人工智能時(shí)代的經(jīng)濟(jì)政策挑戰(zhàn)

1

EconomicPolicyChallengesfortheAgeofAI1

AntonKorinek(UniversityofVirginia,CSHVienna,andGovAI)

September2024

ThispaperexaminestheprofoundchallengesthattransformativeadvancesinAItowardsArtificialGeneralIntelligence(AGI)willposeforeconomistsandeconomicpolicymakers.IexaminehowtheAgeofAIwillrevolutionizethebasicstructureofoureconomiesbydiminishingtheroleoflabor,leadingtounprecedentedproductivitygainsbutraisingconcernsaboutjobdisruption,incomedistribution,andthevalueofeducationandhumancapital.Iexplorewhatrolesmayremainforlaborpost-AGI,andwhichproductionfactorswillgrowinimportance.ThepaperthenidentifieseightkeychallengesforeconomicpolicyintheAgeofAI:(1)inequalityandincomedistribution,(2)educationandskilldevelopment,(3)socialandpoliticalstability,(4)macroeconomicpolicy,(5)antitrustandmarketregulation,(6)intellectualproperty,(7)environmentalimplications,and(8)globalAIgovernance.Itconcludesbyemphasizinghoweconomistscancontributetoabetterunderstandingofthesechallenges.

I.Introduction

Therapidriseofartificialintelligence(AI)willpresentsignificantchallengesforeconomicpolicy.ThereleaseoflargelanguagemodelssuchasChatGPTinlate2022hasheightenedpublic

awarenessofAI'spotentialandstimulateddiscourseonitssocietalimpacts,yetitrepresentsjustonesteppingstoneonthelargertrajectoryofAIadvancement.AsAIcapabilitiescontinuetogrow,itbecomesincreasinglycrucialtore-evaluateoureconomicandpolicyframeworkstoensuretheirrelevanceandappropriatenessfortheAgeofAI.

ThispaperexaminestheeconomicimplicationsofadvancedAI,withparticularemphasisonthe

potentiallyimpendingdevelopmentofArtificialGeneralIntelligence(AGI)–AIsystemscapableofperformingintellectualtasksatorabovethehumanlevelacrossdomains.Iarguethattheadventof

AGIcouldfundamentallyalterthedynamicsofoureconomicsystem,necessitatingacomprehensivereassessmentofeconomictheoryandpolicy.

1ThispaperwaspreparedforaconferenceorganizedbythePetersonInstituteforInternationalEconomics

andtheInternationalMonetaryFundon“RethinkingEconomicPolicy:SteeringStructuralChange.”The

authorwouldliketothankJustinBullock,DuncanCass-Beggs,TomCunningham,BillGates,TimLaseter,

SamManning,DonghyunSuh,JoeStiglitz,andPetiaTopalovaaswellastheparticipantsoftheconferencefortheirthoughtfulcommentsanddiscussions.FinancialsupportfromtheUniversityofVirginia'sBankardFundforPoliticalEconomyandDardenSchoolofBusinessisgratefullyacknowledged.

2

IstartwithabriefdescriptionofthetrajectoryoftechnologicaladvancesdrivingAIprogress,

emphasizingtheexponentialgrowthincomputationalpowerandincreasesinalgorithmic

efficiency.ThenIdrawontheperspectivesofeminentAIresearchersandindustryleaderstolayoutpotentialtrajectoriesforAGIdevelopmentanddiscusskeyindicatorsthatmaysignalitsapproach.

Mymainfocusistheeconomicimplications.IarguethattheemergenceofAGIwillmarktheendoftheIndustrialAgeandusherinaneweconomicera,justliketheIndustrialAgeendedthe

Malthusianera.Thischangeinparadigmwillinvolveasignificantshiftinwhatfactorsofproductionarerelevantfortheeconomy,awayfromlabortoreproduciblefactorssuchascomputational

resourcesandrobots,withconsequentimplicationsforproductivityandoutputgrowth.

