its marke

194、t power.The problem is documented by CSETs Dahlia Peterson230 and by Maya Wang at Human Rights Watch.231 Also,credible cognitive neuroscientists raise the specter of AI-based technologies facilitating mind control in service of totalitarian goalsnot as a metaphor(“influence opera-tions”)but literall

195、y at the neuron level.232 Preventing this implies an ability to track its development.China has a poor track record in safety and post-disaster communication.It is hard to ignore a near-existential calamity that originated in China.Wherever one puts the locus of the COVID-19 pandemic,two datapoints

196、emerge:a massive scourge began in China,despite safeguards,and China responded lethargical-ly to global calls for information.Can one assume an AI disaster will be treated differently?There is little chance China will allow direct safety inspections of its AI infrastructure or eschew technology that

197、 leads to national advantage,hence the need for a comprehensive watchboard able to track unsafe developments.Data-driven knowledge leads to opportunities for collaboration on safety.Although one does not usually think in these terms,there is a need for accurate and timely data to alert U.S.and allie

198、d policymakers for opportunities to engage China on AI matters,not only as they involve safety,but also in areas where both countries and the world stand to benefit,such as AI algorithms that can pinpoint inputs to climate change,promote species preservation,and so on.Spotting these opportunities re

199、quires timely monitoring and communication,which is not fully possible under present circumstances.Center for Security and Emerging Technology23China is already systematically monitoring U.S.AI developments.Watchboards can be viewed as unfriendly acts,akin to espionage when conduct-ed by governments

200、,and serve as grist for United Front propaganda that portrays China as a“victim”of racism or foreign aggression.In fact,China has operated an open-source science and technology intelligence(科技情報,STI)network since 1958 that is mainly U.S.focused,outstrips U.S.efforts by two or three orders of magnitu

201、de,and is immeasurably more effective.233 The present project does not redress the imbalance but may lead to the U.S.government accepting a serious and badly needed open-source-monitoring solution.Monitoring will alert U.S.authorities to technology transfers not in Americas interest.One argument for

202、 expanded open-source monitoring,proposed in various forms and forums,234 is that it is impossible to dissociate Chinas indigenous research(its own business)from predatory technology transfers(the worlds business).The same applies hereany assessment of Chinas AI progress must include an accounting o

203、f what it obtains from abroad.Transactions captured by the watchboard inimical to U.S.national and economic security can and should be passed to cognizant authorities for disposition under legal statutes.Sharing early knowledge of Chinas AI research supports U.S.competitiveness.Chinas STI-monitoring

204、 network operates less as an early-warning system and more as a siphon for transfer opportunities.By contrast,the U.S.government has long refused to share foreign intelligence with private enterprise,even where such knowledge would accrue to U.S.national advantage,in part because there is no fair wa

205、y to apportion the benefit between companies,and in part because the United States could get by without doing it.That situation no longer obtains.China is a peer competitor in many advanced technologies and it may be time to aban-don the hubris that puts U.S.companies at a disadvantage.There is bipa

206、rtisan will to meet Chinas technological challenge and support remedies.As a practical matter,given the U.S.intelligence communitys reluctance to build robust open sourcebased foreign technology detection and analysis programs,235 the only economically viable way to build a risk and threat detec-tio

207、n mechanism is to start“small”(one country),test methodologies,establish a reputation and user base,and thereby attract additional support to expand in increments to a full-blown operation that monitors all state and non-state actors advanced AI developments.In todays milieu,China topics draw the mo

208、st atten-tion and hence are more likely to be funded.Center for Security and Emerging Technology24Chinas development of general AI carries potential existential risk.Students of catastrophic risk agree that even if the chance of“runaway AI”(or a meteor strike,or fatal pathogen)is low,the potential c

209、onsequences may be grave enough to justify strong efforts at mitigation.236 Similarly,if the risk of Chinaor anyone elsecreating or stumbling upon an asymmetrical advantage in AI is small,and confidence in safeguards is high,the results of error,misplaced trust,and bad judgment may be serious enough

210、 to warrant scrutinizing developments in advance.An additional consideration is the technical requirements of open-source ex-ploitation.Tracking Chinas S&T developmentin AI or any fieldcannot be done without Chinese-language sources.To believe otherwise is to imbibe yet another China myth.Indeed,mos

211、t of the“good”items are buried in what frustrated analysts call the“soft encryption”of Chinas logographic script.Working with the native language and orthography presents challenges,which,if solved,take one most of the way to a universal data processing capability.Adding other languages and countrie

212、s is straightforward by comparison.G.CREATING A DATA MODEL OF THE PROBLEMBuilding a scalable indications and warnings(I&W)watchboard of foreign S&T developments begins by populating a relational database of entities,attributes,and connectionsin essence,a software model of the target.The steps includ

