2025+中國學(xué)界對(duì)大語言建模的采購性思考人用工智能+AGI+的多渠道調(diào)查研究報(bào)告(英)-29正式版

ExecutiveSummarylanguagemodelshavegarneredinteresttotheirremarkable

to“generate”human-likeresponsestonaturallanguagequeries—athreshold

onetimewasconsidered“proof”ofsentience—performothertime-saving

tasks.Indeed,LLMsregardedbymanyorthe,pathwaytogeneralartificial

intelligence(GAI)—hypothesizedstatewherecomputers(orevenexceed)

skillsmostortasks.ThelureofAI’sholygrailthroughLLMsdrawninvestmentinthebillions

ofbythosefocusedonthistheUnitedStatesEuropeespecially,big

privatesectorcompaniesledthewayandtheirfocusonLLMsovershadowed

researchonotherapproachestoGAI,despiteLLM’sknowndownsidessuchcost,

powerconsumption,or“hallucinatory”output,deficitsinreasoning

abilities.thesecompanies’betsonLLMsfailtodeliveronexpectationsofprogress

GAI,westernAIdevelopersbepositionedtofallbackon

approaches.contrast,Chinafollowsastate-driven,diverseAIdevelopmentLiketheUnited

States,investsinLLMsbutsimultaneouslypursuestoGAI,

thosemoreexplicitlybrain-inspired.Thisreportdrawsonpublicstatements

bytopscientists,theirassociatedresearch,andongovernment

announcementstodocumentChina’smultifacetedTheChinesegovernmentsponsorsresearchtoinfuse“values”intoAIintendedto

guideautonomousprovideAIsafety,ensureChina’sAI

reflectstheneedsofthepeoplethestate.Thisreportconcludesbyrecommending

U.S.governmentsupportforalternativegeneralAIprogramsforcloserscrutinyof

AIresearch.CenterforSecurityEmergingTechnology|1Introduction:GenerativeAIandAIAchievinggeneralartificialintelligenceorGAI—definedAIthatreplicatesor

exceedsmostcognitiveacrossaoftasks,suchimage/video

understanding,continuallearning,planning,reasoning,skilltransfer,andcreativity1—is

akeystrategicgoalofintenseresearcheffortsbothinChinatheUnitedStates.2

Thereisvigorousdebateintheinternationalscientificcommunityregardingwhichth

leadtoGAImostquicklywhichpathsbestarts.theUnitedStates,

LLMsdominatedthediscussion,yetquestionsremainabouttheirabilityto

achieveGAI.SincechoosingthepathcanpositiontheUnitedStatesa

strategicdisadvantage,thisraisestheurgencyofalternativeapproaches

othercountriesbepursuing.theUnitedStates,expertsbelievethesteptoGAIwilloccur

therolloutofnewversionsofLLMssuchasOpenAI’so1,Google’sGemini,

Anthropic’sClaude,Meta’sLlama.3Othersargue,pointingtopersistentproblems

suchLLMhallucinations,mountofcompute,feedback,ormultimodaldata

sourcesLLMstoachieveGAI.4StillotherAIscientistsseerolesforLLMsin

GAIplatformsbutnottheonly,orevenmain,component.5PonderingthequestionofhowGAIcanbeachievedisimpobecauseittoucheson

optionsavailabletodeveloperspursuingAI’straditionalholygrail—human-level

intelligence.thepath—orapath—toGAIacontinuationofLLMdevelopment,

possiblyaugmentedbymodules?OrLLMsadeadend,necessitating

other,differentapproachesbasedonacloseremulationof

cognitionandbrainfunction?GiventhesuccessofLLMs,thelevelsofinvestment,6endorsementsbyregarded