ThepaperproceedstoraiseeightchallengesthatadvancedAIposesforeconomicsandeconomicpolicy.First,Idiscussthechallengeofincomedistributionandinequalityinaworldwherehumanlabormaybecomeincreasinglysubstitutable.Second,Iassessthedevaluationofhumancapitalandtheneedtorecalibrateapproachestoeducationandskilldevelopment.Third,Ifocusonthe

resultingchallengetosocialandpoliticalstability.Fourth,Idiscussmacroeconomic

considerations,includingshiftsinaggregatedemandinahighlyautomatedeconomyinwhich

labor’srelevancediminishes,andpotentialadaptationsformonetaryandfiscalpolicies.Fifthandsixth,Ibringuptheimplicationsforantitrustpolicyandintellectualpropertyregimes.Seventh,I

examinetheenvironmentalchallengesfromrapidAIdevelopment.Eighth,Iraiseinternational

economicpolicyandgovernancechallenges,includingthemanagementofAI-drivengeopoliticaldynamicsandthemitigationofpotentialglobaldisparitiesinAIcapabilities.SuccessfullypreparingoureconomyandsocietyforthepotentialofAGIwillrequiremeetingallthedescribedchallenges–andlikelymanymore.

Thepaperconcludesbyreflectingontheroleofeconomicanalysisinaddressingthesechallenges,andbyemphasizinghowAIitselfcanbeleveragedasatoolforeconomicanalysis.Throughoutthepaper,IemphasizetheimportanceofproactivepreparationsforthechallengesandopportunitiespresentedbyadvancedAI.Asweapproachapotentiallytransformativetechnologicalshift,itis

imperativethatweexamineandadaptoureconomicpoliciestothenovelchallengesoftheAIage.Thispaperaimstocontributetothiscrucialresearchagenda.

II.TechnologicalCapabilities

ThefieldofAIhaswitnessedunprecedentedgrowthinrecentyears,withprogressacceleratingatapacethatchallengesourtraditionalunderstandingoftechnologicaladvancement.Thissection

exploresthekeydriversbehindthisrapidprogressandexaminesthepotentialtrajectoriesforfutureAIdevelopment.

1.Exponentialprogressincomputing

Moore'sLawhaslongbeenakeybenchmarkforprogressinthedomainofcomputation.Translatedintoeconomicterms,Moore(1965)predictedadoublingintheefficiencyofmicroprocessorseverytwoyears.Thisregularityhasapproximatelyheldformorethanahalf-centurynow,enablingrapidefficiencygainsincomputing,whichinturnunderpinnedtheriseofAIinrecentdecades.

3

ThepaceofgrowthinthecomplexityofAIsystems,however,hasfaroutstrippedeventheserapidefficiencygains.AsobservedbySevillaandRoldan(2024),theamountofcomputationalpower(or“compute”)employedintrainingcutting-edgeAIsystemshasdoubledeverysixmonthsoverthepastdecade.ThistrendisvisuallyrepresentedinFigure1.

Figure1:Thetrainingcomputeemployedbythemostcutting-edgeAImodelsdoubledonaverageeverysixmonthsforthepast15years.DistributedunderaCC-BY4.0licensebyEpoch.

Thetrainingcomputeofthesystemsdepictedinthefigurehasgrownmuchfasterthanwhatis

impliedbyMoore’sLawbecausetheinvestmentsincomputebyfrontierAIcompanieshaveon

averagetripledeveryyearovertheperiod

.2

EstimatesofthecostofcurrentfrontierAIsystemsareintherangeofhundredsofmillionsofdollars,potentiallysoonsurpassingthebilliondollarthreshold.IndustryobserverspredictthattheexponentialgrowthofthecomputeinvestedinfrontierAI

modelswillcontinueforatleastanotherthreetofiveyears,andpossiblylonger,giventherecenteconomicsuccessofAImodels(Sevillaetal,2024).Ifcurrenttrendscontinue,wemayseetrilliondollartrainingrunsforfrontierAImodelsbytheendofthedecade.

