《CSET：2025中國學界對大語言模型的批判性思考：通用人工智能 （AGI） 的多元路徑探索研究報告（英文版）（29頁）.pdf》由會員分享，可在線閱讀，更多相關《CSET：2025中國學界對大語言模型的批判性思考：通用人工智能 （AGI） 的多元路徑探索研究報告（英文版）（29頁）.pdf（29頁珍藏版）》請在三個皮匠報告上搜索。

1、Issue BriefJanuary 2025Chinese Critiques of Large Language ModelsFinding the Path to General Artificial IntelligenceAuthorsWm.C.HannasHuey-Meei ChangMaximilian RiesenhuberDaniel H.ChouChinese Critiques of Large Language ModelsFinding the Path to General Artificial IntelligenceAuthorsWm.C.HannasHuey-

2、Meei ChangMaximilian RiesenhuberDaniel H.Chou Center for Security and Emerging Technology|1 Executive Summary Large language models have garnered interest worldwide owing to their remarkable ability to“generate”human-like responses to natural language queriesa threshold that at one time was consider

3、ed“proof”of sentienceand perform other time-saving tasks.Indeed,LLMs are regarded by many as a,or the,pathway to general artificial intelligence(GAI)that hypothesized state where computers reach(or even exceed)human skills at most or all tasks.The lure of achieving AIs holy grail through LLMs has dr

4、awn investment in the billions of dollars by those focused on this goal.In the United States and Europe especially,big private sector companies have led the way and their focus on LLMs has overshadowed research on other approaches to GAI,despite LLMs known downsides such as cost,power consumption,un

5、reliable or“hallucinatory”output,and deficits in reasoning abilities.If these companies bets on LLMs fail to deliver on expectations of progress toward GAI,western AI developers may be poorly positioned to rapidly fall back on alternate approaches.In contrast,China follows a state-driven,diverse AI

6、development plan.Like the United States,China also invests in LLMs but simultaneously pursues alternate paths to GAI,including those more explicitly brain-inspired.This report draws on public statements by Chinas top scientists,their associated research,and on PRC government announcements to documen

7、t Chinas multifaceted approach.The Chinese government also sponsors research to infuse“values”into AI intended to guide autonomous learning,provide AI safety,and ensure that Chinas advanced AI reflects the needs of the people and the state.This report concludes by recommending U.S.government support

8、 for alternative general AI programs and for closer scrutiny of Chinas AI research.Center for Security and Emerging Technology|2 Introduction:Generative AI and General AI Achieving general artificial intelligence or GAIdefined as AI that replicates or exceeds most human cognitive skills across a wid

9、e range of tasks,such as image/video understanding,continual learning,planning,reasoning,skill transfer,and creativity1is a key strategic goal of intense research efforts both in China and the United States.2 There is vigorous debate in the international scientific community regarding which path wil

10、l lead to GAI most quickly and which paths may be false starts.In the United States,LLMs have dominated the discussion,yet questions remain about their ability to achieve GAI.Since choosing the wrong path can position the United States at a strategic disadvantage,this raises the urgency of examining

11、 alternative approaches that other countries may be pursuing.In the United States,many experts believe the transformative step to GAI will occur with the rollout of new versions of LLMs such as OpenAIs o1,Googles Gemini,Anthropics Claude,and Metas Llama.3 Others argue,pointing to persistent problems

12、 such as LLM hallucinations,that no amount of compute,feedback,or multimodal data sources will allow LLMs to achieve GAI.4 Still other AI scientists see roles for LLMs in GAI platforms but not as the only,or even main,component.5 Pondering the question of how GAI can be achieved is important because

13、 it touches on options available to developers pursuing AIs traditional holy grailhuman-level intelligence.Is the pathor a pathto GAI a continuation of LLM development,possibly augmented by additional modules?Or are LLMs a dead end,necessitating other,fundamentally different approaches that are base

14、d on a closer emulation of human cognition and brain function?Given the success of LLMs,the levels of investment,6 endorsements by highly regarded AI scientists,optimism created by working examples,and the difficulty of reimagining new approaches in the face of models in which companies have great c

15、ommitment,it is easy to overlook the risk of relying on a“monoculture”based on a single research paradigm.7 If there are limitations to what LLMs can deliver,without a sufficiently diversified research portfolio,it is unclear how well western companies and governments will be able to pursue other so

16、lutions that can overcome LLMs problems as pathways to GAI.A diversified research portfolio is precisely Chinas approach to its state-sponsored goal of achieving“general artificial intelligence”(通用人工智能).8 This report will show thatin addition to Chinas known and prodigious effort to field ChatGPT-li

17、ke LLMs,9significant resources are directed in China at alternative pathways to GAI by Center for Security and Emerging Technology|3 scientists who have well-founded concerns about the potential of“big data,small task”(大數據,小任務)approaches to reach human capabilities.10 Accordingly,this paper addresse

18、s two questions:What criticisms do Chinese scientists have of LLMs as paths to general AI?And how is China managing LLMs alleged shortcomings?The paper begins(section 1)with critiques by prominent non-China AI scientists of large language models and their ability to support GAI.The section provides

19、context for views of Chinese scientists toward LLMs(section 2)described in online sources.Section 3 then cites research that supports Chinas public-facing claims about the non-viability of LLMs as a path to GAI.In section 4,we assess these claims as a basis for recommendations in section 5 on why Ch

20、inas alternative projects must be taken seriously.Center for Security and Emerging Technology|4 Large Language Models and Their Critics The term“large language model”captures two facts:they are large networks typically with billions to trillions of parameters,and they are trained on natural language

21、,terabytes of text ingested from the internet and other sources.LLMs,and neural networks(NN)generally,are typologically distinct from“good old-fashioned”(GOFAI)symbolic AI that depends on rule-based coding.In addition,todays large models can manage,to different degrees,multimodal inputs and outputs,

22、including images,video,and audio.11 LLMs debuted in 2017,when Google engineers proposed a NN architecturecalled a transformeroptimized to find patterns in sequences of text by learning to“pay attention”to the co-occurrence relationships between“tokens”(words or parts of words)in the training corpus.

23、12 Unlike human knowledge,knowledge captured in LLMs is not obtained through interactions with the natural environment but depends on statistical probabilities derived from the positional relationships between the tokens in sequences.Massive exposure to corpora during training allows the LLM to iden

24、tify regularities that,in the aggregate,can be used to generate responses to human prompts after the training.Hence,the OpenAI product name“GPT”(generative pre-trained transformer).The ability of LLMs to“blend”different sources of information(which plays to traditional strengths of neural networks i

25、n pattern matching and uncovering similarities in complex spaces)has given rise to applications in areas as diverse as text summarization,translation,code writing,and theorem proving.Yet,it has been hotly debated whether this ability to find and exploit regularities is sufficient to achieve GAI.Init

26、ial enthusiastic reports regarding the“sentience”of LLMs are increasingly supplemented by reports showing serious deficits in LLMs ability to understand language and to reason in a human-like way.13 Some persistent deficits in LLMs,as in basic math,14 appear correctable by plugins,15 i.e.,external p

27、rograms specialized for areas of LLM weaknesses.In fact,such an approachof a network of systems specialized in different aspects of cognitionwould be more like the brain,which has dedicated modules,e.g.,for episodic memory,math,reasoning,etc.,rather than a single process as in LLMs.16 Some scientist

28、s hope that increases in complexity alone might help overcome LLMs defects.For instance,Geoffrey Hinton,crediting an intuition of Ilya Sutskever(OpenAIs former chief scientist,who studied with Hinton),believes scale will solve some of these problems.In this view,LLMs are already“reasoning”by virtue

