2023前沿AI能力和風(fēng)險報告（英文版）

CapabilitiesandrisksfromfrontierAI

AdiscussionpaperontheneedforfurtherresearchintoAIrisk

October2023

Acknowledgements

Wewouldliketothanktheexpertreviewpanel,YoshuaBengio,SaraHooker,Arvind

Narayanan,WilliamIsaac,PaulChristiano,IreneSolaiman,AlexanderBabutaandJohnMcDermidfortheirinsightfulcommentsandfeedback.

ThisreportisadiscussionpapertosupporttheAISafetySummit,anddoesnotrepresentapolicypositionofHMGorrepresenttheviewsoftheexpertreviewpanelabove,whoonlyprovidedcommentsforconsideration.

FrontierAI–CapabilitiesandRisks

Contents

Introduction 4

WhatisthecurrentstateoffrontierAIcapabilities? 5

HowfrontierAIworks 5

FrontierAIcanperformmanyeconomicallyusefultasks 7

FrontierAImodelscanbeaugmentedwithtoolstomakethemmoreautonomous 7

FrontierAIcouldbemorecapablethanevaluationsindicate 8

LimitationsoffrontierAI 9

HowmightfrontierAIcapabilitiesimproveinthefuture? 10

RecentAIprogresshasbeenrapid 10

Recentprogresswasdrivenbysystematictrendsincompute,dataandalgorithms 11

Scalinglaws:performanceimprovespredictablywithincreasedcomputeanddata 12

RapidAIprogressislikelytocontinueforseveralyears 14

Advancedgeneral-purposeAIagentsmightbedevelopedinthefuture 15

WhatrisksdofrontierAIpresent? 15

Crosscuttingriskfactors 16

Itisdifficulttodesignsafefrontiermodelsinopen-endeddomains 16

EvaluatingthesafetyoffrontierAIsystemsisanopenchallenge 16

ItmaybedifficulttotrackhowfrontierAIsystemsaredeployedorused 17

AIsafetystandardshavenotyetbeenestablished 18

InsufficientincentivesforAIdeveloperstoinvestintoriskmitigationmeasures 18

TheremaybesignificantconcentrationofmarketpowerinAI 19

Societalharms 19

Degradationoftheinformationenvironment 19

Labourmarketdisruption 20

Bias,FairnessandRepresentationalHarms 21

Misuserisks 22

DualUseSciencerisks 22

Cyber 23

DisinformationandInfluenceOperations 25

Lossofcontrol 25

HumansmightincreasinglyhandovercontroltomisalignedAIsystems 26

FutureAIsystemsmightactivelyreducehumancontrol 26

Conclusion 28

Glossary 29

FrontierAI–CapabilitiesandRisks

4

Introduction

Weareinthemidstofatechnologicalrevolutionthatwillfundamentallyalterthewaywelive,work,andrelatetooneanother.ArtificialIntelligence(AI)promisestotransformnearlyeveryaspectofoureconomyandsociety.Theopportunitiesaretransformational-advancingdrugdiscovery,makingtransportsaferandcleaner,improvingpublicservices,speedingupandimprovingdiagnosisandtreatmentofdiseaseslikecancerandmuchmore.

DevelopmentsinfrontierAIaretransformingproductivityandsoftwareservices,whichwill

multiplytheproductivityofmanyindustriesandsectors.1ThisprogressinfrontierAIinrecentyearshasbeenrapid,andthemostadvancedsystemscanwritetextfluentlyandatlength,

writewell-functioningcodefromnaturallanguageinstructions,makenewapps,scorehighlyonschoolexams,generateconvincingnewsarticles,translatebetweenmanylanguages,

summariselengthydocuments,amongstothercapabilities.Theopportunitiesarevast,andthereisgreatpotentialforincreasingtheproductivityofworkersofallkinds.

However,thesehugeopportunitiescomewithrisksthatcouldthreatenglobalstabilityand

undermineourvalues.Toseizetheopportunities,wemustunderstandandaddresstherisks.AIposesrisksinwaysthatdonotrespectnationalboundaries.Itisimportantthat

governments,academia,businesses,andcivilsocietyworktogethertonavigatetheserisks,

whicharecomplexandhardtopredict,tomitigatethepotentialdangersandensureAIbenefitssociety.