213、e defining search objects,identifying sources and arranging(continuous)access,compiling an initial dataset,expanding it,expressing different data types in a common format,conditioning data to eliminate errors and variants(entity resolu-tion),purging extraneous data,building a user interface,“tuning”

214、the system to flag data that meet thresholds,establishing routines for supervision and service,and building a customer base and alerts protocol.This sequence is an abstraction.In real-world settings,the steps overlap.Best practices use a combination of humans and machines,one supporting the other,ac

215、cording to needs that develop in practice and change constantly.The notion that an operation like this can be executed from a priori assumptions on the basis of“data science”is a fantasy engaged in by people who never struggled with“dirty”data mined from strategic competitors who speak different lan

216、guages and whose interests usually conflict with those building the database.Taking this principle into account,this papers authors compiled a list of source and method types based on prior success at eliciting Chinese-language data on advanced technologies237 and AI in particular.238 They include a

217、 dozen“general”types accessible to most linguist-researchers that were used to build the starter list of Center for Security and Emerging Technology25entities in this papers Section 2 and another half dozen“advanced”types that are more resource intensive.239Each source type is serviced by a combinat

218、ion of human and artificial agents.In addition,information from the following sources was added to the starter list:Chinese presenters at international and China-based“AGI”conferences from 2009 to present with affiliations and biographic data.Chinese scientists working in any of Chinas BI-AI institu

219、tes from internet searches and confirmed by email address domains.240 metadata on authors,affiliations,and research from limited keyword-based queries of Chinese Academic Journals(CAJ).241 Data from these four sourcesthe starter list and three accretionswas entered into Excel spreadsheets formatted

220、to populate a relational database of organiza-tions and persons(chiefly scientists and science managers).242 This does not make up an I&W watchboard.It allows(1)the accumulation of topical information,(2)access to conditioned(clean,searchable,reliable)data,and(3)building of net-work diagrams based o

221、n location,co-authorship,and citation patterns that support discovery and analysis.There is a learning curve in database constructiona tension between the needs of the analyst and the software engineer,best adjudicated when the volume of data is sufficient to bring most issues to light(e.g.,how many

222、 categories of what type are needed)but before the data grows to a point where major changes to the data structure are unfeasible.These issues are being sorted out in preparation for the next stage of data acquisition,which involves full-scale exploitation of some,but not all,of the source types ide

223、ntified,namely:expanded keyword searches of the CAJ(CNKI)corpus.internet and news searches of key elements already identified.website exploitation(scrapes)of entities surfaced in the initial data pulls.While these tasks can be performed by human operators with requisite lan-guage skills,all can be a

224、ugmented by automated systems to a greater or lesser de-gree.Accordingly,the authors defined and began implementing a monitoring tool specifically tailored to detect and highlight any mention of certain key technologies,persons,or organizations appearing in online news streams.This tool will serve a

225、s an alarm and triaging mechanism for human researchers to create analytic products and identify newly emerging trends as soon as they appear.Center for Security and Emerging Technology26Managing the collected data presents its own problems.The authors extracted,transformed,and loaded(ETL)listed dat

226、a sources into machine-readable structured datasetsa non-trivial taskand are designing and optimizing full-text search criteria to match keywords,persons,and organizations for use in automated search-es against our full-text data collections.Finally,they are collating and curating matched results in

227、to analysis-friendly formats for tabular presentation or visualiza-tions.Automated support for data conditioning and entry is also a priority in this pilot project.For certain structured data,we have begun scraping and processing large volumes of academic publication metadata and academic and commer

228、cial con-ferences and expositions for reference in a single user interface.CSET researchers have also designed data-entry templates that facilitate manual curation of authors and organizations metadata meant to capture their key aspects.Populated tem-plates go through a quality control process for e

229、ntity resolution and system-wide accuracy.H.WATCHBOARDING AND MONITORINGSubsection G(above)describes a database to support I&W on a particular prob-lemChinas development of advanced AI.Creation of a database mirrors work done by the authors elsewhere on similar Asian-language datasets,so success was

230、 expected.A database,however,is not a“watchboard,”which is uncharted territory.While indicators exist for WMD and other technologies,there is no ready-made compilation for“general AI”because there is no consensus on the goal and no precedents to provide guidelines.Also“watchboarding”differs from“mon

231、itoring”in that the former implies progress toward a defined end,while the latter is largely“observation without expectation.”The authors present goal straddles the two:Chi-nas progress toward advanced AI by one of three paths is a“well-defined forecast-ing target”but hardly qualifies as“objectively

232、 and unambiguously evaluable.”243 The aim at this stage,accordingly,is not to elicit quantitative measures of prog-ressalthough that is likely to changebut to determine who in China is taking what steps toward general AI,as shown by overt expressions and other common mea-sures.The AI safety literatu