29、of their ability“to Center for Security and Emerging Technology|5 predict the next symbol and prediction is a pretty plausible theory of how the brain is learning.”17 Indeed,increases in complexity(from GPT-2 through GPT-4)have led to increased performance on various benchmark tasks,such as“theory o

30、f mind”18(reasoning about mental states),where deficits were noted for GPT-3.5.19 Other such deficits are harder to address and persist despite increases in model complexity.Specifically,“hallucinations,”i.e.,LLMs making incorrect claims(a problem inherent to neural networks that are designed to int

31、erpolate and,unlike the brain,do not separate the storage of facts from interpolations)and errors in reasoning have been difficult to overcome,20 with recent studies showing that the likelihood of incorrect/hallucinatory answers increases with greater model complexity.21 In addition,the strategy of

32、increasing model complexity in the hope of achieving novel,qualitatively different“emergent”behaviors that appear once a computational threshold has been crossed likewise has been called into question by research showing that previously noted“emergent”behaviors in larger models were artefacts of the

33、 metrics used and not indicative of any qualitative changes in model performance.22 Correspondingly,claims of“emergence”in LLMs have declined in the recent literature,even as model complexities have increased.23 Indeed,there is the justified concern that the high performance of LLMs on standardized

34、tests could be ascribed more to the well-known pattern matching prowess of neural networks than the discovery of new strategies.24 Still other criticisms of LLMs center on fundamental cognitive and philosophical issues such as the ability to generalize,form deep abstractions,create,self-direct,model

35、 time and space,show common sense,reflect on their own output,25 manage ambiguous expressions,unlearn based on new information,evaluate pro and con arguments(make decisions),and grasp nuance.26 While these deficits are discussed in the western research literature,along with others such as LLMs inabi

36、lity to easily add knowledge beyond the context window without retraining the base model,or the high computational and energy demands of LLM training,most current investment of commercial players in the AI space(e.g.,OpenAI,Anthropic)is continuing down this same road.The problem is not only that“we

37、are investing in an ideal future that may not materialize”27 but rather that LLMs,in Google AI researcher Franois Chollets words,“sucked the oxygen out of the room.Everyone is just doing LLMs.”28 Center for Security and Emerging Technology|6 Chinese Views of LLMs as a Path to General AI(or Not)A rev

38、iew of statements by ranking scientists at Chinas top AI research institutes reveals a high degree of skepticism about LLMs ability to lead,by themselves,to GAI.These criticisms resemble those of international experts,because both groups face the same problems and because Chinas AI experts interact

39、with their global peers as a matter of course.29 Here follow several Chinese scientists views on LLMs as a path to general AI.Tang Jie(唐杰)is professor of computer science at Tsinghua University,the founder of Zhipu(智譜),30 a leading figure in the Beijing Academy of Artificial Intelligence(BAAI),31 an

40、d the designer of several indigenous LLMs.32 Despite his success with statistical models,Tang argues that human-level AI requires the models to be“embodied in the world.”33 Although he believes the scaling law(規模法則)34“still has a long way to go,”that alone does not guarantee GAI will be achieved.35

41、A more fruitful path would take cues from biology.In his words:“GAI or machine intelligence based on large models does not necessarily have to be the same as the mechanism of human brain cognition,but analyzing the working mechanism of the human brain may better inspire the realization of GAI.”36 Zh

42、ang Yaqin(張亞勤,AKA Ya-Qin Zhang)co-founded Microsoft Research Asia,is the former president of Baidu,founding dean of Tsinghuas Institute for AI Industry Research(智能產業研究院)and a BAAI advisor.Zhang cites three problems with LLMs,namely,their low computational efficiency,inability to“truly understand the

43、 physical world,”and so-called“boundary issues”(邊界問題),i.e.,tokenization.37 Zhang believes(with Goertzel)that“we need to explore how to combine large generative probabilistic models with existing first principles of the physical world or real models and knowledge graphs.”38 Huang Tiejun(黃鐵軍)is founde

44、r and former director of BAAI and vice dean of Peking Universitys(PKU)Institute for Artificial Intelligence(人工智能研究院).Huang names three paths to GAI:“information models”based on big data and big compute,“embodied models”trained through reinforcement learning,and brain emulationin which BAAI has a maj

45、or stake.39 Huang agrees that LLM scaling laws will continue to operate but adds“it is not only necessary to collect static data,but also to obtain and process multiple sensory information in real time.”40 In his view,GAI depends on integrating statistical models with brain-inspired AI and embodimen

46、t,that is:Center for Security and Emerging Technology|7 LLMs represent“static emergence based on big data”(是基于大數據的靜態涌現).Brain-inspired intelligence,by contrast,is based on complex dynamics.Embodied intelligence also differs in that it generates new abilities by interacting with the environment.41 Xu

47、 Bo(徐波),dean of the School of Artificial Intelligence at University of Chinese Academy of Sciences(UCAS)(中國科學院大學人工智能學院)and director of the Chinese Academy of Sciences(CAS)Institute of Automation(CASIA,中國科學院自動化研究所),42 and Pu Muming(蒲慕明,AKA Muming Poo),director of CASs Center for Excellence in Brain S

48、cience and Intelligence Technology(CEBSIT,腦科學與智能技術卓越創新中心)43 believe embodiment and environmental interaction will facilitate LLMs growth toward GAI.Although the artificial neural networks on which LLMs depend were inspired by biology,they scale by adding“more neurons,layers and connections”and do no

49、t begin to emulate the brains complexity of neuron types,selective connectivity,and modular structure.In particular,“Computationally costly backpropagation algorithms could be improved or even replaced by biologically plausible learning algorithms.”These candidates include spike time synaptic plasti

50、city,“neuromodulator-dependent metaplasticity”and“short-term vs.long-term memory storage rules that set the stability of synaptic weight changes.”44 Zhu Songchun(朱松純,AKA Song-Chun Zhu)dean of PKUs Institute of Artificial Intelligence and director of the Beijing Institute for General Artificial Intel

51、ligence(北京通用人工智能研究院)founded BIGAI on the premise that big data-based LLMs are a dead-end in terms of their ability to emulate human-level cognition.45 Zhu pulls no punches:“Achieving general artificial intelligence is the original intention and ultimate goal of artificial intelligence research,but c

52、ontinuing to expand the parameter scale based on existing large models cannot achieve general artificial intelligence.”Zhu compares Chinas LLMs achievements to“climbing Mt.Everest”when the real goal is to reach the moon.In his view,LLMs are“inherently uninterpretable,have risks of data leakage,do no

53、t have a cognitive architecture,and lack causal and mathematical reasoning capabilities,and other limitations,so they cannot lead to general artificial intelligence.”46 Zeng Yi(曾毅),director of CASIAs Brain-inspired Cognitive Intelligence Lab(類腦認知智能實驗室)and founding director of its International Resea

54、rch Center for AI Ethics and Center for Security and Emerging Technology|8 Governance,47 is building a GAI platform based on time-dependent spiking neural networks.In his words:“Our brain-like cognitive intelligence team firmly believes that only by mirroring the structure of the human brain and its

55、 intelligent mechanism,as well as the laws and mechanisms of natural evolution,can we achieve artificial intelligence that is truly meaningful and beneficial to humans.”48 Criticisms of LLMs by other Chinese AI scientists are legion.Shen Xiangyang(沈向洋,Harry Shum AKA Heung-Yeung Shum),former Microsof

56、t executive VP and director of the Academic Committee of PKUs Institute of Artificial Intelligence,laments that AI research has no“clear understanding of the nature of intelligence.”Shen supports a view he attributes to New York University professor emeritus and LLM critic Gary Marcus that“no matter

57、 how ChatGPT develops,the current technical route will not be able to bring us real intelligence.”49 Zheng Qinghua(鄭慶華),president of Tongji University and a Chinese Academy of Engineering academician,stated that LLMs have major flaws:they consume too much data and computing resources,are susceptible