TheUKGovernmentbelievesmoreresearchintoAIriskisneeded.Thisreportexplainswhy.ItdescribesthecurrentstateandkeytrendsrelatingtofrontierAIcapabilities,andthenexploreshowfrontierAIcapabilitiesmightevolveinthefutureandreviewssomekeyrisks.Thereis

significantuncertaintyaroundboththecapabilitiesandrisksfromAI,includingsomeexpertswhobelievethatsomeoftheserisksareoverstated.Thisreportfocusesonevidenceforrisksandconcludesthatdoingfurtherresearchisnecessary.

Thisreportcoversmanyrisks,butwewishtoemphasisethattheoverarchingriskisalossoftrustinandtrustworthinessofthistechnologywhichwouldpermanentlydenyusandfuture

generationsitstransformativepositivebenefits.Indiscussingtheotherrisks,wedosoinordertogalvanizeactiontomitigatethem,suchthatwecancapturethefullbenefitsoffrontierAI.

DefiningAIischallengingasitremainsaquicklyevolvingtechnology.ForthepurposesoftheSummitwedefine“frontierAI”ashighlycapablegeneral-purposeAImodelsthatcanperformawidevarietyoftasksandmatchorexceedthecapabilitiespresentintoday’smostadvanced

models(seeFigure1).2Today,thisprimarilyincludeslargelanguagemodels(LLMs)3suchasthoseunderlyingChatGPT,4Claude,5andBard.6However,itisimportanttonotethat,bothtodayandinthefuture,frontierAIsystemsmaynotbeunderpinnedbyLLMs,andcouldbe

underpinnedbyanothertechnology.

5

Figure1:ScopeoftheAISafetySummit-2023

AlphaGo,AlphaFoldorDALLE3whichcannotperformaswideavarietyoftasks.8

ThelimitedfocusofthisreportmeanswedonotcoverpowerfulnarrowAI7systemslike

TherearealreadyanumberofexistinginternationaleffortsandinitiativeswhichtouchuponthecapabilitiesandrisksoffrontierAI.TheupcomingAISafetySummitwillprovidespacefora

focusedanddeepdiscussiononAIsafetyatthefrontierandwhatfurtheractionneedstobetaken,complementingexistinginitiatives,andthisreportisintendedtobearesourceforall.

Thisreportisbynomeansconclusive;therearemanyrisksweomitandweencouragereaderstoviewitasthestartofaconversation.

WhatisthecurrentstateoffrontierAIcapabilities?

FrontierAIcanperformawidevarietyoftasks,isbeingaugmentedwithtoolsto

enhanceitscapabilities,andisbeingincreasinglyintegratedintosystemsthatcanhaveawideimpactontheeconomyandsociety.Althoughthesemodelsstillhavemajor

limitationssuchastheirfactualityandreliability,theircurrentcapabilitiesare

impressive,maybegreaterthanwehavebeenabletoassess,andhaveappearedfasterthanweexpected.

HowfrontierAIworks

FrontierAIcompaniessuchasOpenAI,DeepMindandAnthropicdeveloplargelanguagemodels(LLMs)suchasGPT-4intwophases:pre-trainingandfine-tuning.

Duringpre-training,anLLM“reads”millionsorbillionsoftextdocuments.9Asitreads,wordbyword,10itpredictswhatwordwillcomenext.Atthestartofpre-trainingitpredictsrandomly,but

6

asitseesmoredataitlearnsfromitsmistakesandimprovesitspredictiveperformance.Oncepre-trainingisover,themodelissignificantlybetterthanhumansatpredictingthenextwordofarandomlychosentextdocument.11

Duringfine-tuning,12thepre-trainedAIisfurthertrainedonhighlycurateddatasets,whicharefocusedonmorespecialisedtasks,orarestructuredtodirectmodelbehaviourinwayswhichareinalignmentwithdevelopervaluesanduserexpectations13