233、re is replete with hypothetical risks and disaster scenarios,from which indicators can be inferred.There is also copious literature on AGI and its presumed requirements.Other indicators are generic and apply to any advancing technology enterprise.The tasks come down to:(1)compiling indicators of pro

234、gress toward general AI and(2)applying them in some systematic,sustainable fashion to the sources avail-Center for Security and Emerging Technology27able.Task 1 is fulfilled by three lists of indicators and topical keywords generated from the authors AI and AGI readings,including“indicators”of a gen

235、eral nature,potentially problematic research areas,and a bilingual keyword list(see Appendix).The first two lists orient analysts toward areas that merit attention and serve as a basis for generating topical keywords.These keywords are proxies for the target and are run periodically against three da

236、ta stores,namely,the projects relational database,large proprietary data holdingsin this case,the CSET merged corpus of scholarly literature including Digital Science Dimensions,Clarivates Web of Sci-ence,Microsoft Academic Graph,China National Knowledge Infrastructure,arXiv,and Papers With Code244a

237、nd the internet itself via purpose-built crawlers and newswire databases.The output is local alerts.Associated data may or may not be added to the database depending on what is retrieved,since the goal is to model the target,not replicate it.For select internet searches,we first compile a list of qu

238、ery terms designed to optimize the retrieval of relevant search results.We implement heuristics for result relevancy,and then we query via API calls to major internet search engines.245 Finally,we deduplicate and aggregate relevant search results in a human-friendly format for further triaging of an

239、alysis-relevant results.For news monitoring,we compile a list of query terms designed to optimize the retrieval of relevant news article mentions.We collect Google News RSS and LexisNexis feeds to build custom newswire databases for each topic.This allows us to plot the number of news mentions of to

240、pic keywords over time.The three lists are fluid documents adjusted for their ability to elicit material and our evolving understanding of the target.We invite expert participation in their growth and development.We are also experimenting with algorithms that can nom-inate new search terms from cont

241、ext.The aim is an optimized keyword list able to absorb sophisticated metrics,including genuine numerical benchmarks.To stay relevant in a shifting technological landscape,we propose using intermittent surveys of AI experts from the academic community to inform algorithms that weigh how information

242、and data are displayed in the monitoring dashboard tool.Term weighting will help ensure significant articles and news developments are put front and center for human dashboard monitors.As a practical matter,bounds for the keyword list will fluctuate between 100 and 200 terms,including names of selec

243、ted scientists,by analogy with the Swadesh lists used to isolate key concepts from open-ended vocabularies.246Center for Security and Emerging Technology26Center for Security and Emerging Technology29The China AI Monitoring System in Operation4his pilot project to monitor Chinas progress toward adva

244、nced artificial intelligence will be fully operational when coupled with stable ownership and resourcing,expert benchmarking,and mature processing algorithms.Meanwhile,this section offers two examplesa success and a failureof the authors early efforts to monitor indications of advanced Chinese AI pr

245、ograms,along with a compilation of retrieved statements on“general AIs”trajectory from Chinese experts identified by the project.I.CASE STUDIES IN MONITORING AND DISCOVERYXiamen University and AGISeth Baum,in his 2017 survey of worldwide AGI programs,de-listed a program at Chinas Xiamen University(廈

246、門大學)run by AGI pioneer Hugo de Garis from his inventory.247 The university did not reappear in Baums updated survey a few years later,248 but remained of interest because of the prior activity.Accordingly,the authors ran searches in 2021,both in the aforemen-tioned database and online,aided by knowl

247、edge of Chinas AI programs accumulated during the general study.The searches revealed that the uni-versity had stood up a reinvigorated“Institute of Artificial Intelligence”(人工智能研究院),with an unusually high number of tenure(track)positions among its“AI-related faculty.”249 Other datapoints hinted at

248、the universitys intent to work toward AGI:TCenter for Security and Emerging Technology30 The AI institute researches machine vision,pattern recognition,and cognitive science,and has teams studying hybrid intelligence and cross-media AIall AGI precursors.250 The institute has cooperative research rel

249、ationships with all four of Chinas“BATH”(Baidu,Alibaba,TenCent,Huawei)companies,each of which has AGI research programs.251 AI institute professor Ji Rongrong(紀榮嶸)lamented the inability of AI to deal with the“catastrophic forgetting”problem.Solving it is the“key to AGI”(通用人工智能).252 Xiamen University

250、s AI institute and Tencent,a known AGI developer,253 announced jointly that“multi-modal fusion,”in which the institute specializ-es,will lead to AGI.254 Ben Goertzel is a visiting professor at Xiamen Universitys Fujian Provincial Key Laboratory of Brain-like Intelligent Systems(福建省仿腦智能系統重點實驗室).255 A

