下載本文檔
版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進行舉報或認領
文檔簡介
Feature
HOWCLOSEISAI
N-?EL
LargelanguagemodelssuchasOpenAI’so1have
electrifiedthedebateoverachievingartificialgeneralintelligence.Buttheyareunlikelytoreachthis
milestoneontheirown.ByAnilAnanthaswamy
O
penAI’slatestartificialintelligence(AI)systemdroppedinSeptemberwithaboldpromise.Thecom-panybehindthechatbotChatGPTshowcasedo1—itslatestsuiteoflargelanguagemodels(LLMs)—ashavinga“newlevelofAIcapability”.OpenAI,whichisbasedinSanFran-
cisco,California,claimsthato1worksinawaythatisclosertohowapersonthinksthandopreviousLLMs.
Thereleasepouredfreshfuelonadebatethat’sbeensimmeringfordecades:justhowlongwillitbeuntilamachineiscapableofthewholerangeofcognitivetasksthathumanbrainscanhandle,includinggeneralizingfromonetasktoanother,abstractreasoning,plan-ningandchoosingwhichaspectsoftheworldtoinvestigateandlearnfrom?
Suchan‘a(chǎn)rtificialgeneralintelligence’,orAGI,couldtacklethornyproblems,includingclimatechange,pandemicsandcuresforcan-cer,Alzheimer’sandotherdiseases.Butsuchhugepowerwouldalsobringuncertainty—andposeriskstohumanity.“Badthingscould
happenbecauseofeitherthemisuseofAIorbecausewelosecontrolofit,”saysYoshuaBengio,adeep-learningresearcherattheUniversityofMontreal,Canada.
TherevolutioninLLMsoverthepastfewyearshaspromptedspeculationthatAGImightbetantalizinglyclose.ButgivenhowLLMsarebuiltandtrained,theywillnotbesufficienttogettoAGIontheirown,someresearcherssay.“Therearestillsomepiecesmissing,”saysBengio.
What’sclearisthatquestionsaboutAGIarenowmorerelevantthanever.“Mostofmylife,IthoughtpeopletalkingaboutAGIarecrack-pots,”saysSubbaraoKambhampati,acomputerscientistatArizonaStateUniversityinTempe.“Now,ofcourse,everybodyistalkingaboutit.Youcan’tsayeverybody’sacrackpot.”
WhytheAGIdebatechanged
Thephraseartificialgeneralintelligenceenteredthezeitgeistaround2007afteritsmentioninaneponymouslynamedbookeditedbyAIresearchersBenGoertzelandCassioPennachin.Itsprecisemeaningremains
elusive,butitbroadlyreferstoanAIsystemwithhuman-likereasoningandgeneralizationabilities.Fuzzydefinitionsaside,formostofthehistoryofAI,it’sbeenclearthatwehaven’tyetreachedAGI.TakeAlphaGo,theAIprogramcreatedbyGoogleDeepMindtoplaytheboardgameGo.Itbeatstheworld’sbesthumanplay-ersatthegame—butitssuperhumanqualitiesarenarrow,becausethat’sallitcando.
ThenewcapabilitiesofLLMshaveradicallychangedthelandscape.Likehumanbrains,LLMshaveabreadthofabilitiesthathavecausedsomeresearcherstoseriouslycon-sidertheideathatsomeformofAGImightbeimminent1,orevenalreadyhere.
Thisbreadthofcapabilitiesisparticularlystartlingwhenyouconsiderthatresearch-ersonlypartiallyunderstandhowLLMsachieveit.AnLLMisaneuralnetwork,amachine-learningmodellooselyinspiredbythebrain;thenetworkconsistsofartificialneurons,orcomputingunits,arrangedinlay-ers,withadjustableparametersthatdenotethestrengthofconnectionsbetweentheneurons.Duringtraining,themostpowerful
22|Nature|Vol636|5December2024
ILLUSTRATIONBYPETRAPéTERFFY
LLMs—suchaso1,Claude(builtbyAnthropicinSanFrancisco)andGoogle’sGemini—relyonamethodcallednexttokenprediction,inwhichamodelisrepeatedlyfedsamplesoftextthathasbeenchoppedupintochunksknownastokens.Thesetokenscouldbeentirewordsorsimplyasetofcharacters.Thelasttokeninasequenceishiddenor‘masked’andthemodelisaskedtopredictit.Thetrainingalgorithmthencomparesthepredictionwiththemaskedtokenandadjuststhemodel’sparameterstoenableittomakeabetterpredictionnexttime.Theprocesscontinues—typicallyusing
YOUDON’TSEETHATKINDOFAUTHENTICAGENCYINLARGE
LANGUAGEMODELS.”