Increasingly,frontierAImodelsaremulti-modal.Inadditiontotext,theycangenerateandprocessotherdatatypessuchasimages,video,andsound.14

Thekeyinputstodevelopmentarecomputationalresources(“compute”15)totrainandrunthemodel,dataforittolearnfrom,thealgorithmsthatdefinethistrainingprocess,andtalentandexpertisethatenableallofthis.16Thevastmajorityofcomputeisspentonpre-training,whichiswhenmostcorecapabilitiesarelearntbyamodel.17

ThetotaldevelopmentcostsforthemostcapablefrontierAImodelstodayrunsintothetensofmillionsofpounds,18withcostsexpectedtosoonreachintothehundredsofmillionsorevenbillionsofpounds.19Whilethebestperformingmodelsaredevelopedbyasmallnumberof

well-resourcedorganisations,alargernumberofsmallerentitiesbuildproductsontopofthesefrontiermodelsforspecificmarkets.20

Thebelowdiagramoutlinestheinputsto,andstagesof,thedevelopmentanddeploymentoffrontierAI.

Figure2.Anoverviewoffoundationmodeldevelopment,traininganddeployment.From

AIFoundationModels:initialreview,

CMA,2023.

7

FrontierAIcanperformmanyeconomicallyusefultasks

Simplyfrombeingtrainedtopredictthenextwordacrossdiversedatasets,modelsdevelopsophisticatedcapabilities.21Forexample,frontierAIcan(withvaryingdegreesofsuccessandreliability):

●Conversefluentlyandatlength,drawingonextensiveinformationcontainedintrainingdata.

●Writelongsequencesofwell-functioningcodefromnaturallanguageinstructions,includingmakingnewapps.22

●Scorehighlyonhigh-schoolandundergraduateexaminationsinmanysubjects.23

●Generateplausiblenewsarticles.24

●Creativelycombineideastogetherfromverydifferentdomains.25

●Explainwhynovelsophisticatedjokesarefunny.26

●Translatebetweenmultiplelanguages.27

●Directtheactivitiesofrobotsviareasoning,planningandmovementcontrol.28

●Analysedatabyplottinggraphsandcalculatingkeyquantities.29

●Answerquestionsaboutimagesthatrequirecommon-sensereasoning.30

●Solvemathsproblemsfromhigh-schoolcompetitions.31

●Summariselengthydocuments.32

Thesecapabilitiesshowpotentialtobeappliedacrossawidearrayofeconomicuse-cases.Inadditiontosomeoftheapplicationsabove,frontierAIhasbeenusedto:

●Improvetheperformanceofleadingconsultantsindevelopinggo-to-marketplans.33

●Automateawidevarietyoflegalwork.34

●Supportleadingwealthmanagers.35

●Increasetheproductivityofcall-centreworkers.36

●Accelerateacademicresearch,forexampleineconomics.37

AnnexAprovidesmoredetailonAIcapabilitiesincontentcreation,computervision,theoryofmind,memory,mathematics,physicalintuition,androbotics.

FrontierAImodelscanbeaugmentedwithtoolstomakethemmoreautonomous

FrontierAImodelsaremoreusefulwhenaugmentedwithothertoolsandsoftware.

8

FrontierAImodels,beforetheyareaugmented,respondtoarequestsimplybyproducingasnippetoftext.Bycontrast,autonomous38AIagents39cantakelongsequencesofactionsinpursuitofagoal,withoutrequiringhumaninvolvement.

Researchershavebuiltsoftwareprogramscalled“scaffolds”40thatallowfrontierAImodelstopowerautonomousAIagents.ThescaffoldpromptstheAImodeltocreateaplanforachievingahigh-levelgoalandtothenexecutetheplanstepbystep.ThescaffoldaugmentstheAI

modelwithtoolslikewebbrowsers,allowingittoexecuteeachstepautonomously.Theresultantsystem,builtoutoftheAImodelandthescaffold,isanAIagent.AutoGPTisthemostwell-publicisedexampleofsuchanAIagentasoflate2023.41

Today’sAIagentscurrentlystruggletoperformmosttasks–theyoftengetstuckinloopsandcannotself-correct,orfailatcrucialsteps.However,theydoallowfrontierAItoperformsomeentirelynewtasks.ExamplesoftasksthatAIagentscancurrentlydoinclude:

●Findspecificinformationbybrowsingtheinternet.42

●Organisepartiesinsimulated‘TheSims’-likeenvironments.43

●Solvecomplexproblemsinopen-worldsurvivalgameslikeMinecraft44andCrafter45.