251、 high-profile paper by Xiamen University researchers on“Bioinspired Nanofluidic Iontronics”describes brain-like signal processing for neu-ron-computer interfaces.256 Dean of the College of Humanities Zhu Jing(朱菁)and colleagues provided slides at the 2020 China AGI Annual Conference on the definition

252、 of AGI.257 Jiang Min(江敏),professor in the universitys School of Informatics and a senior IEEE member,co-authors with Goertzel and focuses personally on software and AGI.258Other indicators of an AGI program were found in academic publications and collaboration networks,retrieved from the CSET merge

253、d corpus.Center for Security and Emerging Technology31Figure 1 shows the distribution by year of AGI-themed papers produced by Xiamen University affiliates starting in 2008,when de Garis project began,through its termination in 201 1 and on to mid-2021(data cut-off).The blue areas are papers retriev

254、ed from CNKI and English-language databases based on a half-dozen terms aligned with de Garis China-Brain Project;the red areas represent papers retrieved by searches on Chinese and English expressions for“AGI.”In terms of scholarly output,the programs disappearance had a temporary impact only on(ov

255、ert)AGI research at the university or among its affiliates.FIGURE 1Papers by Xiamen University researchers on AGI topicsTABLE 1Papers co-authored by Xiamen University and majorChinese organizations with known AGI research programsSOURCE:CSET MERGED CORPUSSOURCE:CSET MERGED CORPUS2008PAPERS1614121080

256、6“ARTIFICIAL BRAIN”AGI PAPERSAGI PAPERS WITHOUT“ARTIFICIAL BRAIN”TERMS200920102011201220132014201820192020202142YEAR201520162017 Baidu 24 Alibaba 11 Tencent 93 Huawei 46 CASIA 16 ORGANIZATIONPAPERSCenter for Security and Emerging Technology32Table 1 shows Xiamen University co-author relationships wi

257、th affiliates of Chinas big four AI companies and the Chinese Academy of Sciences Institute of Automation,which all have or will have AGI research programs.The papers show the existence of collaboration networks only.Figure 2 is the same information ex-pressed over years(data available to June 2021

258、only).Finally,we doubt that Professor de Garis failed to bequeath a legacy.His“Chi-na-Brain Project”involved 20 people,several still there,259 building their“network of networks”vision of brain function.260 Although we did not turn up a“hot”AGI research program at Xiamen University,this is likely a

259、distinction without a differ-ence,as there is ample evidence this provincial university is moving(back)toward AGI research,if it ever left it.Wuhans AI Research ComplexWhereas the authors were able to document the likely return of AGI research to Xiamen,an event of similaror greaterimport occurring

260、in Wuhan went unno-ticed until recently for lack of a robust monitoring protocol.The Peoples Republic of China has a history dating from the 1950s of shunting sensitive research to the hinterlands.261 This may or may not apply to artificial intelligence,but habits per-sist,so when significant develo

261、pments in AI happen in the provinces,outside the eastern complexes(Beijing,Shanghai),one should take notice.FIGURE 2Xiamen University Collaboration with BATH+CASIASOURCE:CSET MERGED CORPUS2009PAPERS8002010201120122013201420152016201720182019YEAR20202021604020Center for Security and Emerging Technolo

262、gy33For example,Table 2 was compiled from a corpus of Chinese academic journal articles in mid-2021 as a baseline for identifying key Chinese AI institutes.The table ranks PRC organizations by number of CNKI AI publications in 2020;the 2015 rankings are added for comparison.TABLE 2Organizations with

263、 top AI research output in CNKI journalsSOURCE:CSET MERGED CORPUSORGANIZATIONTABLE 2Organizations with top AI research output in CNKI journals2 20 02 20 0 R Ra an nk k 2 20 02 20 0 P Pa ap pe errs s 2 20 01 15 5 R Ra an nk k 2 20 01 15 5 P Pa ap pe errs s O Orrg ga an niiz za attiio on n 1 1316 2 77

264、7 University of the Chinese Academy of Sciences (?)2 1249 1 1037 W Wu uh ha an n U Un niiv ve errs siitty y S Sc ch ho oo oll o of f IIn nf fo orrm ma attiio on n M Ma an na ag ge em me en ntt(?)3 1132 4 713 Tsinghua University Institute of Automation (?)4 928 6 616 Shanghai Jiao Tong University(?)5

265、 889 18 502 Sichuan University College of Computer Science(?)6 795 31 386 PKU School of Electronics Engineering and Computer Science(?)7 770 80 221 Renmin University of China(?)8 770 5 680 Zhejiang University(?)9 762 14 528 Tongji University(?)10 732 15 527 University of Shanghai for Science and Tec