58、 to catastrophic forgetting,have weak logical reasoning capabilities,and do not know when they are wrong or why they are wrong.50 Li Wu(李武),director of the State Key Laboratory of Cognitive Neuroscience and Learning at Beijing Normal University,stated his belief that“current neural networks are rela

59、tively specialized and do not conform to the way the human brain works.If you desperately hype the large model itself and only focus on the expansion of parameters from billions or tens of billions to hundreds of billions,you will not be able to achieve true intelligence.”51 Recognition of the need

60、to supplement LLM research with alternative paths to GAI is evidenced in statements by Chinas national and municipal governments.On May 30,2023,Beijings city governmentwithin whose jurisdiction much of Chinas GAI-oriented LLM research is taking placeissued a statement calling for development of“larg

61、e models and other general artificial intelligence technology systems”(系統構建大模型等通用人工智能技術體系).52 Section three has five items(7-11),the first four of which pertain to LLMs(algorithms,training data,evaluation,and a basic software and hardware system).Item 11 reads“exploring new paths(新路徑)for general art

62、ificial intelligence”and calls for:Center for Security and Emerging Technology|9 Developing a basic theoretical system(基礎理論體系)for GAI,autonomous collaboration and decision-making,embodied intelligence,and brain-inspired(類腦)intelligence,supported by a unified theoretical framework,rating and testing

63、standards,and programming languages.Embodied systems(robots)will train in open environments,generalized scenarios,and continuous tasks.The plan also mandates the following:“Support the exploration of brain-like intelligence,study the connection patterns,coding mechanisms,information processing and o

64、ther core technologies of brain neurons,and inspire new artificial neural network modeling and training methods.”Alternatives to LLMs were cited at the national level in March 2024,when CAS vice president Wu Zhaohui(吳朝暉,formerly vice minister of Chinas science ministry and president of Zhejiang Univ

65、ersity),53 stated that AI is moving toward“synergy between large and small models”(大小模型協同),adding that China must“explore the development of GAI in multiple ways”(多路徑地探索通用人工智能發展).The latter include“embodied intelligence,distributed group intelligence,human-machine hybrid intelligence,enhanced intell

66、igence,and autonomous decision making.”54 The following month Beijings Haidian District government,with jurisdiction over 1,300 AI companies,more than 90 of which are developing big models,55 issued a three-year plan to facilitate research in embodied(具身)AI.The plan defines“embodiment”as“the ability

67、 of an intelligent system or machine to interact with the environment in real time through perception and interaction”and is meant to serve as a platform for nationwide development.Its details include plans for humanoid robots facilitated by replicating brain functionality.56 Our analysis of public

68、statements by government institutions and ranking Chinese AI scientists indicates that an influential part of Chinas AI community shares the concerns and misgivings held by western critics of LLMs and seeks alternative pathways to general artificial intelligence.Center for Security and Emerging Tech

69、nology|10 What Does the Academic Record Show?Public statements by scientists are one measure of Chinas approach to GAI.Another is their record of scholarship.Prior reviews of Chinese technical literature determined that China is pursuing GAI by multiple means,including generative large language mode

70、ls,57 brain-inspired models,58 and by enhancing cognition through brain-computer interfaces.59 Our present task is to examine the literature for evidence that Chinese scholarsbeyond what positive features brain-based models haveare also driven to seek alternative paths by LLMs shortcomings.Toward th

71、at end,we ran keyword searches in Chinese and English for“AGI/GAI+LLM”and their common variants in CSETs Merged Corpus60 for papers published in 2021 or later with primary Chinese authorship.Some 35 documents were obtained.A separate query using web-based searches recovered 43 more papers.61 15 of t

72、he 78 papers were rejected by the studys lead analyst as off topic.The remaining 63 papers were reviewed by the studys subject matter expert,who highlighted the following 24 as examples of Chinese research addressing LLM problems that stand in the way of large models achieving the generality associa

73、ted with GAI.62 1.CAO Boxi(曹博西),HAN Xianpei(韓先培),SUN Le(孫樂),“Can Prompt Probe Pretrained Language Models?Understanding the Invisible Risks from a Causal View,”arXiv preprint arXiv:2203.12258v1(2022).2.CHENG Bing(程兵),“Artificial Intelligence Generative Content(AIGC)including ChatGPT Opens a New Big P

74、aradigm Space of Economics and Social Science Research”(以 ChatGPT 為代表的大語言模型打開了經濟學和其他社會科學研究范式的巨大新空間),China Journal of Econometrics(計量經濟學報)3,no.3(July 2023).3.CHENG Daixuan(程岱宣),HUANG Shaohan(黃少涵),WEI Furu(韋福如),“Adapting Large Language Models to Domains via Reading Comprehension,”arXiv preprint arXiv:

75、2309.09530v4(2024).4.DING Ning(丁寧),ZHENG Hai-Tao(鄭海濤),SUN Maosong(孫茂松),“Parameter-efficient Fine-tuning of Large-scale Pre-trained Language Models,”Nature Machine Intelligence,March 2023.5.DONG Qingxiu(董青秀),SUI Zhifang(穗志方),LI Lei(李磊),“A Survey on In-context Learning,”arXiv preprint arXiv:2301.00234

76、v4(2024).6.HUANG Jiangyong(黃江勇),YONG Silong(雍子隆),63 HUANG Siyuan(黃思遠),“An Embodied Generalist Agent in 3D World,”Proceedings of the 41st International Conference on Machine Learning,Vienna,Austria,PMLR 235.2024.Center for Security and Emerging Technology|11 7.JIN Feihu(金飛虎),ZHANG Jiajun(張家俊),“Unifie

77、d Prompt Learning Makes Pre-trained Language Models Better Few-shot Learners,”IEEE International Conference on Acoustics,Speech and Signal Processing,June 2023.8.LI Hengli(李珩立),ZHU Songchun(朱松純),ZHENG Zilong(鄭子隆),“DiPlomat:A Dialogue Dataset for Situated Pragmatic Reasoning,”37th Conference on Neura

78、l Information Processing Systems(NeurIPS 2023).9.LI Jiaqi(李佳琪),ZHENG Zilong(鄭子隆),ZHANG Muhan(張牧涵),“LooGLE:Can Long-Context Language Models Understand Long Context?”arXiv preprint arXiv:2311.04939v1(2023).10.LI Yuanchun(李元春),ZHANG Yaqin(張亞勤),LIU Yunxin(劉云新),“Personal LLM Agents:Insights and Survey ab

79、out the Capability,Efficiency and Security,”arXiv preprint arXiv:2401.05459v2(2024).11.MA Yuxi(馬煜曦),ZHU Songchun(朱松純),“Brain in a Vat:on Missing Pieces towards Artificial General Intelligence in Large Language Models,”arXiv preprint arXiv:2307.03762v1(2023).12.NI Bolin(尼博琳),PENG Houwen(彭厚文),CHEN Min

80、ghao,ZHANG Songyang (張宋揚),LING Haibin(凌海濱),“Expanding Language-image Pretrained Models for General Video Recognition,”arXiv preprint arXiv:2208.02816v1(2022).13.PENG Yujia(彭玉佳),ZHU Songchun(朱松純),“The Tong Test:Evaluating Artificial General Intelligence through Dynamic Embodied Physical and Social In

81、teractions,”Engineering 34,(2024).14.SHEN Guobin(申國斌),ZENG Yi(曾毅),“Brain-inspired Neural Circuit Evolution for Spiking Neural Networks,”PNAS 39(2023).15.TANG Xiaojuan(唐曉娟),ZHU Songchun(朱松純),LIANG Yitao(梁一韜),ZHANG Muhan(張牧涵),“Large Language Models Are In-context Semantic Reasoners Rather than Symboli