●Supportthesynthesisofchemicalsbysearchingthewebforrelevantinformationandwritingcodetooperaterobotichardware.46

ManyleadingAIresearchersandcompaniesexplicitlyaimtobuildAIagentswhosegeneralcapabilitieswouldexceedthoseofhumans.47

FrontierAIcouldbemorecapablethanevaluationsindicate

ResearchersandusersfrequentlyuncoversurprisingcapabilitiesforfrontierAImodelswhichpre-deploymentevaluationdidnotuncover.48

ThecapabilitiesoffrontierAImodelsarelikelytobefurtherenhancedinmanywaysinthefuture,suchasthrough:

●Betterprompts.49ThewaythataquestionisphrasedcansignificantlyaffectafrontierAIsystem’sresponse.Forexample,encouragingamodeltothinkthroughitsanswer“stepbystep”significantlyimprovesperformanceonmathsandlogicproblems.50

●Bettertools.FrontierAImodelscanbetrainedtousetoolslikewebbrowsers,

calculators,knowledgedatabases,orrobotactuators,andcancompetentlyuseentirelynewtoolswhenprovidedtextinstructionsonhowtousethem51.Thesetoolsand

resourcescansignificantlyimprovecapabilitiesatrelevanttasksorendowthemwithentirelynovelcapabilities,suchastheabilitytodirectlymanipulatephysicalsystems.52

●Betterscaffolds.Scaffoldingsoftwareprograms(“scaffolds”)structuretheinformationflowofanAImodel,leavingthemodelitselfunchanged.53Betterscaffoldscould,forexample,helpanAIagentself-correctwhentheyhavemadeamistake,54orimprovetheirlong-termmemory.

●Newfine-tuningdata.Fine-tuningonhigh-qualitydatacansignificantlyimproveAIcapabilitiesinagivendomain,atatinyfractionofthecostofpre-training.

9

●Team-workbetweenAIsystems.MultipledifferentAIsystems,includingbothnarrowmodelsandmoregeneralmodels,couldcollaboratetoperformtasks.55

Unlikepre-training,theseimprovementsdonotrequiresignificantcomputationalresourcesandsoawiderangeofactorscouldcheaplyimprovefrontierAIcapabilities,providedthey

haveeasyaccesstopre-trainedmodels.

LimitationsoffrontierAI

ThereisongoingdebateaboutthelimitationsoffrontierAIsystems,includingwhethertheirperformanceisdrivenmorebygeneralreasoningorbyacombinationofmemorisationandfollowingbasicheuristics56.

GeneralreasoningabilitiesareevidencedbyfrontierAIproducingremarkablyaptresponsestonovelquestions,Forexample,PaLM’sabilitytounderstandthehumourbehindjokeswhich

hadneverbeforebeentold.57

However,thereisalsoevidencethatmodelsrelyheavilyonmemorisationandbasicheuristics:

●LLMsperformlesswellwhenaquestionisrewordedtomakeitdifferentfromtextthatisintheirtrainingdata.58

●LLMsoftensolvecomplexproblemsusingoverly-simpleheuristicsthatwouldfailtosolveothersimilarproblems.59

●ThereareinstanceswhereLLMsfailtoapplyinformationfromtheirtrainingdatainverybasicways.60

Beyondanuncertainabilitytogeneralisetonewcontexts,otherkeylimitationsofcurrentfrontierAImodelsinclude:

●Hallucinations:AIsystemsregularlyproduceplausibleyetincorrectanswersandstatetheseanswerswithhighconfidence.61Thismightbeaddressedbysystemsusing

knowledgerepositories,62improvedfine-tuning,ornewmethodsforteachingthemodelwhatitdoesanddoesnotknow.