266、hnology(?)11 718 46 311 Beijing Normal University(?)12 715 27 435 Nanjing University(?)13 701 3 748 NUAA(?)14 688 16 522 Jilin University(?)15 686 12 531 Southeast University(?)16 658 10 567 Tianjin University(?)17 632 42 345 W Wu uh ha an n U Un niiv ve errs siitty y o of f T Te ec ch hn no ollo og

267、 gy y(?)18 607 20 492 North China Electric Power University (?)19 605 23 469 Southwest Jiaotong University School of Electrical Engineering(?)20 595 25 459 H Hu ua az zh ho on ng g U Un niiv ve errs siitty y o of f S Sc ciie en nc ce e a an nd d T Te ec ch hn no ollo og gy y(?)2020 PAPERS 2015 RANK2

268、015 PAPERSORGANIZATION2020 RANKCenter for Security and Emerging Technology34The authors,who have followed AI for years,were surprised by Wuhan Uni-versitys School of Information Managements outsized contributionwhich should have,but did not,surface in prior researchcomplemented by two other Wuhan in

269、stitutes in the top 20(Wuhan-based organizations are in bold).262 Wuhans Huazhong University of Science and Technology(number 20)attract-ed our attention earlier when its HUST-Suzhou Institute of Brainsmatics received disproportionately large funding from Chinas National Natural Science Foundation c

270、ompared to other recipients.Table 3 is a schedule of grants by NNSF for AGI-pre-cursor projects,compiled by the authors in 2020 from incomplete data.263 TABLE 3NNSF academic grants for AI-brain projects 201819The grants,small by U.S.standards,are typically matched by local resources or supplemented

271、in other ways,but the main point is that two other NNSF grants were excluded from the tallies because of their unusual size:“Whole-brain mapping system based on morphology and omics spatial information”(基于形態與組學空間信息的細胞分型全腦測繪系統)for USD$10,224,342 awarded to Luo Qingming(駱清銘),264 and“High-resolution op

272、tical imaging and visualization of brain connections in brain spatial information”(腦空間信息中腦連接的高分辨光學成像與可視化研究)for USD$2,800,821 to Li Pengcheng(李鵬程).NNSF Individual Grants 2018 2019 Grants Funding(USD)Grants Funding(USD)BI-AI 31 2,524,531 19 2,002,811 Connectomics 18 3,959,650 11 1,398,454 Brain-comput

273、er Interfaces 20 1,655,108 19 2,202,753 Total(AI-Brain)69 8,139,289 49 5,604,018 Average Grant Amount per Project$117,961.00$114,368.00 NNSF INDIVIDUAL GRANTS20182019FUNDING(USD)GRANTSGRANTSFUNDING(USD)SOURCE: for Security and Emerging Technology35Professors Luo and Li were both at HUST in Wuhan.Luo

274、,founder of its Brains-matics Institute,was also dean of the Wuhan Optoelectronics National Research Center of HUSTechnology,265 engaged mainly in photonics research including“multi-modal molecular imaging”in support of BI-AI and connectomics.266 The import of this funding anomaly did not register w

275、ith us then,either,and we treated it as an outlier.In retrospect,other signs of Wuhans emergence as a potential AGI center were there had we been looking.A professor at Huazhong Normal University(華中師范大學),Liu Kai(劉凱),267 has sponsored China AGI Annual Conferences(中國通用人工智能年會)since 2016,268 a result of

276、 Lius collaboration with U.S.AGI expert Pei Wang.269 The conferences supplement the Artificial General Intelligence Societys series,which has run worldwide since 2008 to showcase local developments,and a disproportionate number of its attendees are from Wuhan institutes.270 The city made AI news in

277、May 2021,which we also missed,when the Wuhan Artificial Intelligence Computing Center(武漢人工智能計算中心)began operation,billed as the start of“a new generation of AI innovation.”271 That month,the CAS In-stitute of Automation(a declared AGI developer),Huawei,and the citys East Lake Hi-tech Development Zone

278、(東湖高新區)announced plans to build a Wuhan Gener-al Artificial Intelligence Platform(武漢通用人工智能平臺),create a“perception-cog-nition-decision whole chain ecology,”and an open platform for“autonomous and controllable general AI”(自主可控通用人工智能).272 This development was impossible to ignore.In July 2021,CASIA ann

279、ounced a“multi-modal general artificial intelligence”project called“Zidong Taichu”(紫東太初)based on Huaweis“Shengteng”(昇騰)platform,aimed at integrating image,text,and audio data to“replicate human semantic processing.”Computing support is via the Wuhan AI Computing Center.273 Wuhans role as a key node(

280、核心節點)and next step(下一步)in AI develop-ment was made explicit,for anyone listening,on December 20,2021 by CASIAs Wang Jinqiao(王金橋),who acknowledged that OpenAIs GPT-3(subsection A above),from which Zidong Taichu draws its inspiration,“opens a new beginning for artificial intelligence from dedicated in