82、c Reasoners,”arXiv preprint arXiv:2305.14825v2(2023).16.WANG Junqi(王俊淇),PENG Yujia(彭玉佳),ZHU Yixin(朱毅鑫),FAN Lifeng(范麗鳳),“Evaluating and Modeling Social Intelligence:a Comparative Study of Human and AI Capabilities,”arXiv preprint arXiv:2405.11841v1(2024).17.XU Fangzhi(徐方植),LIU Jun(劉軍),Erik Cambria,“A

83、re Large Language Models Really Good Logical Reasoners?”arXiv preprint arXiv:2306.09841v2(2023).18.XU Zhihao(徐智昊),DAI Qionghai(戴瓊海),FANG Lu(方璐),“Large-scale Photonic Chiplet Taichi Empowers 160-TOPS/W Artificial General Intelligence,”Science,April 2024.19.YUAN Luyao(袁路遙),ZHU Songchun(朱松純),“Communica

84、tive Learning:a Unified Learning Formalism,”Engineering,March 2023.Center for Security and Emerging Technology|12 20.ZHANG Chi(張馳),ZHU Yixin(朱毅鑫),ZHU Songchun(朱松純),“Human-level Few-shot Concept Induction through Minimax Entropy Learning,”Science Advances,April 2024.21.ZHANG Tielin(張鐵林),XU Bo(徐波),“A

85、Brain-inspired Algorithm that Mitigates Catastrophic Forgetting of Artificial and Spiking Neural Networks with Low Computational Cost,”Science Advances,August 2023.22.ZHANG Yue(章岳),CUI Leyang(崔樂陽),SHI Shuming(史樹明),“Sirens Song in the AI Ocean:a Survey on Hallucination in Large Language Models,”arXiv

86、 preprint arXiv:2309.01219v2(2023).23.ZHAO Zhuoya(趙卓雅),ZENG Yi(曾毅),“A Brain-inspired Theory of Mind Spiking Neural Network Improves Multi-agent Cooperation and Competition.”Patterns,August 2023.24.ZOU Xu(鄒旭),YANG Zhilin(楊植麟),TANG Jie(唐杰),“Controllable Generation from Pre-trained Language Models via

87、Inverse Prompting,”arXiv preprint arXiv:2103.10685v3(2021).The studies collectively address the litany of LLM deficits described in this papers sections 1 and 2,namely,those associated with theory of mind(ToM)failures,inductive,deductive,and abductive reasoning deficits,problems with learning new ta

88、sks through analogy to previous tasks,lack of grounding/embodiment,unpredictability of errors and hallucinations,lack of social intelligence,insufficient understanding of real-world input,in particular in video form,difficulty in dealing with larger contexts,challenges associated with the need to fi

89、ne tune outputs,and cost of operation.Proposed solutions to these problems range from adding modules,emulating brain structure and processes,rigorous standards and testing,and real-world embedding,to replacing the computing substrate outright with improved chip types.Several prominent Chinese scient

90、ists cited in this studys section 2,who made public statements supporting alternate GAI models,including Tang Jie,Zhang Yaqin,Xu Bo,Zhu Songchun,and Zeng Yi,are on the bylines of many of these papers,adding authenticity to their declarations.In addition,virtually all of Chinas top institutions and c

91、ompanies engaged in GAI research,including the Beijing Academy of Artificial Intelligence(北京智源人工智能研究院),the Beijing Institute for General Artificial Intelligence(北京通用人工智能研究院),the Chinese Academy of Sciences Institute of Automation(中國科學院自動化研究所),Peking University(北京大學),Tsinghua University(清華大學),Univers

92、ity of Chinese Center for Security and Emerging Technology|13 Academy of Sciences(中國科學院大學)and Alibaba,ByteDance,Huawei,and Tencent AI lab,are represented in the selected corpus,in most cases on multiple papers.64 The record of metadata adduced here,and conclusions drawn in prior CSET research65 supp

93、ort the present studys contention that major elements in Chinas AI community question LLMs potential to achieve GAIthrough increases in scale or modalitiesand are contemplating or pursuing alternative pathways.Center for Security and Emerging Technology|14 Assessment:Do All Paths Lead to the Buddha?

94、When LLM-based chatbots first became available,early claims that LLMs might be sentient,i.e.,experience feelings and sensations,or even show self-awareness,66 were prevalent and much discussed.Since then,cooler heads have prevailed,67 and the focus has shifted from philosophical speculations about t

95、he interior lives of LLMs to more concrete measurements of LLM abilities on key indicators of“intelligent”behavior and the strategically important question of whether LLMs might be capable of general artificial intelligence(GAI).While it is far from clear whether consciousness and the capacity for e

96、motions are critical to GAI,what is clear is that a GAI system must be able to reason and to separate facts from hallucinations.As things stand,LLMs have no explicit mechanisms that would enable them to perform these core requirements of intelligent behavior.Rather,the hope of LLM enthusiasts is tha

97、t,somehow,reasoning abilities will“emerge”as LLMs are trained to become ever better at predicting the next word in a conversation.Yet,there is no theoretical basis for this belief.To the contrary,research has shown that LLMs vast text memory has masked deficiencies in reasoning.68 Heuristic attempts

98、 to improve reasoning(e.g.,chain-of-thought),69 likely the basis for improved performance in OpenAIs new“o1”LLM,and more recent approaches such as“rephrase and respond,”70“tree-of-thoughts”71 or“graph-of-thoughts”72 have yielded improvements,but do not solve the underlying problem of the absence of

99、a core“reasoning engine.”By the same token,multiple attempts to fix LLMs hallucination problem73 have run into dead ends because they fail to address the core problem that is inherent to LLMs ability to generalize from training data to new contexts.Indeed,current efforts to improve reasoning abiliti

100、es and fix hallucinations are a bit like playing“whack-a-mole”but with moles hiding in a billion-dimensional weight space and with a mallet that is uncertain to hit where intended.The resulting systems might be sufficient for situations where humans can assess the quality of LLM output,e.g.,writing

101、cover letters,designing travel itineraries or creating essays on topics that are perennial favorites of high school teachers.Yet,these capabilities are a far cry from GAI.The public debates in the western world on the appropriate path to GAI tend to be drowned out by companies with financial interes

102、ts in promoting their latest LLMs with claims of“human-like intelligence”or“sparks of artificial general intelligence,”74 even in the face of ever more apparent shortcomings of LLMs,as detailed in section 1.The dominance of commercial interests that promote LLMs as sure paths to GAI has Center for S

103、ecurity and Emerging Technology|15 already negatively affected the ability of academic research in the U.S.to pursue alternative approaches to GAI.75 The situation is different in China.While there are also companies in China developing LLMs for commercial purposes,leading Chinese AI scientists and

104、government officials,as detailed in this paper,realize that LLMs have fundamental limitations that make it important to investigate other approaches to GAI or supplement LLM performance using“brain-like”algorithms.The latter strategy,of pursuing“brain-inspired”AI has led to major breakthroughs in th

105、e past,for example,by combining deep learning76modeled on the brains sensory processing hierarchyand reinforcement learning77modeling how the brain learns strategies from rewardsinto“deep reinforcement learning,”78 which,for instance,formed the basis of AlphaGo,79 the first artificial neural network

106、 that beat human champions in the game of Go.This difference in research directions may give China an advantage in the race to achieve GAI.It might be helpful to compare the current situation to how China came to dominate the global market for photovoltaic(PV)panels(or,more recently,battery technolo

107、gy and electric vehicles),based on Chinese government decisions made at the beginning of the millennium to become a world leader in PV.The ensuing policy decisions and investments to build up the domestic PV industry and increase the efficiency of PV panels led to innovation and economies of scale t