AIscientists,optimismcreatedbyexamples,thedifficultyofreimagining

newapproachesinthefaceofmodelsinwhichcompaniesgreatcommitment,itis

easytooverlooktheofrelyingona“monoculture”basedonasingleresearch

m.7therearelimitationstoLLMsdeliver,withoutasufficiently

diversifiedesearchportfolio,itisunclearhowwellwesterncompaniesgovernmentsbetopursueothersolutionscanovercomeLLMsproblems

pathwaystoGAI.AdiversifiedresearchportfolioispreciselyChina’sapproachtoitsstate-sponsored

goalofachieving“generalartificialintelligence”通用人工智能.8Thisreportshow

that—inadditiontoChina’sknownprodigiousefforttofieldChatGPTLLMs,9—significantresourcesaredirectedinChinaalternativepathwaystoGAIbyCenterforSecurityEmergingTechnology|2scientistshavewell-foundedconcernsaboutthepotentialof“bigsmall

task”大數(shù)據(jù),小任務(wù))approachestoreachhumancapabilities.10Accordingly,thispaperaddressesquestions:criticismsdoChinesescientists

ofLLMstogeneralAI?howisChinamanagingLLMs’alleged

shortcomings?Thepaperbegins(section1)critiquesbyprominentnon-ChinaAIscientistsof

languagemodelstheirtosupportGAI.Thesectionprovidescontext

forviewsofChinesescientiststowardLLMs(section2)describedinsources.

Section3thencitesresearchsupportsChina’spublic-facingclaimsaboutthenon-

viabilityofLLMsapathtoGAI.section4,assesstheseclaimsasafor

recommendationsinsection5onChina’salternativeprojectsmustbetaken

seriously.CenterforSecurityEmergingTechnology|3LargeLanguageModelsTheirCriticsThetermlanguagemodel”capturestheylargenetworkstypically

billionstotrillionsofparameters,theytrainedonnaturallanguage,

terabytesoftextingestedfromtheinternetothersources.LLMs,neural

networks(NN)generally,typologicallydistinctfrom“goodoldfashioned”(GOFAI)

symbolicAIthatdependsonrule-basedcoding.addition,today’slargemodelscan

todifferentdegrees,multimodalinputsoutputs,includingimages,video,

audio.11LLMsdebutedin2017,whenGoogleengineersproposedaNNarchitecture—a

transformer—optimizedtopatternsinsequencesoftextbytoattention”tothecooccurrencerelationshipsbetween“tokens”orof

words)inthetrainingcorpus.12Unlikehumanknowledge,knowledgecapturedinLLMs

isnotobtainedthroughinterionsthenaturalenvironmentbutdependson

probabilitiesderivedfromthepositionalrelationshipsbetweenthetokensin

sequences.MassiveexposuretocorporatrainingallowstheLLMtoidentify

regularitiesintheaggregate,beusedtogenerateresponsestopromptsafterthetraining.Hence,theOpenAIproduct“GPT”(generativepre-

trainedtransformer).TheofLLMsto“blend”differentsourcesofinformation(whichplaysto

strengthsofneuralnetworksinpatternmatchinganduncovering

similaritiesincomplexspaces)hasgiventoapplicationsindiversetext

summarization,translation,codetheoremproving.Yet,itbeenhotlydebatedwhetherthisabilitytoexploitregularitiesis

sufficienttoachieveGAI.Initialenthusiasticreportsthe“sentience”ofLLMs

increasinglysupplementedbyreportsshowingseriousdeficitsinLLMs’abilityto

understandlanguagetoreasoninahuman-like.13SomepersistentinLLMs,inbasicmath,14ppearcorrectablebys,15

i.e.,externalprogramsspecializedforofLLMeaknesses.such—ofanetworkofsystemsspecializedindifferentaspectsofcognition—

wouldbemorelikethewhichhasdedicatedmodules,e.g.,forepisodicmemory,

reasoning,etc.,ratherthanasingleprocessinLLMs.16SomescientistshopeincreasesincomplexityhelpovercomeLLMs’

defects.ForGeoffreyHinton,creditingintuitionofIlyaSutskever(OpenAI’s

formerchiefscientist,studiedHinton),believessolvesomeof

theseproblems.thisview,LLMs“reasoning”byvirtueoftheirability“toCenterforSecurityEmergingTechnology|4predictthenextsymbolpredictionisaprettytheoryofhowtheis

g.”17Indeed,increasesincomplexity(fromGPT-2throughGPT-4)ledto

increasedpeonvariousbenchmarktasks,such“theoryof18

aboutmentalstates),deficitswerenotedforGPT-3.5.19Othersuchdeficitsarehardertoaddressandpersistdespiteincreasesinmodel