Therapidgrowthinthecomputationalresourcesdeployediscomplementedbysignificant

improvementsinalgorithmicefficiency.Hoetal(2024)reportthatthealgorithmicefficiencywith

whichAIsystemsutilizecomputehasgrownatarateofabout2.5xperyear.Whenmultiplyingtheseefficiencygainswiththe4xincreaseincomputeusestemmingfromMoore’sLawandincreased

investmentincompute,thesefactorsresultinastaggering10xincreaseintheeffectivecompute

2Adoublingofcomputeeverysixmonthsimpliesa16-foldincreaseevery2years.SinceMoore’sLawimplies“only”adoublingofchipefficiencyeverytwoyears,an8-foldincreaseinspendingoncomputeisrequired

everytwoyearstoachievetheobservedoverallgrowthofcompute–almostamountingtoatriplingperyear.

4

usedfortrainingfrontierAIsystemseverysingleyear.Inaddition,themostrecentgenerationofAImodelsthatcanperformreasoning-relatedtasks,suchasOpenAI’so1,alsoemploysubstantialamountsofcomputeattheinferencestage,i.e.,whentheyrespondingtouserinquiries.

Theimplicationsofthisrapidgrowthincomputeareprofound.AsAIsystemsbecomemore

powerful,theycantackleincreasinglycomplextasks,leadingtobreakthroughsinmoreandmoredomainsofhumancognitivework.ThecomplexityoffrontierAImodelsisrapidlyapproachingourbestestimatesofthecomplexityoftheneuralnetworksinthehumanbrain(Carlsmith,2020).

Moreover,advancesinroboticssuchashumanoidrobotsarealsoproceedingatarapidpace(see,e.g.,Duetal.,2024).

2.Computeandcapabilities

WhileinputstoAIandalgorithmicefficiencyareimportant,whatmattersfromaneconomic

perspectiveishowtheseinputstranslateintooutputs,i.e.,intoAIsystems’capabilities.Thisiswhatweturntonext.

ScalinglawsAdvancesinAIcapabilitiesfollowpredictableregularities–so-calledscalinglaws–thatdescribehowmuchbetterAIsystemsbecomeatpredictionasthenumberofmodel

parametersandtheamountoftrainingdataincrease(Kaplanetal,2020;Hoffmanetal,2022).

Together,theparametercountandamountoftrainingdatadeterminehowmuchcomputeis

neededtotrainamodel.ThedescribedscalinglawshaveheldformorethanadecadenowandareattheheartofthestrategiesemployedbyfrontierAIlabsintheirpursuitofAGI.Thepredictive

poweroftheselawsallowsresearcherstoestimatehowmuchimprovementinAIcapabilitiescanbeexpectedfromfurtherincreasesincomputationalresources,guidingbothresearchand

investmentbyAIlabs.Morerecently,OpenAI(2024)foundthatscalinglawsapplynotonlytothemodeltrainingstage,butalsowhenlettingAImodelsdeliberatebeforegivinganswersatthe

inferencestage,implyingthatmorecomputeispredictablyassociatedwithbetteranswers.

Moregenerally,neuralnetworks,thefundamentalbuildingblocksofmodernAIsystems,are

universalapproximatorsofarbitraryfunctions(Cybenko,1989;Horniketal,1989).Thismeansthat,intheory,theycanperformanyinformationprocessingtaskarbitrarilywell,givensufficient

resourcesandtrainingdata.Theexperienceofthepastdecadehasshownthattheyareinfactveryefficientfunctionapproximatorsforawiderangeofreal-worldusecases.ThispropertyunderpinsthepotentialofAIsystemsacrossawiderangeofapplications–includingthepossibilityof

performingalltheintellectualfunctionsofthehumanbrain.

3.PredictionsofAGI

OneofthecrucialquestionsregardingthefutureimpactofAIiswhetherandwhenAIsystemswillmatchhumanlevelsofintelligenceacrossallcognitivetasksthathumanscanperform,asthis

wouldcreatethepotentialforwidespreadlaborsubstitution.AIsystemsthatcandothisare

termedartificialgeneralintelligence(AGI)todistinguishthemfromnarrowAIsystemsthatcanonlyperformindividualtasksorasmallrangeoftasks,suchasimagerecognitionorvoicetranscription(Morrisetal.,2024).

5

Justafewyearsago,allAIsystemswerenarrow.However,modernlargelanguagemodels(LLMs)andothertypesofgenerativeAIsincetheearly2020sareincreasinglygeneralpurpose,performingagrowingrangeofcognitivefunctions,fromcreativetaskstosimpleformsofreasoning.