108、hat now have China producing at least 75%of the worlds solar panels.A decision by China to strategically invest in non-LLM-based approaches to GAI80 may repeat this success,albeit in a field of even greater importance than photovoltaics.Center for Security and Emerging Technology|16 Managing a China

109、 First-Mover Advantage Geoffrey Hinton,recent Nobel Prize winner and recipient of a Turing Award for his work on multilayer neural networksthe first AI NN architecture that led to superhuman performance on a range of benchmark tasks in computer vision and other areasacknowledges“a race,clearly,betwe

110、en China and the U.S.,and neither is going to slow down.”81 This race to general AI is typically characterized as a competition for data,chips,talent,and energy,82 with success measured on benchmarks meant to assess“human-level intelligence.”The assumption underlying these comparisons is that both s

111、ides are competing in the same arena.This view is dangerously misleading.The present study shows major elements in Chinas AI community pursuing alternative paths to GAI in which model complexitytaken by many in the U.S.as a proxy for performance,conditioned by companies focus on the number of parame

112、ters of their models as a distinguishing featureplays only a subsidiary role.These non-traditional approaches,moreover,have Chinese state backing.Utility aside,pragmatism likely also motivates PRC support for a general AI that avoids the inherent uncontrollability of large statistical models,83 whic

113、h along with their hallucinations and other pesky foibles also resist top-down government censorship,since their inner workings are and will likely remain a“black box.”84 The Chinese governments early concern with LLM“safety”(安全,which also means“security”)should be understood in this context.85 So h

114、ow do alternative models improve things from the states perspective?BIGAI director Zhu Songchun,whose influence extends well past his Beijing circle of colleagues,86 has an answer.Zhu argues that for an AI to be general it must ingest principles that guide its exploration of the environment in which

115、 it is embedded.In Zhus system,an AI:(1)must manage unlimited tasks including those not predefined;(2)have autonomy(自主性),including the ability to generate its own tasks;and(3)be“value-driven,”not“data-driven”as in todays large models.87 Zhu correctly notes that current LLMs“do not have human cogniti

116、ve and reasoning capabilities,and also lack human emotions and values.”Center for Security and Emerging Technology|17 “From the perspective of values,whether large models can understand human value orientation determines whether large models can be safely and reliably applied to important areas rela

117、ted to the national economy and peoples livelihood.”Hence“values”are needed not only to drive the system as it learns but to ensure that what it does learn meets the needs of the nation and the people.As Zhu explains:The core of peoples concerns about the threat of artificial intelligence is their d

118、istrust of“big models.”There are two levels of trust.The first is trust in the systems ability.The second is recognition of values.The core of trust between people is value identity.88 Zhus test of a general AI system,on which BIGAI is founded,besides evaluating vision,language,cognition,movement,an

119、d learning also assesses its adherence to embedded values along five dimensions:elementary self-value,advanced self-value,“primary social value,advanced social value,and group value.89 The difficulty of guaranteeing that LLM outputs are consistent with a particular set of values has also been recogn

120、ized in the western literature.Earlier,ham-fisted approaches to ensuring that LLM outputs were aligned with specific sets of values have been widely ridiculed.90 At the core of the challenge of“alignment”is the absence of an explicit“moral engine”in LLMs.This has forced developers to resort to labor

121、ious“fine-tuning”of LLM parameters based on human feedback on problematic LLM-generated responses.91 This approach based on tweaking undesirable or“non-aligned”answers and hoping for results that generalize to novel prompts has no guarantee of success.92 A case in point is a recent study presenting

122、the same ethical dilemmas in different languages to different LLMs.93 That study found widely divergent behavior for ethical decisions not just across LLMs,but even for the same LLM when presented with the same ethical dilemma in different languages.New approaches such as training not based on human

123、 feedback but using different value models might be suitable for specific,well-defined scenarios,94 but it is unclear how such a strategy could generalize to broader sets of ethical dilemmas,let alone lead to LLM responses that are consistent with a specific set of values.It is therefore currently f

124、ar from clear if and how particular sets of values can be“trained into”an LLM,given that,for LLMs,“good”and“bad”are just words to be predicted,with no grounding in any kind of valuative framework.In the end,Zhus argument for an alternative GAI approach cuts three ways:Center for Security and Emergin

125、g Technology|18 Zhu claims that LLMs cannot achieve GAI because they lack human-like sensibilities that underlie the motivation to explore and learn.A cognitive architecture able to assimilate values is needed,in this view,for AI to achieve generality.He proposes to manage the AI safety problem by r

126、eplacing largely unworkable“guardrails”and ad hoc fixes to LLM output with an internal requirement to behave according to first principles,i.e.,a value system matching that of its users.He addresses the fear of Chinas ruling elite that large models will subvert the autocracy.Driven by socialist Part

127、y values,the GAI that China needs to stay competitive remains within comfortable bounds and reinforces State ideology,potentially forever.Given these considerations,the AI“race”takes on a new dimension with challenges not only in the economic and military spheres but in human value orientation.A fin

128、al distinction between Chinese and western AI research evidenced throughout this paper needs to be made explicit,namely,the real possibility that Chinas directed,strategic approach may be more effectiveall other things being equalthan the western profit-driven approach that focuses on quick wins at

129、the possible expense of more successful strategies that require a longer time horizon.95 Our recommendations,accordingly,are twofold:(1)replace the monoculture of LLM research with government and institutional support for a multifaceted approach to general AI,and(2)take seriously the need to monitor

130、 Chinese technical developments through open sources.96 Center for Security and Emerging Technology|19 Authors William C.Hannas is CSETs lead analyst and formerly the CIAs senior expert for China open-source analysis.He is currently focused on U.S.-China technology competition,community outreach,and

131、 data discovery methodologies.Huey-Meei Chang is CSETs senior China S&T specialist,co-editor of Chinese Power and Artificial Intelligence:Perspectives and Challenges(Routledge,2023),and co-author of several papers on Chinas AI development.Maximilian Riesenhuber,PhD,is professor of neuroscience at Ge

132、orgetown University and codirector of its Center for Neuroengineering.His research focuses on understanding brain function and how these insights can be translated to augmented cognition applications and neuromorphic AI.Daniel H.Chou is a data scientist at CSET.He has collected,enhanced,and analyzed

133、 data for multiple studies on China AI and technology development while supporting government and private sector projects.Acknowledgements The authors are grateful to CSETs Helen Toner,who served as“red teamer,”and to John Chen of the RAND Corporation and Dr.Mike Wolmetz of Johns Hopkins Universitys

134、 Applied Physics Laboratory for serving as outside reviewers.The authors are also grateful to CSETs Dr.Igor Mikolic-Torriera,Dr.Catherine Aiken,Matthew Mahoney,Shelton Fitch,and Ben Murphy for their generous support during the review and publication process.2025 by the Center for Security and Emergi

135、ng Technology.This work is licensed under a Creative Commons Attribution-Non Commercial 4.0 International License.To view a copy of this license,visit https:/creativecommons.org/licenses/by-nc/4.0/.Document Identifier:doi:10.51593/20230048 Center for Security and Emerging Technology|20 Endnotes 1 Wm

136、.C.Hannas,Huey-Meei Chang,Catherine Aiken,Daniel Chou,“China AI-Brain Research,”(Center for Security and Emerging Technology,September 2020),50,https:/cset.georgetown.edu/publication/china-ai-brain-research/.2 We avoid the term“artificial general intelligence”(AGI)here due to its historical associat

137、ion with features of biological brains and minds(such as emotional and social intelligence,affect,consciousness)that may or may not be necessary elements of a“general”AI.See Wm.C.Hannas,Huey-Meei Chang,Daniel Chou,Brian Fleeger,“Chinas Advanced AI Research,”(CSET,July 2022),1-3,for a description of