complexity.Specifically,“hallucinations,”i.e.,LLMsmakingincorrectclaims(aproblem

inherenttoneuralnetworksthataredesignedtointerpolateunlikethebrain,do

notseparatethestorageoffrominterpolations)errorsinreasoningbeendifficulttoovercome20recentstudiesthatthelikelihoodof

incorrect/hallucinatorysweincreasesgreatermodelcomplexity.21addition,thestrategyofincreasingmodelcomplexityinthehopeofhievingnovel,

qualitativelydifferent“emergent”behaviorsappearonceacomputational

thresholdbeencrossedlikewisebeencalledintoquestionbyresearch

thatpreviouslynoted“emergent”inmodelswereartefactsof

themetricsusednotindicativeofanyqualitativeinmodelperformance.22

Correspondingly,claimsof“emergence”inLLMsdeclinedintherecentliterature,

evenmodelcomplexitiesincreased.23Indeed,thereisthejustifiedconcerntheighperformanceofLLMson

standardizedtestscouldbeascribedmoretothewell-knownpatternmatching

prowessofneuralnetworksthanthediscoveryofnewstrategies.24StillotherofLLMscenteronfundamentalcognitiveandphilosophicalissues

suchtheabilitytogeneralize,formdeepabstractions,create,self-direct,modeltime

space,showcommonsense,reflectontheirownoutput,25manageambiguous

expressions,unlearnbasedonnewinformation,evaluateproconarguments

decisions),graspnuance.26thesedeficitsdiscussedinthewesternresearchliterature,others

suchLLMs’inabilitytoeasilyknowledgebeyondthecontextwindowwithout

thebasemodel,orthecomputationalenergydemandsofLLM

mostcurrentinvestmentofcommercialplayersintheAIspace(e.g.,OpenAI,

Anthropic)iscontinuingdownthissameroad.Theproblemisnotonly“weinvestinginidealfuturemaynotmaterialize”27butratherLLMs,inGoogle

AIresearcherFranoisChollet’swords,“suckedtheoxygenoutoftheroom.Everyone

isdoingLLMs.”28CenterforSecurityEmergingTechnology|5ChineseViewsofasaPathtoGeneralAI(orNot)AreviewofstatementsbyscientiststopAIresearchinstitutes

revealsahighdegreeofskepticismaboutLLMs’tolead,bythemselves,toGAI.

Theseresemblethoseofinternationalexperts,becausebothgroupsthe

problemsbecauseChina’sAIexpertsinteractwiththeirglobalpeersa

matterofcourse.29HerefollowseveralChinesescientists’viewsonLLMsapathtogeneralTang唐杰)isprofessorofcomputerscienceTsinghuaUniversity,thefounderof

智譜),30aleadingfigureintheAcademyofIntelligence(BAAI),31

thedesierofseveralindigenousLLMs.32Despitesuccessstatistical

models,argueshuman-levelAIrequiresthemodelstobe“embodiedinthe

d.”33Althoughbelievesthescalinglaw(規(guī)模法則34“stillalongwaytogo,”

onedoesnotguaranteeGAIwillbeachieved.35Amofruitfulpathwouldtake

cuesfrombiology.hiswords:“GAIormachineintelligencebasedonlargemodelsdoesnotnecessarilytobe

thethemechanismofhumanbraincognition,butanalyzingtheofthebrainmaybettertherealizationofGAI.”36ZhangYaqin張亞勤,AKAYa-QinZhang)co-foundedMicrosoftResearchAsiisthe

formerpresidentoffoundingdeanofTsinghua’sInstituteforAIIndustry

Research智能產(chǎn)業(yè)研究院)aadvisor.ZhangcitesthreeproblemsLLMs,

namely,theirlowcomputationalefficiency,inabilityto“trulyunderstandthephysical

world,”socalled“boundaryissues”邊界問題i.e.,tokenization.37Zhangbelieves

Goertzel)“weneedtoexplorehowtocombinegenerativeprobabilistic

modelsexisting‘firstprinciples’[ofthephysicalworld]orrealmodelsand

knowledges.”38HuangTiejun黃鐵軍)isfounderformerdirectorofandvicedeanofPeking

University’sInstituteforIntelligence(人工智能研究院Huangnames

threetoGAI:“informationmodels”basedonbigdatabigcompute,

“embodiedmodels”trainedthroughreinforcementbrainemulation—in

astake.39HuangagreesLLMscalinglawswillcontinueto

operatebut“itisnotonlyecessarytocollectstaticdata,butalsotoobtainprocessmultiplesensoryinformationinrealtime.”40Inview,GAIdependson

integratingstatisticalmodelsbrain-inspiredAIandembodiment,CenterforSecurityEmergingTechnology|6LLMsrepresent“staticemergencebasedonbigdat”是基于大數(shù)據(jù)的靜態(tài)涌現(xiàn).