TherapidprogressoverthepastfewyearshascompressedthetimelinesforachievingAGI.Inthe2010s,themedianestimateofAIresearchersforwhenAGIwouldbereachedwasinthesecond

halfofthe21stcentury(Graceetal.,2018).Morerecently,agrowingnumberofAIresearchersandindustryleadershaveofferedmuchshortertimelines.GeoffreyHinton,oneofthethree"godfathersofdeeplearning"whowontheTuringawardfortheircontributionstothefield,proclaimedthat"[hehad]suddenlyswitched[his]viewsonwhetherthesethingsaregoingtobemoreintelligentthanus"andexpectsAGItobereachedin“5to20yearsbutwithoutmuchconfidence[since]weliveinveryuncertaintimes”(Hinton,2023).Likewise,SamAltman,CEOofOpenAI,statedthat"AGIwillbea

realityin5years,giveortake"inearly2024.Thesepredictionsarealsoinlinewiththeexpectationsofthegeneralpublic(Pauketatetal.,2023).

AGIwouldbetransformativeforoureconomyandsociety,creatingbothopportunitiesandrisksofunprecedentedscope.Thishasledtoagrowingsenseofalarmamongtechnologyexpertswhoseepotentiallyradicalchangesoccurringwithinashortperiod.TheyseetheexponentialgrowthofAIandassociatedscalinglawsandwonderwhytheworldisnotmorealarmed.

GeoffreyHinton’swarningsareakintoAlbertEinstein’slettertoUSPresidentFranklinDelano

Roosevelt,inwhichhewarnedofthepossibilitythatnuclearpowercouldbeharnessedforatomicweapons,givingrisetothenuclearage.Therewassignificantuncertaintyabouthowrealisticthe

predictionsofEinsteinandotherexpertswereatthetime,andhowsoontheirpredictionswouldbe

realized.In1933,ErnestRutherfordcalledthenotionthatnuclearpowercouldeverbeharnessed“moonshine.”Wearefacingsimilaruncertaintytoday,butthescalingofAIisnonetheless

proceedingrelentlessly.

4.HarbingersofAGI

EconomistshaveanimportantroletoplayinshapingthedirectionofprogresstowardsAGI(KorinekandStiglitz,2022)andinhelpingoursocietytoprepareforandadapttotheresultingdisruptionsaswellastointegratetheincreasinglycapableAIsystemsintooureconomicsystem.

GiventheuncertaintyaboutwhetherandhowquicklythedevelopmentofAGImightcometopass,itisimportanttocloselyfollowadvancesinAIandregularlyupdatepredictions.Todoso,itisusefultomonitorfourkeycategoriesofindicators:

1.AIResearchBreakthroughs:TrackingmajoradvancesinAIresearchanddevelopmentis

perhapsthemostdirectindicatorofprogressinAI,althoughitisdifficulttoappreciatethe

overallimpactofaseriesofrelativelysmallindividualadvances.Forexample,leadingAGIlabsarecurrentlyfocusedonimprovingthereasoningcapabilitiesofAIsystems–aweakpointof

LLMs.ThismayrequireapproachesthatgobeyondthecurrentarchitectureofLLMs.Someof

therecentadvancesinthisdomainwereGoogleDeepMind'sAlphaProofandAlphaGeometry2,whichsolvedmathproblemsofequivalentdifficultytoSilverMedalistsatthe2024InternationalMathOlympiad(DeepMindAlphaProofandAlphaGeometryTeams,2024).Moreover,monitoringthetrajectoryandspeedofAImodelsimprovementallowsobserverstorecognizetrendsand

6

makebetter-informedpredictions.