138、the terminological ambiguities with“AGI”overall and as relates to China,https:/cset.georgetown.edu/publication/chinas-advanced-ai-research/.The standard Chinese term 通用人工智能 equates literally to“general artificial intelligence.”A summary of recent Chinese policy documents issued by the Centre for the

139、 Governance of AI notes editorially that“通用人工智能 is sometimes,depending on context,translated as artificial general intelligence(AGI),sometimes as general artificial intelligence,and sometimes as general-purpose artificial intelligence.”This description gels with our own observations.Fynn Heide,“Beij

140、ing Policy Interest in General Artificial Intelligence Is Growing,”(Centre for the Governance of AI,June 8,2023),https:/www.governance.ai/post/beijing-policy-interest-in-general-artificial-intelligence-is-growing.3 Geoffrey Hinton,conversation with Joel Hellermark,April 2024,https:/ Robert Hart,“Met

141、as AI Chief:AI Models Like ChatGPT Wont Reach Human Intelligence,”Forbes,May 22,2024,https:/ Ben Goertzel,The Consciousness Explosion(prepublication copy,2024),12.“This sort of technology large language models,on its own,seems clearly not capable of producing HLAGI human-level AGI but it does seem v

142、ery promising as a component of integrated multi-module AGI systems.”Also see,TWIML AI Podcast(09:03),https:/ 2023 venture-capital investors poured over$36bn into generative AI,more than twice as much as in 2022.”Guy Scriven,“Generative AI Will Go Mainstream in 2024,”The Economist,November 13,2023,h

143、ttps:/ Jon Kleinberg and Manish Rashavan,“Algorithmic Monoculture and Social Welfare,”PNAS 118,no.22(February 2021),https:/www.pnas.org/doi/10.1073/pnas.2018340118.8 Hannas,Chang,Chou,and Fleeger,“Chinas Advanced AI Research.”9 Huey-Meei Chang,“Chinas Bid to Lead the World in AI,”The Diplomat,July 6

144、,2024,https:/ for Security and Emerging Technology|21 10“Big data,small task”refers to the design philosophy of LLMs,which are trained on large datasets(on the order of trillions of“tokens”)with a simple task,viz.to predict the next word in a text.This is in contrast to traditional AI systems traine

145、d on specific tasks such as recognizing faces,with architectures optimized for the problem domainwhich enabled these systems to learn from smaller datasets.11 Note that the scope of this report is limited to text-based models.Yet,all considerations regarding text-based LLMs as paths to GAI also dire

146、ctly apply to their multimodal extensions that process not just text but also images,video and audio,which are differences in input/output modalities that do not fundamentally augment the models“intelligence.”12 Ashish Vaswani,Noam Shazeer,Niki Parmar,Jakob Uszkoreit,et al.,“Attention Is All You Nee

147、d,”arXiv preprint arXiv:1706.03762v7(2023).13 Kyle Mahowald,Anna A.Ivanova,Idan A.Blank,Nancy Kanwisher,et al.,“Dissociating Language and Thought in Language Models,”Trends in Cognitive Sciences 28,no.6(March 2024);Boshi Wang,Xiang Yue,Huan Sun,“Can ChatGPT Defend its Belief in Truth?Evaluating LLM

148、Reasoning via Debate,”Findings of the Association for Computational Linguistics,EMNLP(2023).14 Nouha Dziri,Ximing Lu,Melanie Sclar,Xiang Lorraine Li,et al.,“Faith and Fate:Limits of Transformers on Compositionality,”37th Conference on Neural Information Processing Systems,NeurIPS(2023).15 E.g.,Wolfr

149、am GPT.See https:/ S.M.Rivera,A.L.Reiss,M.A.Eckert and V.Menon,“Developmental Changes in Mental Arithmetic:Evidence for Increased Functional Specialization in the Left Inferior Parietal Cortex,”Cerebral Cortex 15(November 2005).17 Geoffrey Hinton,conversation with Joel Hellermark.18 Winnie Street,Jo

150、hn Oliver Siy,Geoff Keeling,Adrien Baranes,et al.,“LLMs Achieve Adult Human Performance on High-order Theory of Mind Tasks,”arXiv preprint arXiv:2405.18870v2(2024).19 Tomer D.Ullman,“Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks,”arXiv preprint arXiv.2302.08399v5(2023).20

151、 Jie Huang,Xinyun Chen,Swaroop Mishra,Huaixiu Steven Zheng,et al.,“Large Language Models Cannot Self-Correct Reasoning Yet,”arXiv preprint arXiv:2310.01798v1(2023).21 Lexin Zhou,Wout Schellaert,Fernando Martinez-Plumed,Yael Moros-Daval,et al.,“Larger and More Instructable Language Models Become Less

152、 Reliable,”Nature,September 25,2024.22 Chaeffer,Brando Miranda,Sanmi Koyejo,“Are Emergent Abilities of Large Language Models a Mirage?”37th Conference on Neural Information Processing Systems,NeurIPS(2023).23 Adrian de Wynter,“Awes,Laws,and Flaws from Todays LLM Research,”arXiv preprint arXiv:2408.1

153、5409v2(2024).Center for Security and Emerging Technology|22 24 While LLMs excel on standardized tests that are similar to those in its massive training set(e.g.,the SAT or the bar exam),recent research has shown that figuring out novel tasks based on one or two examplessomething humans are quite goo

154、d atcontinues to present a challenge for LLMs.For instance,the influential“Abstraction and Reasoning Corpus”notes that humans can solve an average of 80%of ARCs pattern matching tasks,while the best AI systems perform at around 30%,https:/lab42.global/arc/.25 Goertzel,“Generative AI vs.AGI:The Cogni

155、tive Strengths and Weaknesses of Modern LLMs,”arXiv preprint arXiv:2309.10371v1(2023),63.26 Doug Lenat and Gary Marcus,“Getting from Generative AI to Trustworthy AI:What LLMs Might Learn from Cyc,”arXiv preprint ArXiv:2308:04445(2023).27 Julia Angwin,“Press Pause on the Silicon Valley Hype Machine,”

156、The New York Times,May 15,2024,https:/ Dwarkesh Patel,“Francois Chollet,Mike Knoop LLMs Wont Lead to AGI-$1,000,000 Prize to Find True Solution,”Dwarkesh Podcast,June 11,2024.Chollet added“I see LLMs as more of an off-ramp on the path to AGI actually.All these new resources are actually going to LLM

157、s instead of everything else they could be going to.”https:/ Liu Yangnan(劉楊楠),“Before Achieving AGI,Global AI Leaders Are Arguing over These 4 Key Issues”(實現 AGI 之前，全球 AI 大佬在這 4 個關鍵問題上吵起來了),S(搜狐),June 13,2023,https:/ Zhipu(智譜)was founded in 2019,has a staff of 800,and a market valuation of$2.5 billi

158、on as of May 2024.It ranks at the top of Chinas 260 generative AI companies.Eleanor Olcott,“Four Start-ups Lead Chinas Race to Match OpenAIs Chat GPT,”Financial Times,May 2,2024,https:/ BAAI(北京智源人工智能研究)was founded in 2018 by Huang Tiejun(see below),vice-dean at Peking Universitys Institute for Artif

159、icial Intelligence(人工智能研究院),with a goal of building strong AI.https:/ See Tang Jies CV at https:/ Celeste Biever,“Chians ChatGPT:Why China Is Building Its Own AI Chatbots,”Nature,May 22,2024,https:/ A rule of thumb,similar to Moores Law for semiconductor density,which holds that the power of a neura