●Coherenceoverextendeddurations:AImodelsarelessreliableontasksthatrequirelong-termplanningortakingalargenumberofsequentialsteps(e.g.writinganovel).63Thisispartiallyduetotheirrestrictedcontextlengthandthescarcityoflong-duration

tasktrainingdata.64TheselimitationsmightbeaddressedbyalgorithmicinnovationstogiveAIasourceoflong-termmemory,creatingmoredataonlong-horizontasks,betterscaffoldsthathelpAIagentsspotandcorrecttheirownerrors,65orimprovedtechniquesforbreakinglongtasksintomultiplesmallsteps66.

●Lackofdetailedcontext:Manytasksintherealeconomyrequireextensivecontextaboutaparticularcompany,project,orcode-base.Currentfrontiersystemsare

genericallycompetent,butlackthisspecificcontextandcannotlearnitfromthe

availabledata.Thismightbeaddressedbyaccesstoadditionalprivatedatasources,newdatagenerationtechniques,moredata-efficientfine-tuningtechniques,new

“model-based”learningmethods,67orsimplybyincreasingthecomputeanddatausedtodevelopthesystem.

10

Itremainsuncertainhowtheselimitationswillevolve.SomearguethattheselimitationswillpermanentlylimitfrontierAIdevelopmentincertainapplications.Ontheotherhand,recentprogressinAIhasgreatlysurpassedexpertpredictionsinmanydomains,while

underperforminginotherareas.68

HowmightfrontierAIcapabilitiesimproveinthefuture?

RecentAIprogresshasbeenrapidandwilllikelycontinue.Thisisduetopredictable

improvementsintheperformanceoffrontierAImodelswhendevelopedwithmore

compute,moredataandbetteralgorithms.Unexpectednewcapabilitiesmayalso

emerge.Advancedgeneral-purposeAIagentscouldbedevelopedinthenottoodistantfuture–althoughthisisasubjectofdebate,especiallyregardingthetiming.

RecentAIprogresshasbeenrapid

TherecentpaceofAIprogresshassurprisedforecastersandmachinelearningexpertsalike.69ProblemsthatfrustratedtheAIcommunityfordecadeshaverapidlyfallentoever-more-

capablemodels.

Figure3.AnoverviewofnotableAIachievementsfrom2022-2023acrossdiversedomains,Epoch2023

RecentadvancesinfrontierAIarethecontinuationofalonger-runningtrend:therapid

progresssince2012initsparentfieldofdeeplearningacrosscomputervision,gameplaying,andlanguagemodelling.70In2014,AIcouldonlygeneratesimple,blurryimages.However,by2022,modelslikeDALL-E2andImagencouldgeneratehigh-quality,creativeimagesfromtextprompts(seefigure4a).SubstantialadvanceswereseenintheshiftfromGPT-3.5toGPT-4,releasedjustmonthsapart.Forexample,oncalculusquestionsGPT-3.5scoredbelowmost

humans,butGPT-4improvedsignificantlyandscoredaroundthemedianhumanlevel.

11

Figure4b.CompletionsfromGPT-2to

Figure4a.Timelineofimagesgeneratedbyimagemodelsfrom

OurWorldinData

GPT-4.GPT-4completionfrom

Bubeck

etal.,2023.

Recentprogresswasdrivenbysystematictrendsincompute,dataandalgorithms

AstandardanalysisofprogressinAIcapabilitiesconsidersthreekeyfactors:computingpower,data,andimprovementsintheunderlyingalgorithms.71