Brain-inspiredintelligence,bycontrast,isbasedoncomplexdynamics.Embodied

intelligencediffersinthatitgeneratesnewabilitiesbyinteractingthe

environment41Bo徐波,deoftheSchoolofArtificialIntelligenceUniversityofChinese

AcademyofSciences(UCAS)中國科學(xué)院大學(xué)人工智能學(xué)院)directorofthe

ChineseAcademyofSciences(CAS)InstituteofAutomation(CASIA,中國科學(xué)院自動(dòng)化

研究所,42Muming蒲慕明,AKAMumingPoo),directorofCAS’sCenterfor

ExcelleinBrainScienceIntelligenceTechnology腦科學(xué)與智能技術(shù)卓

越創(chuàng)新中心43believeembodimentenvironmentalinteractionfacilitateLLMs’

growthtodGAI.AlthoughtheartificialneuralnetworksonwhichLLMsdepend

wereinspiredbybiology,theybyadding“moreneurons,layersconnections”

donotbegintoemulatethecomplexityofneurontypes,selective

connectivity,modularstructure.particular,“Computationallycostlybackpropagationalgorithms…couldbeimprovedoreven

replacedbyplausiblealgorithms.”Thesecandidatesinclude

spiketimesynapticplasticity,“neuromodulatordependentmetaplasticity”“short-

termvs.long-termmemorystoragerulessetthestabilityofsynaptics.”44ZhuSongchun朱松純,AKASong-deanofPKU’sInstituteofIntelligencedirectoroftheInstituteforGeneralArtificialIntelligence北京

通用人工智能研究院)foundedonthepremisebigdata-basedLLMsa

dead-endintermsoftheirtoemulatehuman-levelcognition.45pullspunches:“Achievinggeneralartificialintelligenceistheoriginalintentionandultimategoalof

artificialintelligenceresearch,butcontinuingtoexpandtheparameterbasedon

existinglargemodelscannotachievegeneralartificialintelligence.”comparesChina’sLLM’sachievementstoMt.Everest”whenthereal

goalistoreachthemoon.Inview,LLMs“inherentlyuninterpretable,ofdataleakage,donotacognitivearchitecture,lackcausalandmathematical

reasoningcapabilities,otherlimitations,sotheyleadto‘generalartificial

intelligence’.”46ZengYi曾毅,directorofCASIA’sBrain-inspiredCognitiveIntelligence類腦認(rèn)知

智能實(shí)驗(yàn)室foundingdirectorofitsInternationalResearchCenterforAIEthicsCenterforSecurityEmergingTechnology|7Governance,47isbuildingaGAIplatformbasedontime-dependentspikingneural

networks.hiswords:“Ourbraincognitiveintelligenceteamfirmlybelievesonlybythe

structureofthebrainitsintelligentwellthelawsofnaturalevolution,achieveintelligenceistruly

meaningfulbeneficialtohumans.”48ofLLMsbyotherChineseAIscientistslegion.?ShenXiangyang沈向洋,HarryShumAKAHeung-YeungShum),former