billionsoffragmentsoflanguage,scientifictextandprogrammingcode—untilthemodelcanreliablypredictthemaskedtokens.Bythisstage,themodelparametershavecapturedthestatisticalstructureofthetrainingdata,andtheknowledgecontainedtherein.Theparametersarethenfixedandthemodelusesthemtopre-dictnewtokenswhengivenfreshqueriesor‘prompts’thatwerenotnecessarilypresentinitstrainingdata,aprocessknownasinference. Theuseofatypeofneuralnetworkarchitec-tureknownasatransformerhastakenLLMssignificantlybeyondpreviousachievements.Thetransformerallowsamodeltolearnthatsometokenshaveaparticularlystronginfluenceonothers,eveniftheyarewidelyseparatedinasampleoftext.ThispermitsLLMstoparselanguageinwaysthatseemtomimichowhumansdoit—forexample,dif-ferentiatingbetweenthetwomeaningsoftheword‘bank’inthissentence:“Whentheriver’sbankflooded,thewaterdamagedthebank’sATM,makingitimpossibletowithdrawmoney.” Thisapproachhasturnedouttobehighlysuccessfulinawidearrayofcontexts,
includinggeneratingcomputerprogramstosolveproblemsthataredescribedinnaturallanguage,summarizingacademicarticlesandansweringmathematicsquestions.
Andothernewcapabilitieshaveemergedalongtheway,especiallyasLLMshaveincreasedinsize,raisingthepossibilitythatAGI,too,couldsimplyemergeifLLMsgetbigenough.Oneexampleischain-of-thought(CoT)prompting.ThisinvolvesshowinganLLManexampleofhowtobreakdownaproblemintosmallerstepstosolveit,orsimplyaskingtheLLMtosolveaproblemstep-by-step.CoTpromptingcanleadLLMstocorrectlyanswerquestionsthatpreviouslyflummoxedthem.Buttheprocessdoesn’tworkverywellwithsmallLLMs.
ThelimitsofLLMs
CoTpromptinghasbeenintegratedintotheworkingsofo1,accordingtoOpenAI,andunderliesthemodel’sprowess.FrancoisChollet,whowasanAIresearcheratGoogleinMountainView,California,andleftinNovembertostartanewcompany,thinks
Nature|Vol636|5December2024|23
Feature
thatthemodelincorporatesaCoTgeneratorthatcreatesnumerousCoTpromptsforauserqueryandamechanismtoselectagoodpromptfromthechoices.Duringtraining,o1istaughtnotonlytopredictthenexttoken,butalsotoselectthebestCoTpromptforagivenquery.TheadditionofCoTreasoningexplainswhy,forexample,o1-preview—theadvancedversionofo1—correctlysolved83%ofprob-lemsinaqualifyingexamfortheInternationalMathematicalOlympiad,aprestigiousmathe-maticscompetitionforhigh-schoolstudents,accordingtoOpenAI.Thatcompareswithascoreofjust13%forthecompany’spreviousmostpowerfulLLM,GPT-4o.
But,despitesuchsophistication,o1hasitslimitationsanddoesnotconstituteAGI,sayKambhampatiandChollet.Ontasksthatrequireplanning,forexample,Kambhampati’steamhasshownthatalthougho1performsadmirablyontasksthatrequireupto16plan-ningsteps,itsperformancedegradesrapidlywhenthenumberofstepsincreasestobetween20and40(ref.2).Cholletsawsimilarlimita-tionswhenhechallengedo1-previewwithatestofabstractreasoningandgeneralizationthathedesignedtomeasureprogresstowardsAGI.Thetesttakestheformofvisualpuzzles.Solvingthemrequireslookingatexamplestodeduceanabstractruleandusingthattosolvenewinstancesofasimilarpuzzle,somethinghumansdowithrelativeease.
LLMs,saysChollet,irrespectiveoftheirsize,arelimitedintheirabilitytosolveproblemsthatrequirerecombiningwhattheyhavelearnttotacklenewtasks.“LLMscannottrulyadapttonoveltybecausetheyhavenoabilitytobasicallytaketheirknowledgeandthendoafairlysophisticatedrecombinationofthatknowledgeontheflytoadapttonewcontext.”