160、l network increases in proportion to its size,training data,and computational resources.Center for Security and Emerging Technology|23 35“Tsinghua Universitys Tang Jie:From GPT to GPT Zero Will Be a Major Milestone This Year”(清華大學唐杰:從 GPT 到 GPT Zero 會是今年重大階段性成果),Tencent Network(騰訊網),March 1,2024,htt

161、ps:/ Tang Jie(唐杰),“Big Models and Superintelligence”(大模型與超級智能),Communications of CCF(中國計算機學會通訊)20,no.6,(2024),also see https:/ Chinese orthography lacks word division,i.e.,white space between words.Accordingly,these distinctions must be made on the fly,complicating the process of nominating the“toke

162、ns”on which LLMs operate.38 Li Anqi(李安琪),“Kai-Fu Lee and Ya-Qin Zhang Fireside Chat:Chinas Big Models Have More Opportunities on the C-end,and Technology Will Not Bring Permanent Leadership”(李開復、張亞勤爐邊談話:中國大模型在 C 端的機會更多,技術不會帶來永久領先),Tencent Technology(騰訊科技),June 14,2024,https:/ Liu Yangnan,“Before ach

163、ieving AGI.”40“Take-Aways from WAIC Keynote Speeches by Leading Figures at the Science Frontier Conference”(WAIC 科學前沿會議大佬演講干貨!),(智東西),July 6,2024,https:/ Tiejun,the Earliest Promoter of Chinas Large-scale Model:The World May Only Need Three LLM Ecosystems”(中國大模型的最早推行者黃鐵軍：全球可能只需要三個大模型生態),Tencent Tech

164、nology(騰訊科技),June 9,2023,https:/ CASs Institute of Automation is one of Chinas top institutes developing GAI through LLM and brain-inspired research.Its scientists are tied with Peking University for the highest number of GAI-related studies.Wm.C Hannas,Huey-Meei Chang,Max Riesenhuber,and Daniel H.C

165、hou,“Chinas Cognitive AI Research:Emulating Human Cognition on the Way to General Purpose AI,”(Center for Security and Emerging Technology,July 2023),11,https:/cset.georgetown.edu/publication/chinas-cognitive-ai-research/.43 CEBSIT is an umbrella organization for some 39 research institutes in China

166、.See http:/ Bo Xu and Muming Poo,“Large Language Models and Brain-inspired General Intelligence,”National Science Review,October 2023,https:/ Chang and Hannas,“Spotlight on Beijing Institute for General Artificial Intelligence.”46 Zhu Songchun(朱松純),“Zhu Songchun:Will the Rapid Development of Artific

167、ial Intelligence Definitely Pose a Threat?”(朱松純:人工智能高速發展一定會產生威脅嗎?),ScienceNet(科學網),September 12,2023,https:/ Yangmei(韓揚眉),“Zhu Songchun:20 Years of Exploration Has Given China an Advantage in Moving towards the Era of Center for Security and Emerging Technology|24 General Artificial Intelligence”(朱松

168、純:20 年探索,為我國邁向通用人工智能時代贏得先機),China Science Daily(中國科學 報),July 25,2024,https:/ See Zeng Yis home page at https:/braincog.ai/yizeng/.See also https:/brain-work/cn.48“The Chinese Academy of Sciences Team Is Determined to Invest in the Next 20 Years or Even Longer to Build a General Brain-like Artificial

169、 Intelligence Infrastructure”(中科院團隊:決心投入未來 20 年乃至更長時間,打造通用類腦人工智能基礎設施),Science&Technology Review(科技導報),October 10,2022,https:/ Xiangyang:Rethinking the Human-machine Relationship in the Era of General Purpose Large Models”(沈向洋:通用大模型時代 重新思考人機關系),EastM(東方財富網),March 23,2024,https:/ president Zheng Qingh

170、ua:Big Models Have Become the Pinnacle of Current Artificial Intelligence,but They Still Have Four Major Flaws”(同濟校長鄭慶華:大模型已成當前人工智能巔峰,但存四大缺陷),Sohu(搜狐),April 23,2024,https:/ Comment:Big Models and General Artificial Intelligence|Turing Conference SIGAI Roundtable Forum held in Wuhan”(焦點評論:大模型與通用人工智能|

171、圖靈大會 SIGAI 圓桌論壇在武漢舉行),Beijing Institute for General Artificial Intelligence,July 30,2024,https:/www.bigai.ai/blog/news/焦點評論：大模型與通用人工智能-圖靈大會 bigai 圓桌/.52 Several Measures for Promoting the Innovation and Development of General Artificial Intelligence in Beijing(北京市促進通用人工智能創新發展的若干措施),May 30,2023.https

172、:/ Zhaohui Appointed as vice president of the Chinese Academy of Sciences”(吳朝暉任中國科學院副院長),April 11,2024,http:/ Jiao Yifei(繳翼飛),“Wu Zhaohui,Vice Minister of the Ministry of Science and Technology:Big Models Push Artificial Intelligence toward the 3.0 Stage,and We Need to Explore the Development of Gen

173、eral Artificial Intelligence through Multiple Paths”(科技部副部長吳朝暉：大模型推動人工智能邁向 3.0 階段 要多路徑探索通用人工智能發展),21st Century Business Herald(21 世紀經濟報道),March 24,2024,https:/ The figures omit universities and government-sponsored institutes engaged in AI research in Haidian.56 Sun Ying(孫穎)and Wang Haixin(王海欣),“Bei

174、jing Releases Three-year Plan and Lays Out Six Major Actions to Build a National Embodied Intelligence Innovation Highland”(發布三年計劃、布局六大行動,北京打造全國具身智能創新高地),Beijing Daily(北京日報),April 27,2024,https:/ For example,LLMs by Zhipu,BAAI,iFlytek,Huawei,Baidu,Shanghai Artificial Intelligence Laboratory(上海人工智能實驗

175、室),Baichuan AI(百川智能),Moonshot AI(月之暗面),etc.with declared GAI goals.See“2023 H1 Chinas Top 50 Most Valuable AGI Innovation Institutions Officially Released”(2023 H1中 Center for Security and Emerging Technology|25 國最具價值 AGI 創新機構 TOP 50正式發布),Geek Park(極客公園),July 27,2023,https:/ Hannas,Chang,Riesenhuber

176、,and Chou,“Chinas Cognitive AI Research.”59 William C Hannas,Huey-Meei Chang,Rishika Chauhan,Daniel H.Chou,et al.,“Bibliometric Analysis of Chinas Non-Therapeutic Brain-Computer Interface Research,”(Center for Security and Emerging Technology,March 2024),https:/cset.georgetown.edu/publication/biblio

177、metric-analysis-of-chinas-non-therapeutic-brain-computer-interface-research/.60 CSET merged corpus of scholarly literature including Web of Science,OpenAlex,Semantic Scholar,The Lens,arXiv,and Papers With Code.Searches included English and Chinese variants of the terms.The LLM keywords were:大語言模型,大型

178、語言模型,large language model,LLM,GPT,LLaMA;the AGI keywords were:通用人工智能,artificial general intelligence,人工通用智慧,強人工智能 strong artificial intelligence,strong AI,AGI,GAI.61 Online sources were primarily Chinese institute and scientific research(科研)websites.62 These papers bylines typically list a half-doze

179、n or more authors.For brevity we provide only the lead author,last author(often a senior advisor),the corresponding author if different from the first and last authors,and finally all authors claiming a non-China affiliation(marked in italics).63 YONG Silongs pinyin spelling(“si”for 子)is idiosyncrat

180、ic.See https:/www.bigai.ai/blog/news/聯結場景理解和具身智能，首個智能體具身推理能/.64 Here are the scores:BAAI(3 papers),BIGAI(10),CASIA(5),other Beijing-located CAS institutes(10),Peking University(18)and Tsinghua University(11)for a total of 57 appearances versus 18 for all other Chinese institutes combined,i.e.,76 per

181、cent of the cited affiliations.Prior CSET analysis of Chinese GAI research using a different(and much larger)corpus showed Beijing-based institutes on 70 percent of the papers.See Hannas,Chang,Riesenhuber,and Chou,“Chinas Cognitive AI Research,”12.Some 6 foreign organizations were among the corpuss

182、cited affiliations,most prominently Carnegie Mellon University(3)and Microsoft Research Asia(2).65 Hannas,Chang,Aiken and Chou“Chinas AI-Brain Research;”Hannas,Chang,Chou and Fleeger“Chinas Advanced AI Research;”Hannas,Chang,Riesenhuber and Chou,“Chinas Cognitive AI Research;”Huey-Meei Chang and Wm.