Aparticularlyimportanttypeofresearchbreakthroughareadvancesthatautomatethe

researchprocessitself,whichcansignificantlyacceleratethecreationofbetterAIsystemsinthefuture.IfAIsystemscanproducebetterAIsystemswithlessandlesshumaninput,itmarksanimportantsteptowardsAGI.AIsystemsarealreadycontributingtotheirowndevelopment.Pengetal.(2023)demonstratethataccesstoAI-poweredcodingassistantsallows

programmerstocompletecodingtasks56percentfaster–andthisalsoappliestothe

programmersoffutureAIsystems.Luetal.(2024)describeanAIagentthattheyterm“AI

scientist,”whichcanautonomouslyproduceideasforpapersincomputerscience,execute

them,andwriteuptheresults,givingrisetonovelresearchinsights.AIisalsobeingusedinthedesignofAIchips,potentiallyacceleratinghardwaredevelopmentforAIsystems(Mok,2024).

2.AIProductReleases:TrackingtheperformanceofnewAIproductsonestablishedbenchmarkscanprovideadditionalconcreteevidenceofAIcapabilities.Thisishighlyusefulforeveryone

whoperformscognitivework–forexample,ImyselfcontinuallytrackprogressingenerativeAIsystemstobetterunderstandhowIcanoptimallydeploytheminmyresearch.Thisallowsmetohaveafirst-personperspectiveonhowAIisrevolutionizingtheconductofeconomic

research(Korinek,2023a,2024).Thereisnogoodsubstituteforhavingthistypeofsubjectiveexperience.

3.AIDeploymentAcrosstheEconomy:Lastly,monitoringtheadoptionandimpactofAI

technologiesinworksettingscanofferareal-worldperspectiveonthepracticalAI

advancements.Thiscanmosteasilybeobservedinone’sownareaofwork–forexample,in

economics,generativeAIhasbeenadoptedheavilybyyoungerresearcherssuchasPhD

studentsandassistantprofessors,whereasoldergenerationsofeconomistslagbehind,

makingitharderforthemtoassessadvancesinAI.OftentimesIlearnaboutnewAIcapabilitiesbylookingovertheshouldersofmyyoungercolleagues.Naturally,acompletepicturecanonlybeobtainedbymonitoringthedeploymentofAIacrossdifferentsectorsoftheeconomytogetabetterunderstandingofwhattaskscanorcannotbeperformedbyAIandhowclosewearetoAGI(see,e.g.,Bonneyetal,2024).

4.FinancialMarketTrends:Financialmarketsareamirroroftheexpectationsoftheir

participantsandthereforereflectpublicexpectationsofAGI.Inprimarymarkets,the

expectationofAGIislikelytoleadtoagrowinginfluxofinvestmentsintoAI,bothonthe

hardwareandsoftwareside.Insecondarymarkets,stockpricesofexistingcorporationsthat

produce,incorporate,orotherwisebenefitfromAIarelikelytoseelargerun-upsinassetprices.Moreover,interestratesmayriseinparallelwithapick-upingrowthexpectations(Chowetal.,2023).Althoughfinancialmarketsareinherentlynoisyandpronetoboomsandbusts,their

forward-lookingnatureprovidesausefulcomplementtorealindicators.

III.EconomicimplicationsofAGI

ThissectiondiscussestheimplicationsofAGIforeconomicsandeconomicpolicy.Forthesakeofclarity,IwillcutthroughtheuncertaintyregardingthedevelopmentofAGIandtakeitasgiventhatAGIwillbereachedwithintheforeseeablefuture,inlinewithHinton’sprediction.Moreover,Iwill

7

assumethatprogressinroboticswillproceed,andthatAGI-poweredrobotscanperformanyphysicalworkthathumanscanperformsoonafterAGIisreached

.3

RigorouslypreparingeconomicpolicyfortheAgeofAGIwillrequirescenarioplanningtoexaminetheimpactofalternativeeconomicpoliciesunderdifferentscenariosfortechnological

advancement.Korinek(2023b)andKorinekandSuh(2024)provideexamples.However,eveniffullAGIisnotachievedinthenearfuture,theeconomicchallengesdiscussedbelowaredirectionallycorrectandrelevant.

1.Historicaleconomicparadigms

TounderstandtheeconomicimplicationsofachangeastransformativeasAGI,itisusefultotakeastepbackandlookatthelonger-termtrajectoryofoureconomicsystem,startingwiththe

MalthusianAge.IwillthenexplorewhatchangedintheIndustrialAgeandwhatislikelytochangeintheAgeofAGI,focusingontheeffectsonoutputandwages.