CanLLMsdeliverAGI?
So,willLLMseverdeliverAGI?Onepointintheirfavouristhattheunderlyingtransformerarchitecturecanprocessandfindstatisticalpatternsinothertypesofinformationinadditiontotext,suchasimagesandaudio,providedthatthereisawaytoappropriatelytokenizethosedata.AndrewWilson,whostudiesmachinelearningatNewYorkUni-versityinNewYorkCity,andhiscolleaguesshowedthatthismightbebecausethedif-ferenttypesofdataallshareafeature:suchdatasetshavelow‘Kolmogorovcomplexity’,definedasthelengthoftheshortestcomputerprogramthat’srequiredtocreatethem3.Theresearchersalsoshowedthattransformersarewell-suitedtolearningaboutpatternsindatawithlowKolmogorovcomplexityandthatthissuitabilitygrowswiththesizeofthemodel.Transformershavethecapacitytomodelawideswatheofpossibilities,increasingthechancethatthetrainingalgorithmwilldiscoveranappropriatesolutiontoaproblem,andthis‘expressivity’increaseswithsize.Theseare,
saysWilson,“someoftheingredientsthatwereallyneedforuniversallearning”.AlthoughWilsonthinksAGIiscurrentlyoutofreach,hesaysthatLLMsandotherAIsystemsthatusethetransformerarchitecturehavesomeofthekeypropertiesofAGI-likebehaviour.
Yettherearealsosignsthattransformer-basedLLMshavelimits.Forastart,thedatausedtotrainthemodelsarerunningout.ResearchersatEpochAI,aninstituteinSanFranciscothatstudiestrendsinAI,estimate4thattheexistingstockofpubliclyavailabletextualdatausedfortrainingmightrunoutsomewherebetween2026and2032.TherearealsosignsthatthegainsbeingmadebyLLMs
HUMANSAND
OTHERANIMALS
AREAPROOFOF
PRINCIPLETHAT
YOUCANGETTHERE.”
astheygetbiggerarenotasgreatastheyoncewere,althoughit’snotclearifthisisrelatedtotherebeinglessnoveltyinthedatabecausesomanyhavenowbeenused,orsomethingelse.ThelatterwouldbodebadlyforLLMs.
RaiaHadsell,vice-presidentofresearchatGoogleDeepMindinLondon,raisesanotherproblem.Thepowerfultransformer-basedLLMsaretrainedtopredictthenexttoken,butthissingularfocus,sheargues,istoolimitedtodeliverAGI.BuildingmodelsthatinsteadgeneratesolutionsallatonceorinlargechunkscouldbringusclosertoAGI,shesays.Thealgorithmsthatcouldhelptobuildsuchmodelsarealreadyatworkinsomeexisting,non-LLMsystems,suchasOpenAI’sDALL-E,whichgeneratesrealistic,sometimestrippy,imagesinresponsetodescriptionsinnaturallanguage.ButtheylackLLMs’broadsuiteofcapabilities.
Buildmeaworldmodel
TheintuitionforwhatbreakthroughsareneededtoprogresstoAGIcomesfromneuroscientists.Theyarguethatourintelli-genceistheresultofthebrainbeingabletobuilda‘worldmodel’,arepresentationofoursurroundings.Thiscanbeusedtoimaginedifferentcoursesofactionandpredicttheirconsequences,andthereforetoplanandrea-son.Itcanalsobeusedtogeneralizeskillsthathavebeenlearntinonedomaintonewtasksbysimulatingdifferentscenarios.
Severalreportshaveclaimedevidencefortheemergenceofrudimentaryworldmodels
insideLLMs.Inonestudy5,researchersWesGurneeandMaxTegmarkattheMassachusettsInstituteofTechnologyinCambridgeclaimedthatawidelyusedopen-sourcefamilyofLLMsdevelopedinternalrepresentationsoftheworld,theUnitedStatesandNewYorkCitywhentrainedondatasetscontaininginfor-mationabouttheseplaces,althoughotherresearchersnotedonX(formerlyTwitter)thattherewasnoevidencethattheLLMswereusingtheworldmodelforsimulationsortolearncausalrelationships.Inanotherstudy6,KennethLi,acomputerscientistatHarvardUniversityinCambridgeandhiscolleaguesreportedevi-dencethatasmallLLMtrainedontranscriptsofmovesmadebyplayersoftheboardgameOthellolearnttointernallyrepresentthestateoftheboardandusedthistocorrectlypredictthenextlegalmove.