183、C.Hannas,“Spotlight on Beijing Institute for General Artificial Intelligence,”(Center for Security and Emeriging Technology,May 2023),https:/cset.georgetown.edu/publication/spotlight-on-beijing-institute-for-general-artificial-intelligence/;Wm.C.Hannas,Huey-Meei Chang,Rishika Chauhan,Daniel H.Chou,e

184、t al.,“Bibliometric Analysis of Chinas Non-Therapeutic Brain-Computer Interface Research”(Center for Security and Emerging Technology,March 2024),https:/cset.georgetown.edu/publication/bibliometric-analysis-of-chinas-non-therapeutic-brain-computer-interface-research/.Center for Security and Emerging

185、 Technology|26 66 Leonardo de Cosmo,“Google Engineer Claims AI Chatbot Is Sentient:Why That Matters,”Scientific American,(July 2022),https:/ David J.Chalmers,“Could a Large Language Model Be Conscious?”arXiv preprint arXiv:2303.07103(2024).68 Tang Xiaojuan,Zhu Songchun,Liang Yitao,Zhang Muhan,“Large

186、 Language Models Are In-context Semantic Reasoners Rather than Symbolic Reasoners,”arXiv preprint arXiv:2305.14825v2(2023).69 Jason Wei,Xuezhi Wang,Dale Shuurmans,Maarten Bosma,et al.,“Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,”arXiv preprint arXiv:2201.11903(2022).70 Yih

187、e Deng,Weitong Zhang,Zixiang Chen,Quanquan Gu,“Rephrase and Respond:Let Large Language Models Ask Better Questions for Themselves,”arXiv preprint arXiv:2311.04205(2024).71 Shunyu Yao,Dian Yu,Jeffrey Zhao,Izhak Shafran,et al.,“Tree of Thoughts:Deliberate Problem Solving with Large Language Models,”ar

188、Xiv preprint arXiv:2305.10601(2023).72 Maciej Besta,Nils Blach,Ales Kubicek,Robert Gerstenberger,et al.,“Graph of Thoughts:Solving Elaborate Problems with Large Language Models,”arXiv preprint arXiv:2308.09687(2024).73 For a recent survey,see S.M.Towhidul Islam Tonmoy,S.M.Mehedi Zaman,Vinija Jain,An

189、ku Rani,et al.,“A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models,”arXiv preprint arXiv:2401.01313(2024).74 Sbastien Bubeck,Varun Chandrasekaran,Ronen Eldan,Johannes Gehrke,et al.,“Sparks of Artificial General Intelligence:Early Experiments with GPT-4,”arXiv prep

190、rint arXiv:2303.12712(2023).75 Brian Owens,“Rage against Machine Learning Driven by Profit,”Nature,September 18,2024,https:/ Yann LeCun,Yoshua Bengio and Geoffrey Hinton,“Deep Learning,”Nature,May 28,2015.77 Emre O.Neftci and Bruno B.Averbeck,“Reinforcement Learning in Artificial and Biological Syst

191、ems,”Nature Machine Intelligence,March 2019.78 Matthew Botvinick,Jane X.Wang,Will Dabney,Kevin J.Miller,et al.,“Deep Reinforcement Learning and Its Neuroscientific Implications,”Neuron,August 19,2020.79 David Silver,Aja Huang,Chris J.Maddison,Arthur Guez,et al.,“Mastering the Game of Go with Deep Ne

192、ural Networks and Tree Search,”Nature,January 28,2016.80 It was noteworthy that a number of papers surveyed emphasized the potential of“human-like”or“brain-like”algorithms to overcome shortcomings of current AI systems,given the brains status as the past and current“gold standard”for intelligent sys

193、tems.Center for Security and Emerging Technology|27 81 Geoffrey Hinton,conversation with Joel Hellermark,April 2024,https:/ Joyce Guo,Zoey Duan,Chance Flanigen,Ethan Hsu,et al.,“Energy Consumption Ramifications of U.S.-China AI Competition,”SSRN(May 2024),http:/dx.doi.org/10.2139/ssrn.4839772.83 As

194、discussed,for example,in Zou,et al.,“Controllable Generation from Pre-trained Language Models via Inverse Prompting”(item 24 in this papers Section 3).84“I honestly dont think we will get to an AI that we can trust if we stay on the current path LLMs are recalcitrant,and opaque by natureso-called“bl

195、ack boxes”that we can never fully rein in.”Gary Marcus,“OpenAIs Sam Altman Is Becoming One of the Most Powerful People on Earth.We Should Be Very Afraid,”The Guardian,August 3,2024,https:/ Cyberspace Administration of China,et al.,“Interim Measures for the Administration of Generative Artificial Int

196、elligence Services”(生成式人工智能服務管理暫行辦法),July 10,2023,https:/ Technical Committee 260 on Cybersecurity of Standardization Administration of China(SAC/TC260,全國信息安全標準化技術委員會),“Technical Documentation of National Technical Committee 260 on Cybersecurity of Standardization Administration of China:Basic Secur

197、ity Requirements for Generative Artificial Intelligence Services,”(全國網絡安全標準化技術委員會技術文件：生成式人工智能服務安全基本要求),October 11,2023,translated by CSET,https:/cset.georgetown.edu/publication/china-safety-requirements-for-generative-ai-final/.86 Chang and Hannas,“Spotlight on Beijing Institute.”87 Zhu Songchun,“Wi

198、ll the Rapid?”88 Zhu Songchun,“Will the Rapid?”Our emphasis.89 Yujia Peng,Jiaheng Han,Zhenliang Zhang,Lifeng Fan,et al.,“The Tong Test:Evaluating Artificial General Intelligence through Dynamic Embodied Physical and Social Interactions,”Engineering 34,(2024):12-22.90 Nico Grant,“Google Chatbots A.I.

199、Images Put People of Color in Nazi-Era Uniforms,”The New York Times,February 26,2024,https:/ The process is called“reinforcement learning from human feedback”(RLHF).92 Cf.the“whack-a-mole”challenge referred to in this papers section 4 in the context of efforts to reduce hallucinations or improve rea

200、soning abilities of LLMs.Center for Security and Emerging Technology|28 93 Utkarsh Agarwal,Kumar Tanmay,Aditi Khandelwal and Monojit Choudhury,“Ethical Reasoning and Moral Value Alignment of LLMs Depend on the Language We Prompt Them in,”arXiv preprint arXiv:2404.18460v1(2024).94 Elizaveta Tennant,S

201、tephen Hailes,and Mirco Musolesi,“Moral Alignment for LLM Agents,”arXiv preprint arXiv:2410.01639v1(2024).95 The authors are grateful to RANDs John Chen for this insight.96 Wang Yue(王悅),“After Nearly 1,500 Days of Wrestling with Large Models,Beijing Academy of Artificial Intelligence Is Still Insisting on Original Innovation”(與大模型交手近 1500 天，智源仍在堅持原始創新),Leiphone(雷峰網),June 20,2024,https:/ see https:/scout.eto.tech/?id=3558.