MicrosoftexecutiveVPdirectoroftheAcademicCommitteeofPKU’s

InstituteofIntelligence,lamentsAIresearch“clear

understandingoftheofintelligence.”Shensupportsaviewattributes

toNewYorkUniversityprofessoremeritusLLMcriticMarcusthat“no

matterhowChatGPTdevelops,thecurrenttechnicalroutenotbeto

usrealintelligence.”49?Qinghua(鄭慶華presidentofTongjiUniversityaChineseAcademy

ofEngineeringacademician,statedthatLLMshaveflaws:theyconsume

toomuchdatacomputingresources,susceptibletocatastrophic

forgetting,logicalreasoningcapabilities,donotknowwhenthey

ortheyareg.50?LiWu李武directoroftheStateKeyLaboratoryofCognitiveNeuroscienceBeijingNormalUniversity,statedhisbelief“currentneural

networksrelativelyspecializeddonotconformtothethehuman

works.youdesperatelyhypethemodelitselfonlyfocusonthe

expansionofparametersfrombillionsortensofbillionstohundredsofbillions,

younotbetoachievetrueintelligence.”51RecognitionoftheneedtosupplementLLMresearchwithalternativetoGAIisevidencedinstatementsbynationalandmunicipalgovernments.On30,2023,citygovernment—whosejurisdictionmuchof

GAI-orientedLLMresearchisplace—issuedastatementcallingfor

developmentofmodelsothergeneralartificialintelligencetechnology

systems”系統(tǒng)構(gòu)建大模型等通用人工智能技術(shù)體系.52Sectionthreefiveitems(7-

11),thefirstfourofwhichpertaintoLLMs(algorithms,trainingdata,evaluation,a

softwarehardwaresystem).Item11reads“exploringnew新路徑)for

generalartificialintelligence”andcallsfor:CenterforSecurityEmergingTechnology|8Developingabasictheoreticalsystem基礎(chǔ)理論體系)forGAI,autonomous

collaborationdecision-making,embodiedintelligence,brain-inspired類腦)

intelligence,supportedbyaunifiedtheoreticalframework,ratingandtesting

programminglanguages.Embodiedsystems(robots)[trainopenenvironments,generalizedscenarios,continuoustasks.Themandatesthefollowing:“Supporttheexplorationofbrainintelligence,studytheconnectionpatterns,

codingmechanisms,informationprocessingandothercoretechnologiesofneurons,inspirenewartificialneuralnetworkmodelingandtrainingmethods.”AlternativestoLLMswerecitedthenationallevelin2024,whenvice

presidentWu吳朝暉,formerlyviceministerofChina’ssciencepresidentofUniversity),53statedAIismovingtoward“synergybetween

andsmallmodels”大小模型協(xié)同,addingChinamust“explorethe

developmentofGAIinmultipleways”多路徑地探索通用人工智能發(fā)展Thelatter

“embodiedintelligence,distributedgroupintelligence,humanhybrid

intelligence,enhancedintelligence,autonomousdecisionmaking.”54ThefollowingmonthHaidianDistrictgovernment,sdictionover1,300

AIcompanies,morethan90ofdevelopingbigmodels,55issuedathree-year

tofacilitateresearchinembodied具身)AI.Thedefines“embodiment”“theofintelligentsystemormachinetointeracttheenvironmentinreal

timethroughperceptioninteraction”andismeanttoserveaplatformfor

development.Itsdetailsplansforhumanoidrobotsfacilitatedby

replicatingbrainfunctionality.56OuranalysisofpublicstatementsbygovernmentinstitutionsandrankingChineseAI

scientistsindicatesinfluentialpartofChina’sAIcommunitysharestheconcerns

misgivingsheldbywesternofLLMsseeksalternativepathwaysto

generalartificialintelligence.CenterforSecurityEmergingTechnology|9本報(bào)告來源于三個(gè)皮匠報(bào)告站（）,由用戶Id:349461下載,文檔Id:611736,下載日期:2025-02-17WhatDoesAcademicRecordstatementsbyscientistsonemeasureofapproachtoGAI.Anotheris

theirrecordofscholarship.reviewsofChineseliteraturedetermined

ChinaispursuingGAIbymultiplemeans,includinggenerativelanguage

models,57brain-inspiredmodels,58byenhancingcognitionthroughbrain-computer

interfaces.59OurpresenttaskistotheliteratureforevidenceChinese

—beyondwhatpositivefeaturesbrain-basedmodelshave—drivento

seekalternativebyLLM’sshortcomings.end,rankeywordsearchesinChineseEnglishfor“AGI/GAI+