281、telligence to general intelligence”274the whole point of our discovery project.Could we have forecast Wuhans emergence as an advanced AI center earlier,with the proper data,protocols,and incentive?What other potentially transforma-tive developments have we and other observers been missing?Center for

282、 Security and Emerging Technology36 J.CHINESE EXPERT VIEWS ON“GENERAL AI”This subsection assesses Chinas willingness to accept advanced AI,based on ex-pert views discovered while executing this project.Chinese AI scientists,like their counterparts worldwide,are aware of the controversies surrounding

283、 the definition,attainability,and risks of“AGI.”Their views on these matters will impact practical outcomes.First some caveats:The authors acknowledge the larger issue of whether acceptance of advanced AI per se by Chinas governing class,S&T policymakers,AI scientists,and informed public even matter

284、s.An argument can be made that technology follows its own dynamicwhat can be done is done,driven by pressure to accept innovations that offer a higher“utility function”no matter what its long-term outcome may be.“Choice”on a macroscale may be as illusory as the alleged“free will”Chinese AI scientist

285、s(E.9 above)propose to investigate.Technical innovation may be delayed by government,for good reasons(en-vironmental degradation,other existential threats),bad reasons(elite paranoia,societal collapse),or in the case of the United States,for no reason(hubris,compla-cency).In an authoritarian country

286、 especially,the speed at which events happen is determined to a greater degree by the attitude of its governing body.We will argue that Chinas political elite not only support advanced AI but see it as in their class interest to continue doing so.Finally,acceptance itself becomes a self-fulfilling p

287、rophecy.Declaring a goal of world domination by 2030 may spur Chinas AI scientists to act in ways that create the desired end,for example,by ignoring safety concerns that hobble other nations or simply by believing in a given outcome.As Goertzel put it,China“has no chip on its shoulder about AI or A

288、GIit has the same status as any other advanced technology.”275 So Zhu Songchun,Nikos Logothetis,and others gravitate there and help drive Chinas success.What do Chinese scientists think of general AIs prospect and how comfort-able are they with it?The authors cite two surveys,bearing in mind that th

289、e focus is“advanced AI”with general capabilities,not the human level AGI addressed by the surveys,which may(or may not)differ.A paper by Grace et al.in 2018 reports a survey of international researchers who published at two top machine learning conferences in 2015.276 The forecasts per se(guesswork

290、in any case)are less interesting than the differential findings for Asian and,especially,Chinese respondents,who are more optimistic about AGIs emergence:“(R)espondents from different regions had striking differences in HLMI hu-man-level machine intelligence predictions.Fig.3 shows an aggregate pre-

291、diction for HLMI of 30 years for Asian respondents and 74 years for North Center for Security and Emerging Technology37Americans.Fig.B.1 displays a similar gap between the two countries with the most respondents in the survey:China(median 28 years)and USA(median 76 years).”277 Mindful that beliefs m

292、atter,the present authors surveyed two groups of Chinese AI researchers in 2020 on issues related primarily to BI-AI.A majority(74 percent)stated that brain-inspired AI will lead to AGI(83 percent of the BI-AI specialists).AI generalists were split 50/50 on whether AGI would be achieved(through any

293、means)in 510 years or more than 10 years,while BI-AI specialists were more aligned that AGI is more than 10 years away.278 While this data is indeterminate,the probability that Chinese scientists imagine and contemplate general AIs appearance at a relatively early date seems hard to refute.Some of C

294、hinas leading AI scientists lend credence to the argument that Chinas research establishment both welcomes and is working toward advanced(general)AI:Dong Le(董樂),deputy director,Beijing Institute for General Artificial Intelli-gence,at the BEYOND International Technology Innovation Expo in Macao,Dece

295、mber 2021.The trend of AI is from perceptual to cognitive intelligence.We are following a pathway in the“evolution of general artificial intelligence”with the poten-tial to“become the most reliable and important aid in the human-machine merger of the future.”279 Gao Wen(高文),dean of Peking University

296、s School of Information Science and Technology and director of Pengcheng Lab(C.6 above).AI is at a critical stage in the development of a new generation of artificial intelligence to strong artificial intelligence.By 2030,Chinas artificial intelli-gence will reach the worlds leading level.280 Huang

297、Tiejun(黃鐵軍),vice dean of Peking Universitys Institute for Artificial Intelligence and an outspoken advocate of ASI.Artificial superintelligence is the futurean“evolutionary trend.”Technologi-cal innovation and development should not be restricted by the limitations of human beings.We can treat the d

298、evelopment of artificial intelligence with a more open mind.281 Center for Security and Emerging Technology38Ming Dong(明東),Tianjin University professor and director of its Brain Science and Brain-like Research Center(E.4 above).“The future is not about replacing human beings with artificial intellig