Otherresults,however,showhowworldmodelslearntbytoday’sAIsystemscanbeunreliable.Inonesuchstudy7,computersci-entistKeyonVafaatHarvardUniversity,andhiscolleaguesusedagiganticdatasetoftheturnstakenduringtaxiridesinNewYorkCitytotrainatransformer-basedmodeltopredictthenextturninasequence,whichitdidwithalmost100%accuracy.
Byexaminingtheturnsthemodelgener-ated,theresearcherswereabletoshowthatithadconstructedaninternalmaptoarriveatitsanswers.Butthemapborelittleresem-blancetoManhattan(see‘TheimpossiblestreetsofAI’),“containingstreetswithimpos-siblephysicalorientationsandflyoversaboveotherstreets”,theauthorswrite.“Althoughthemodeldoesdowellinsomenavigationtasks,it’sdoingwellwithanincoherentmap,”saysVafa.Andwhentheresearcherstweakedthetestdatatoincludeunforeseendetoursthatwerenotpresentinthetrainingdata,itfailedtopredictthenextturn,suggestingthatitwasunabletoadapttonewsituations.
Theimportanceoffeedback
Oneimportantfeaturethattoday’sLLMslackisinternalfeedback,saysDileepGeorge,amemberoftheAGIresearchteamatGoogleDeepMindinMountainView,California.Thehumanbrainisfulloffeedbackconnectionsthatallowinformationtoflowbidirectionallybetweenlayersofneurons.Thisallowsinfor-mationtoflowfromthesensorysystemtohigherlayersofthebraintocreateworldmod-elsthatreflectourenvironment.Italsomeansthatinformationfromtheworldmodelscanripplebackdownandguidetheacquisitionoffurthersensoryinformation.Suchbidirec-tionalprocesseslead,forexample,topercep-tions,whereinthebrainusesworldmodelstodeducetheprobablecausesofsensoryinputs.Theyalsoenableplanning,withworldmodelsusedtosimulatedifferentcoursesofaction. ButcurrentLLMsareabletousefeedbackonlyinatacked-onway.Inthecaseofo1,the
24|Nature|Vol636|5December2024
TruestreetsinManhattan,NewYork
Non-existent‘streets’reconstructed
Directionbyanartificial-intelligencesystem
oftravel
attheDalleMolleInstituteforArtificialIntelligenceStudiesinLugano-Viganelllo,Switzerland,reported9buildinganeuralnet-workthatcouldefficientlybuildaworldmodelofanartificialenvironment,andthenuseittotraintheAItoracevirtualcars.
IfyouthinkthatAIsystemswiththislevelofautonomysoundscary,youarenotalone.AswellasresearchinghowtobuildAGI,BengioisanadvocateofincorporatingsafetyintothedesignandregulationofAIsystems.Hearguesthatresearchmustfocusontrainingmodelsthatcanguaranteethesafetyoftheirownbehaviour—forinstance,byhavingmech-anismsthatcalculatetheprobabilitythatthemodelisviolatingsomespecifiedsafetycon-straintandrejectactionsiftheprobabilityistoohigh.Also,governmentsneedtoensuresafeuse.“Weneedademocraticprocessthatmakessureindividuals,corporations,eventhemilitary,useAIanddevelopAIinwaysthataregoingtobesafeforthepublic,”hesays.
SOURCE:REF.7
THEIMPOSSIBLESTREETSOFAI
Theabilitytobuildrepresentationsofour
environment,calledworldmodels,helpshumansto
reasonandplan.ItisthoughtthatAIsystemswillneedthiscapacity,too,iftheyaretodevelophuman-level
intelligence.InthecaseofanAIsystemthatwas
trainedtopredictroutestakenbytaxisinManhattan,NewYork,itsinternalmapdidnotresemblethereal
world.Inlatertesting,thisledtoaninabilitytohandledetoursthatwerenotpresentinthetrainingdata.
TheAIsystem’smap
containsstreetswith
impossibleorientations
andbridgesthatdon’texist.