LLM”theircommonvariantsinCSET’sMergedCorpus60forpaperspublishedin

2021orlaterprimaryChineseauthorship.Some35documentswereA

separatequeryweb-basedsearchesrecovered43morepapers.6115ofthe78

paperswererejectedbythestudy’sleadanalystofftopic.Theremain63papers

werereviewedbythestudy’ssubjectmatterexpert,highlightedthefollowing24

examplesofChineseresearchaddressingLLMproblemsstandintheof

modelsachievingthegeneralityassociatedGAI.621.曹博西HAN韓先培SUNLe(孫樂“CanPromptProbe

PretrainedLanguageModels?UnderstandingtheRisksfromaCausal

View,”preprintarXiv:2203.12258v12.CHENG程兵,“ArtificialIntelligenceGenerativeContentincluding

OpensaNewBigParadigmSpaceofEconomicsSocialScience

Research”以ChatGPT為代表的大語言模型打開了經(jīng)濟(jì)學(xué)和其他社會(huì)科學(xué)研究范

式的巨大新空間ChinaJournalofEconometrics計(jì)量經(jīng)濟(jì)學(xué)報(bào))3,no.3(July

2023).3.CHENG程岱宣HUANG黃少涵WEIFuru韋福如“AdaptingLargeLanguageModelstoDomainsviaReadingComprehension,”

preprintarXiv:2309.09530v44.DINGNing丁寧ZHENGHai-Tao鄭海濤SUNMaosong孫茂松“Parameter-efficientFine-tuningofLarge-scalePre-trainedLanguageModels,”

NatureIntelligence,March2023.5.DONGQingxiu董青秀SUIZhifang穗志方LILei李磊,“ASurveyonIn-

contextarXivpreprintarXiv:2301.00234v4(2024).6.HUANGJiangyong黃江勇YONGSilong雍子隆,63HUANGSiyuan黃思遠(yuǎn)“AnEmbodiedGeneralistAgentin3DWorld,”Proceedingsofthe41st

ConferenceonMachineLearning,Austria,235.

2024.CenterforSecurityEmergingTechnology|107.JINFeihu金飛虎ZHANG張家俊,“UnifiedPromptMakesPre-

trainedLanguageModelsBetterFew-shotLearners,”IEEEInternational

ConferenceonAcoustics,SpeechSignalProcessing,June2023.8.LIHengli李珩立ZHUSongchun朱松純ZHENG鄭子隆,“DiPlomat:

ADialogueDatasetforSituatedPragmaticReasoning,”37thConferenceon

NeuralInformationProcessingSystems(NeurIPS2023).9.LIJiaqi(李佳琪ZHENG鄭子隆ZHANG張牧涵,“LooGLE:Can

Long-ContextLanguageModelsUnderstandLongContext?”preprint

arXiv:2311.04939v1(2023).10.LIYuanchun李元春ZHANGYaqin張亞勤Yunxin劉云新,“Personal

LLMAgents:InsightsandSurveyabouttheCapability,EfficiencySecurity,”

preprintarXiv:2401.05459v211.MAYuxi馬煜曦ZHUSongchun朱松純,“BraininaonPieces

towardsArtificialGeneralIntelligenceinLargeModels,”arXiv

preprintarXiv:2307.03762v112.NIBolin尼博琳PENGHouwen彭厚文CHENZHANGSongyang

張宋揚(yáng)),LINGHaibin凌海濱),“ExpandingLanguagePretrainedModels

forGeneralVideoRecognition,”preprintarXiv:2208.02816v1(2022).13.PENGYujia彭玉佳ZHUSongchun朱松純,“TheTongTest:GeneralIntelligencethroughDynamicEmbodiedSocial

Interactions,”Engineering34,(2024).14.SHENGuobin申國斌ZENGYi曾毅,“Brain-inspiredNeuralCircuitEvolution

forSpikingNeuralNetworks,”PNAS39(2023).15.TANG唐曉娟ZHUSongchun朱松純LIANGYitao梁一韜ZHANG張牧涵“LargeLanguageModelsAreIn-contextSemantic

ReasonersRatherthanSymbolicReasoners,”arXivpreprintarXiv:2305.14825v2(2023).16.WANGJunqi王俊淇PENGYujia彭玉佳ZHUYixin朱毅鑫Lifeng范

麗鳳,“EvaluatingModelingSocialIntelligence:aComparativeStudyof

HumanAICapabilities,”arXivpreprintarXiv:2405.11841v1(2024).17.Fangzhi徐方植Jun(劉軍,ErikCambria,“AreLargeLanguageModels