299、ence,but making AI a part of human beings through interconnection and interoper-ability.A blend of human and computer without barriers is the inevitable end of the future.”282 Pan Gang(潘綱),Zhejiang University professor of computer science,AI,computer vision,and BCI expert.BCI will support“a new type

300、 of intelligencebrain-computer hybrid intelli-gence”(腦機混合智能)able to effect comprehensive enhancement of ones perceptual,cognitive,and behavioral ability,a“mixture of biological and machine brain.”283 Shi Luping(施路平),director of the Center for Brain-inspired Computing Research,Tsinghua University and

301、 head of the Tianjic team.“Nanodevices have enabled us to develop electronic devices such as neu-rons and synapses at the level of human brain energy consumption,so now is the best time to develop general artificial intelligence.”284 Tan Tieniu(譚鐵牛),former deputy director of the Chinese Academy of S

302、ciences and deputy chief of the PRCs liaison office in Hong Kong.“How to make the leap from narrow artificial intelligence to general artificial intelligence is the inevitable trend in the development of the next generation of artificial intelligence.It is also a major challenge in the field of rese

303、arch and application.”285 Xu Bo(徐波),director of CAS Institute of Automation and chair of Chinas Next Generation Artificial Intelligence Strategic Advisory Committee.“General artificial intelligence has always been a dream in the technology world.We believe autonomous evolution a topic of the institu

304、tes research is a bridge from weak artificial intelligence to general artificial intelligence.”286 Yao Dezhong(堯德中),director of the MOE Lab for Neuroinformation(神經信息教育部重點實驗室)in Chengdu on development of a“digital twin brain(數字孿生腦)model.”287 Center for Security and Emerging Technology39“It is believe

305、d that the future digital twin brain can become a general intelli-gent system in a certain sense by imitating the working principle of the human nervous system,bringing AI to a new level.Thinking and decision-making are close to the human brain.”288 Zhu Songchun(朱松純),director of BIGAI,former UCLA co

306、mputer vision expert,proponent of the“small data,big task”approach to AI development.“The goal of general AI research is to create general agents with autono-mous perception,cognition,decision-making,learning,execution,and social collaboration capabilities that conform to human emotions,ethics,and m

307、oral concepts.”289 Zeng Yi(曾毅,C.9 above),prominent AI safety advocate,speaker at the Future of Life Institutes“Beneficial AGI 2019”conference,and head of a BAAI group that developed the“Beijing AI Principles,”290 whom one would expect to be against general AI,is on record backing it.“The development

308、 of narrow(專用)AI does not completely avoid risk,be-cause the system is likely to encounter unexpected scenarios in its applica-tion.Having a certain general ability may improve the robustness and adap-tiveness of an intelligent system.”291 Support from the scientific establishment does not guarantee

309、 acceptance by society.Outside of China,AI(not to mention AGI or ASI)conjures up dystopian scenarios among many informed persons(Hawking,Musk,Bostrom,and Gates)concerns that are well-founded.While the authors share these concerns,overcau-tion can mean ceding to ones competitor the“first mover advant

310、age”(先發優勢)China demonstrably seeks,292 which in AI may be difficult or impossible to close.So it is worth contemplating how receptive China as a whole is to an expanded role of AI in human life.While the authors know of no popular surveys,Chinas acceptance of general AI is the topic of two recent bo

311、oks.Song Bing(宋冰),vice president of the Berggru-en Institute and director of its PKU-based China Center(北京大學博古睿研究中心)published an edited book by 17 Chinese philosophers and scientists titled 智能與智慧:人工智能遇見中國哲學家(Intelligence and Wisdom-Artificial Intelligence Meets Chinese Philosophers),293 billed as“th

312、e first systematic endeavor by prom-inent Chinese philosophers.to address challenges and opportunities posed by frontier technologies such as artificial intelligence and robotics.”294 Center for Security and Emerging Technology40Song,a former Goldman Sachs executive and capital markets lawyer presum

313、-ably not given to flights of fantasy,summarized the books content in interviews.As she sees it,“Western media and intellectual elites generally have more vigilance and fear of cutting-edge technology,”and of AI especially,than Chinese,which she attributes to Western human-centrismin contrast to the

314、“non-anthropocentrism”(非人類中心主義)of China.So“the emergence of an existence stronger than human is not a problem.”295“What is wrong coexisting with machine superintelligence(超級機器智能)?”296Song is correct ascribing to China a preference for holistic thought that denies humanitys privileged status.297 Thes

315、e contrasting tendencies of Eastern and West-ern thought have been known for decades and observed in careful psychological studies.298 The second book,by Fudan University philosopher Xu Yingjin(徐英瑾)titled 人工智慧哲學十五講(Fifteen Lectures on the Philosophy of Artificial Intelligence),299 likewise comes dow