Computingpower(“compute”forshort)referstothenumberofoperationsthatareperformed,usuallyinthecontextoftrainingAIsystems.Theamountofcomputeusedduringtraininghasexpandedoverthepastdecadebyafactorof55million:fromsystemstrainedbysingle

researchersatthecostofafewpounds,tosystemstrainedonmultipleGPUclustersby

companiesatthecostofmanymillionsofpounds.72Thistrendismostlytheresultofspendingmoremoneyoncompute,aswellastheresultofsignificanttechnologicalimprovementsto

computinghardware.73

Trainingalgorithmshavealsoimprovedsubstantiallyoverthepastdecade,sothattoday’s

machinelearningmodelscanachievethesameperformancewithlesscomputeanddatathan

thoseofthepast.Researchsuggeststhatbetteralgorithmsroughlyhalvedcompute

requirementseachyearforvisionandlanguagemodels.74MassiveamountsofdatahavealsoplayedanimportantroleinrecentAIprogress.AIdevelopershavetappedintoreadilyavailabledatasetsscrapedfromtheinternet,withtheamountoftrainingdatausedgrowingatover50%peryear.75

Enhancementsappliedafterinitialtraininghavefurtheraugmentedsystemcapabilities.Thesepost-trainingenhancementsincludeimproveddataforfine-tuning,76equippingmodelswith

toolslikecalculators,77webbrowsers78,andbetterprompts.79Post-trainingenhancementscansignificantlyimproveperformanceinspecificdomainsatasmallfractionoftheoriginaltrainingcost,80andsoawiderangeofactorscanusethemtoimprovefrontierAIcapabilities.

12

Scalinglaws:performanceimprovespredictablywithincreasedcomputeanddata

ThekeydriverfortheincreaseincomputeanddataisthatfrontierAImodelperformance

predictablyimproveswithmodelscale.Researchershavediscoveredso-called“scaling

laws”,81whichcanpredict,givenaparticularamountofcomputeanddata,afrontierAImodel’sperformanceatthespecifictaskofpredictingthenextword(thetaskusedtotrainthese

models).

Figure5a.Trainingerrorreduces

predictablywithcomputeacrossa

broadrangeofempirically-studied

trainingruns.Figurefrom

Hoffmannet

al,2022.

Figure5b.ExponentialincreaseintrainingcomputeforOpenAI'sGPTmodelsfrom2018to2023.82Epoch.

Nextwordpredictionhascontinuallyimprovedovertimeasdevelopershavescaledtheir

trainingcomputeanddata.Itisuncertainhowlongthistrendwillcontinue,butithasheldovermanyordersofmagnitudeofcomputeanddatasetsizeincreaseswithoutbreaking.

Whilethenextwordpredictiontaskisnotitselfwhatwecareabout,itisusedasanindicatorofmodelcapabilitiessinceitisstronglycorrelatedwithperformanceinmanydownstreamtasks.83Forexample,ifamodelisextremelygoodatnextwordpredictiononcodeandmathematics

data,itismorelikelytobegoodatsolvingprogrammingpuzzlesandmathematicsproblems.

13

Figure6.PerformanceonbroadbenchmarkssuchasBIG-BenchandMMLUimproveswithmoretrainingcompute.ThisfigurewastakenfromOwen2023.

Althoughaverageperformance,aggregatedacrossmanydownstreamtasks,improvesfairlypredictablywithscale,itismuchhardertopredictperformanceimprovementsatspecificreal-worldproblems.ThedevelopmentoffrontierAIsystemshasinvolvedmanyexamplesof

surprisingcapabilities,unanticipatedbymodeldevelopersbeforetrainingandoftenonly

discoveredbyusersafterdeployment.Therearedocumentedexamplesofunexpected

capabilitieswheremodelswerenotshowinganysignsofimprovementbeforeacertainscaleandthenrapidlyimprovedsuddenly84–thoughtheinterpretationoftheseexamplesis

contested.85Inanycase,wecannotcurrentlyreliablypredictaheadoftimewhichspecificnewcapabilitiesafrontierAImodelwillgainwhenitistrainedwithmorecomputeanddata.

14

Figure7.IndividualcapabilitiesmayappearsuddenlyorunexpectedlyasthecomputeusedtodevelopAIincreases.Figurefrom

Weietal,2022.

RapidAIprogressislikelytocontinueforseveralyears

TherecentimprovementinAIcapabilitiesisnottheresultofasinglebreakthroughbutratheraconcertedadvancementacrossmultipledimensions,includingalgorithms,spendingon

compute,improvementsinhardwareperformance,andpost-trainingenhancements.Allofthesefactorscanindependentlyenhanceprogress,meaningthatchallengesorlimitsinanysingleoneofthemisunlikelytostopprogressinAIasawhole.