GoodReasoners?”arXivpreprintarXiv:2306.09841v218.Zhihao徐智昊DAIQionghai(戴瓊海FANGLu方璐,“Large-scale

PhotonicChipletEmpowers160-TOPS/WGeneralIntelligence,”

Science,April2024.19.YUANLuyao袁路遙ZHUSongchun朱松純),“CommunicativeLearning:a

UnifiedFormalism,”Engineering,March2023.CenterforSecurityEmergingTechnology|1120.ZHANG張馳ZHUYixin朱毅鑫ZHUSongchun朱松純),“Human-level

shotConceptInductionthroughMinimaxEntropyScience

Advances,April2024.21.ZHANGTielin張鐵林徐波,“ABrain-inspiredAlgorithmthat

MitigatesCatastrophicForgettingofArtificialandSpikingNeuralNetworksLowComputationalCost,”ScienceAdvances,August2023.22.ZHANGYue章岳Leyang崔樂陽SHIShuming史樹明),“Siren’sSongin

theAIOcean:aSurveyonHallucinationinLargeModels,”arXiv

preprintarXiv:2309.01219v223.ZHAOZhuoya趙卓雅ZENGYi曾毅,“ABrain-inspiredTheoryofSpikingNeuralNetworkImprovesMulti-CooperationCompetition.”

Patterns,August2023.24.ZOU鄒旭YANG楊植麟TANGJie唐杰,“ControllableGeneration

fromPre-trainedLanguageModelsviaInversePrompting,”arXivpreprint

arXiv:2103.10685v3(2021).ThestudiescollectivelyaddresstheofLLMdeficitsdescribedinthispaper’s

sections12,namely,thoseassociatedtheoryof(ToM)failures,

inductive,deductive,abductivereasoningdeficits,problemslearningnew

tasksthroughanalogytoprevioustasks,ofgrounding/embodiment,

unpredictabilityoferrorsandhallucinations,lackofintelligence,insufficient

understandingofreal-worldinput,inparticularinvideoform,difficultyindealingcontexts,challengesassociatedtheneedtotuneoutputs,costof

operation.Proposedsolutionstotheseproblemsfrommodules,emulatingbrain

structureprocesses,rigorousstandardsandtesting,real-worldembedding,to

thecomputingsubstrateoutrightwithimprovedtypes.SeveralprominentChinesescientistscitedinthisstudy’ssection2,madepublic

statementssupportingGAImodels,includingTangJie,ZhangYaqin,Bo,

Songchun,ZengYi,areonthebylinesofofthesepapers,adding

authenticitytotheirdeclarations.addition,vofChina’stopinstitutionscompaniesengagedinGAI

research,includingtheAcademyofArtificialIntelligence北京智源人工智能研

究院theInstituteforGeneralArtificialIntelligence北京通用人工智能研究院theChineseAcademyofSciences’InstituteofAutomation中國科學(xué)院自動(dòng)化研究所PekingUniversity北京大學(xué)TsinghuaUniversity清華大學(xué)UniversityofChineseCenterforSecurityEmergingTechnology|12AcademyofSciences中國科學(xué)院大學(xué))andAlibaba,ByteDance,Huawei,TencentAIrepresentedintheselectedcorpus,inmostcasesonmultiplepapers.64Therecordofmetadataadducedhere,conclusionsdrawninpriorCSETresearh65

supportthepresentstudy’scontentionmajorelementsinChina’sAIcommunity

questionLLMs’potentialtoachieveGAI—throughincreasesinscaleormodalities—

arecontemplatingorpursuingalternativeCenterforSecurityEmergingTechnology|13Assessment:DoAllPathstotheBuddha?WhenLLM-basedchatbotsfirstbecameavailable,earlyclaimsLLMsmightbe

sentient,i.e.,experiencefeelingssensations,orevenshowself-awareness,66were

prevalentmuchdiscussed.Sincethen,coolerheadsprevailed,67tfocus

shiftedfromphilosophicalspeculationsabouttheinteriorlivesofLLMstomore

concretemeasurementsofLLMabilitiesonkeyof“intelligent”behaviortheimportantquestionofwhetherLLMsmightbecapableofgeneral

artificialintelligenceitisfarfromwhetherconsciousnessthecapacityforemotionstoGAI,whatisisthataGAIsystemmustbetoreasonto

separatefromhallucinations.Asthingsstand,LLMsexplicitmechanisms

wouldenablethemtoperformthesecorerequirementsofintelligentbehavior.