SowilliteverbepossibletoachieveAGI?Computerscientistssaythereisnoreasontothinkotherwise.“Therearenotheoreticalimpediments,”saysGeorge.MelanieMitchell,acomputerscientistattheSantaFeInstituteinNewMexico,agrees.“Humansandsomeotheranimalsareaproofofprinciplethatyoucangetthere,”shesays.“Idon’tthinkthere’sanythingparticularlyspecialaboutbiologicalsystemsversussystemsmadeofothermaterialsthatwould,inprinciple,preventnon-biologicalsystemsfrombecomingintelligent.”
internalCoTpromptingthatseemstobeatwork—inwhichpromptsaregeneratedtohelpansweraqueryandfedbacktotheLLMbeforeitproducesitsfinalanswer—isaformoffeed-backconnectivity.But,asseenwithChollet’stestsofo1,thisdoesn’tensurebullet-proofabstractreasoning.
Researchers,includingKambhampati,havealsoexperimentedwithaddingexternalmod-ules,calledverifiers,ontoLLMs.ThesecheckanswersthataregeneratedbyanLLMinaspe-cificcontext,suchasforcreatingviabletravelplans,andasktheLLMtorerunthequeryiftheanswerisnotuptoscratch8.Kambhampati’steamshowedthatLLMsaidedbyexternalverifi-erswereabletocreatetravelplanssignificantlybetterthanwerevanillaLLMs.Theproblemisthatresearchershavetodesignbespokeverifi-ersforeachtask.“Thereisnouniversalverifier,”saysKambhampati.Bycontrast,anAGIsystemthatusedthisapproachwouldprobablyneedtobuilditsownverifierstosuitsituationsastheyarise,inmuchthesamewaythathumanscanuseabstractrulestoensuretheyarereasoningcorrectly,evenfornewtasks.
EffortstousesuchideastohelpproducenewAIsystemsareintheirinfancy.Bengio,forexample,isexploringhowtocreateAIsys-temswithdifferentarchitecturestotoday’stransformer-basedLLMs.Oneofthese,which
useswhathecallsgenerativeflownetworks,wouldallowasingleAIsystemtolearnhowtosimultaneouslybuildworldmodelsandthemodulesneededtousethemforreasoningandplanning.
AnotherbighurdleencounteredbyLLMsisthattheyaredataguzzlers.KarlFriston,athe-oreticalneuroscientistatUniversityCollegeLondon,suggeststhatfuturesystemscouldbemademoreefficientbygivingthemtheabilitytodecidejusthowmuchdatatheyneedtosam-plefromtheenvironmenttoconstructworldmodelsandmakereasonedpredictions,ratherthansimplyingestingallthedatatheyarefed.This,saysFriston,wouldrepresentaformofagencyorautonomy,whichmightbeneededforAGI.“Youdon’tseethatkindofauthen-ticagency,insay,largelanguagemodels,orgenerativeAI,”hesays.“Ifyou’vegotanykindofinte
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負責。
- 6. 下載文件中如有侵權(quán)或不適當內(nèi)容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 2025年度建筑材料加工生產(chǎn)合同范本4篇
- 專業(yè)出國留學輔導協(xié)議樣本(2024)版B版
- 2025年度醫(yī)療器械緊急運輸服務協(xié)議3篇
- 2025年度數(shù)據(jù)中心場地租賃合作協(xié)議4篇
- 2025年度食品試用及消費者滿意度調(diào)查合同4篇
- 2025年度綠色建筑設計與施工一體化服務合同4篇
- 2025年度市政基礎設施改造鏟車租賃協(xié)議書4篇
- 二零二四全新建筑工程施工聯(lián)營協(xié)議書下載3篇
- 2024重慶離婚協(xié)議書標準范文
- 二婚再婚2024年度財產(chǎn)共有協(xié)議
- 2024年黑河嫩江市招聘社區(qū)工作者考試真題
- 第22單元(二次函數(shù))-單元測試卷(2)-2024-2025學年數(shù)學人教版九年級上冊(含答案解析)
- 藍色3D風工作總結(jié)匯報模板
- 安全常識課件
- 河北省石家莊市2023-2024學年高一上學期期末聯(lián)考化學試題(含答案)
- 2024年江蘇省導游服務技能大賽理論考試題庫(含答案)
- 2024年中考英語閱讀理解表格型解題技巧講解(含練習題及答案)
- 新版中國食物成分表
- 浙江省溫州市溫州中學2025屆數(shù)學高二上期末綜合測試試題含解析
- 2024年山東省青島市中考生物試題(含答案)
- 保安公司市場拓展方案-保安拓展工作方案
評論
0/150
提交評論