InvestmentsinAIwillcontinuetogrowrapidlyoverthenextfewyears.86LeadingAIdeveloperslikeAnthropicandOpenAIhavegarneredsignificantfundingandestablishedcloud

partnerships,inlargeparttosupportfurtherscalingofcompute.87HardwaremanufacturerslikeTSMCarereportedlyexpandingtheirproductionofAIchips,againsuggestingthatmore

computationalresourceswillbeavailablefortraining.88

However,sustainingtherateofrecentrapidscaleupofcomputeanddatapast2030islikelytorequirenewapproaches.Developerswouldhavetoi)spendhundredsofbillionsofpoundsoncomputeforasingletrainingrun89andii)findwaystogeneratesufficienthigh-qualitydata

goingbeyondwhatisreadilyavailableontheinternet.90Havingsaidthis,improvementsinalgorithmicefficiencymayreducecomputeneeds,suchthatcomputemightnotbeabindingconstraint.

NovelresearchdirectionsthatcouldfurtheracceleratefrontierAIprogressinclude:

15

●Enrichedtrainingdata–e.g.experthumanfeedback,AIgeneratedsyntheticfeedback,anddatapruning–mayincreasedataefficiency,improvecapabilitiesonchallengingscientificproblems,andreducecosts.91

●Multimodaltraining,whichmayofferincreasingsynergiesbetweenthedifferent

modalitiesandthepotentialforfrontierAItoprocessandproducetext,images,audioandvideo.92

●TrainingfrontierAItoactasanautonomousagentthatnavigatestheinternetasahumanandperformslongsequencesofactions,usingtheabovetechniquesto

generatecheapdataforlearningtheseskills.93

Importantly,thereisalsotheprospectthatAIsystemsthemselvesaccelerateAIprogress.

FrontierAIisalreadyhelpingAIresearcherstocreatesyntheticdatafortraining,94writenewcode,95andevenimprovemodelarchitectures.96WhileAIresearchiscurrentlymostlynon-automated,increasedautomationbyfuturefrontierAIsystemsmayacceleratethepaceofAIprogresssignificantly.97ThiscouldmeanwedevelopverycapableAIsystemssoonerthatwewouldotherwiseexpect,andhavelesstimetopreparefortheassociatedrisks.

Advancedgeneral-purposeAIagentsmightbedevelopedinthefuture

RecentprogressinAIhasprompteddiscussionregardingthepotentialnear-termdevelopment

ofadvancedgeneral-purpose,highlyautonomousAIagentsthatcanperformmosteconomicallyvaluabletasksbetterthanhumanexperts.

SeveralleadingAIcompaniesexplicitlyaimtobuildsuchsystems,98andbelievethattheymaysucceedthisdecade.99Somesurveysofpublishedmachinelearningresearchershavefoundthemedianrespondentpredictsagreaterthan10%chanceofhuman-levelmachine

intelligenceby2035,thoughthesesurveyshavebeencritiqued.100Attemptsatforecastingthedevelopmentofhuman-levelmachineintelligencebasedonhistorictrendsincomputingcostsandgrowthinAIresearchinputssometimesconcludethatthereisagreaterthan10%

probabilityby2035.101

However,thereisalargeamountofuncertaintyaboutthetimelinetothesecapabilities.Many,

ifnotmost,otherresearchersdonotexpectAIsystemsthatgenerallymatchhuman

performancewithintwentyyearsanddonotagreethatitisaconcern.102Historically,andfrequently,therehavebeenpredictionsofimminentAIbreakthroughsthatdidnotcometopass.103

WhatrisksdofrontierAIpresent?

WemustunderstandtherisksassociatedwithfrontierAItosafelyaccessandseizetheopportunitiesandbenefitsthetechnologybrings.

Inthissection,wefirstreviewseveralcross-cuttingriskfactors–technicalandsocietal

conditionsthatcouldaggravatean