316、n unequivocally in favor of strong AI:“Does the ethical benefit of researching general AI outweigh the cons,or does the harm outweigh the benefits?My answer is the pros outweigh the cons(利大于弊).300 The ultimate gauge of AIs acceptance is the attitude of Chinas ruling class,as measured by institutiona

317、l support,public declarations,and by the relationship of artificial intelligence to the needs of political dictatorship.The first of these,state sup-port,is documented in a 2022 book edited by the present authors,301 especially its chapter on“State plans,research,and funding.”302 Support for advance

318、d AI is also explicit in Chinas 2017 foundational AI document,the“New Generation AI Devel-opment Plan,”303 which lays the groundwork for“general artificial intelligence,”(通用人工智能),while calling for a merger(混合)of AI and human intelligence.304 Endorsements by Chinas top political leaders are easily fo

319、und.On May 30,2016,President Xi Jinping in a speech to a MOST-sponsored assembly titled“Striv-ing to Build a World S&T Superpower”made the following reference to one of the three advanced AI pillars in China:“Connectomics is at the scientific forefront for understanding brain function and further ex

320、ploring the nature of consciousness.Exploration in this area not only has important scientific significance,but also has a guiding role in the prevention and treatment of brain disease and the development of intelligent technology.”305 Center for Security and Emerging Technology41Xis report to the 1

321、9th Party Congress in October 2017 highlighted progress in AI and quantum science.306 Since Xi assumed power in 2012,only three of 70-odd Politburo study sessions have been on specific technologiesAI(2018),blockchain(2019),and quantum computing(2020).307 CAS Institute of Automations website lists 13

322、 instances from May 2018 to November 2020 where Xi publicly expressed support for AI.308 For example:“Driven by new theories and technologies such as mobile internet,big data,supercomputing,sensor networks,and brain science,artificial intelligence has accelerated its development,causing the emergenc

323、e of deep learning,cross-discipline integration,human-machine collaboration,group intelli-gence,and autonomous control.”309 In Xis estimate,AI has the effect of a“lead goose”(頭雁),i.e.,a spillover function driving the current S&T“revolution”and“industrial transformation(產業變革).”310Are there other reas

324、ons for Chinese leaders embrace of AI?It is an open ques-tion whether advances in information technology benefit freedom(cryptocurrency,blockchain,virtual private networks)or dictatorship(censorship,predictive policing,ubiquitous surveillance).AI-enabled influence operations are also a two-way stree

325、t.What seems certain is that whichever“side”has a technical advantage is primed to prevail.As the struggle for information dominance becomes automated,the poten-tial for AI surprise will increase,raising the stakes.The Chinese governments use of AI as an instrument of oppression is document-ed by Da

326、hlia Peterson and Maya Wang,among others,and requires no elabora-tion here.31 1 It is also a topic of study at CSET,potentially as a follow-up database exercise.312 The point is that Chinas Communist Party leaders acceptance of ad-vanced AI,beyond enhancing national power,may be motivated by a need

327、to own it just to hold the line.Center for Security and Emerging Technology52Center for Security and Emerging Technology43Recommendationshina,through complementary approaches and state backing,is vigorously pursuing advanced general AI,understood as broadly capable software that can function autonom

328、ously(creatively)in novel environments.Whether this research leads to human-level AIhow-ever definedis irrelevant,as the impact of these advances on Chinas ability to survive as an authoritarian state while projecting its will globally will be comparable on many levels to scenarios associated with h

329、igh-level machine intelligence.Indeed,if Chinas programs in BCI and connectom-ics succeed,the distinction between human and machine intelligence may disappear.What might frustrate Chinas goals?We acknowledge,with Chinas S&T community,313 that basic science has not been the countrys strong suit,but f

330、ind little comfort given(1)a paradigm shift in the quality and quantity of China-authored scientific papers,(2)Chinas success in facilitating the return,importation,and cooperation of foreign-trained scientists,and(3)Chinas storied ability to operationalize scientific discovery in advance of those m

331、aking the discoveries,which in the end is all that matters.314 These circumstances take on added urgency in light of AIs function as an enabler of most other scientific endeavors.Chinas intent to lead the world in AI by 2030 and achieve a“first mover advantage”cannot be dismissed as rhetoric but is

332、a serious pros-pect that will have practical consequences.Of course,Chinas future capa-bilities relative to other nations is a key element this study does not address,although it is obvious that projections of any sort depend on the availability of data and ones ability to derive from it indications

333、 that can be assessed.CCenter for Security and Emerging Technology44This paper outlines a monitoring project meant to kickstart an I&W regime.The need is not limited to China or AI but applies to foreign S&T in general.Sadly,such an open-source collection-and-analysis system does not exist anywhere among the liberal democraciesneither inside nor outside the U.S.intelligence communityan issue the a