Rather,thehopeofLLMenthusiastsisthat,somehow,reasoningabilitieswill

“emerge”LLMstrainedtobecomeeverbetterpredictingthenextwordina

conversation.Yet,thereistheoreticalforthisbelief.Tothecontrary,research

shownthatLLMs’vasttextmemorymaskeddeficienciesinreasoning.68Heuristicattemptstoimprovereasoning(e.g.,chain-of-thought),69likelythesfor

improvedperformanceinOpenAI’snew“o1”LLM,morerecentapproachessuch

“rephraserespond,”70“tree-of-thought71orthoughts”72yieldedimprovements,butdonotsolvetheunderlyingproblemofthebseofa

core“reasoningengine.”thetoken,multipleattemptstofixLLMs’hallucinationproblem73run

intodeadendsbecausetheytoaddressthecoreproblemisinherenttoLLMs’

togeneralizefromtrainingdatatonewcontexts.Indeed,currenteffortsto

improvereasoningabilitiesfixhallucinationsabitlikeplaying“whack-a-mole”

butmoleshidinginabillion-dimensionalamalletis

uncertaintointended.Theresultingsystemsbesufficientfor

situationshumansassessthequalityofLLMoutput,e.g.,cover

letters,designingtravelitinerariesorcreatingessaysontopicsthatareperennial

favoritesofschoolteachers.Yet,theseafarfromGAI.ThepublicdebatesinthewesternontheappropriatepathtoGAItendtobe

drownedoutbycompaniesfinancialinterestsinpromotingtheirlatestLLMsof“humanlikeintelligence”or“sparksofartificialgeneralintelligence,”74even

intheofevermoreshortcomingsofLLMs,detailedinsection1.The

ofcommercialinterestspromoteLLMssuretoGAICenterforSecurityEmergingTechnology|14negativelyaffectedtheofacademicresearchintheU.S.topursue

alternativeapproachestoGAI.75Thesituationisdifferentinina.WhiletherecompaniesinChinadeveloping

LLMsforcommercialpurposes,leadingChineseAIscientistsandgovernmentofficials,

detailedinthispaper,thatLLMsfundamentallimitationsmakeit

importanttoinvestigateotherapproachestoGAIorsupplementLLMperformance

“brainlike”Thelatterstrategy,ofpursuing“braininspired”AIledtobreakthroughsintheforexample,bydeeplearning76—

modeledonthesensoryprocessinghierarchy—reinforcementlearning77—

modelinghowthebrainstrategiesfromrewards—into“deepreinforcement

g,”78which,forinstance,formedthebasisofAlphaGo,79thefirstartificialneural

networktbeathumanchampionsinthegameofGo.Thisdifferenceinresearch

directionsgiveChinaadvantageintheracetoachieveGAI.behelpfultothecurrentsituationtohowChinatodominate

theglobalmarketforphotovoltaicpanels(or,morerecently,batterytechnology

electricvehicles),basedonChinesegovernmentdecisionstheofthemillenniumtobecomeaworldleaderinTheensuingpolicydecisionsinvestmentstobuildupthedomesticindustryincreasetheefficiencyofpanelsledtoinnovationeconomiesofscalenowhaveChinaproducingleast75%oftheworld’spanels.AdecisionbyChinatostrategicallyinvestin

non-LLM-basedapproachestoGAI80repeatthissuccess,albeitinafieldofeven

greaterthanphotovoltaics.CenterforSecurityEmergingTechnology|15ManagingaChinaFirst-MoverGeoffreyHinton,recentNobelwinnerrecipientofaTuringAwardforworkonmultilayerneuralnetworks—thefirstAINNarchitectureledto

superhumanperformanceonarangeofbenchmarktasksincomputervisionother

—acknowledges“arace,clearly,betweenChinatheU.S.,neitherisgoing

toslowdown.”81Thisraceto