TheMalthusianAge

IntheMalthusianAge,theeconomywascharacterizedbyaproductionfunctionoftheformY=AF(T,L)

wheretechnology,A,waslargelystagnant(orgrowingveryslowly),land,T,wasthescarceresource,

andhumanswhoprovidedlabor,L,werereproducible.Inthisera,populationgrowthwas

constrainedbytheresourcestheeconomycouldproduce.Giventhefixedfactorland,anincreaseinpopulationimpliedadecreasingmarginalproducttohumanlabor.Humansreproducedtothepointwhereourmarginalproductequaledoursubsistencelevel.Alargerpopulationcouldnotbesupportedsoanyadditionalpopulationgrowthledtodestitution,hunger,anddeath.Giventhis

situation,landwasthemostvaluableproductionfactor–farmorevaluableinrelativetermsthanitistoday.Humansearnedasubsistenceincomethatwasjustenoughtosurvive.Fromapurely

economicperspective,themarginalhumanwasdispensable.

TheIndustrialAge

Theeconomyoftheindustrialagehasbeenthemainfocusofmoderneconomics–itcapturesthe

worldwearestilllivingin.ItischaracterizedbyaproductionfunctionoftheformY=AF(K,L)

wheregrowthintechnology,A,isthekeydriverofeconomicgrowth,capital,K,isthereproduciblefactorandaccompaniedthegrowthinA,andlabor,L,growssoslowlythatisusuallytakenas

exogenous.SincetheIndustrialRevolution,output,Y,inadvancedcountrieshasgrownabout20-fold,drivenbycomparativelyrapidtechnologicaladvances.

3RoboticshasrecentlyexperiencedbreakthroughsbycombiningrobotswithmorecapablebrainsintheformofmodernAIsystems.Robotsthatcanperformanyhumantaskswilllikelytakelongerthan5years,i.e.,theywillmaterializemoreslowlyifAGIhappenssoon.However,AGIwilllikelydesignrobotsthatcanperformanyphysicaltaskifhumanshavenotgottentheirfirst.

8

Giventheseadvances,laborsuddenlybecamethescarcefactorofproductionandexperiencedlargeincreasesinitsmarginalproductinparallelwiththe20-foldincreaseinoutput,givingrisetothelevelofwagesthatweseetoday.Inshort,labortransitionedfrombeingdispensableto

becomingthebottleneckproductionfactoroftheeconomy.

TheAgeofAGI

WhenAGIisreached,AIsystemscan,bydefinition,performanycognitiveworkthathumanscanperform,and,perourearlierassumption,willsoonperformanyphysicalworkaswell,givingrisetoaproductionfunctionoftheform

Y=AF(K,L+M)

inwhichthenewvariable,M,capturesmachines,intheformofAIcomputeandrobots.Inthissimplemodel,machinesareaperfectsubstituteforhumanlabor,i.e.,forbothhumancomputeandhumanphysicalcapabilities

.4

InanAGI-poweredworld,growthintechnologyAwillalsoaccelerate,asartificialbrainpowerandrobotscandrivescientificprogressandinnovation.BothtraditionalphysicalcapitalKandthenewmachineswillbereproducibleresources.Thedistinctionbetweenthetwowillincreasinglyblur.

Theywillaccumulatewithoutboundsandgenerateevermoreeconomiccapacity,inthespiritofRomer(1986)-styleAK-models.

Animmediateimplicationisthatthegrowthrateofoutputwillrise.Iftechnologyisacceleratingandallfactorsofproductionarereproducible,therearenomorebottleneckstooutputgrowthinthe

formoffixedfactors.

Theimplicationsforthemarginalproductoflaborarealsostraightforward.IntheMalthusianage,landwasthebottleneckfactor,andthemarginalproductoflaborwascompeteddowntothe

humansubsistencelevelbythecruelprocessofhumansmultiplyinguntiltheavailablelivingspaceandarablelandwereexhausted.TheIndustrialAgeusheredinaGoldenAgeforhumans:laborwassuddenlythebottleneckfactor,andtheaveragereturnstolaborgrewabout20-foldaboveour

subsistenceneeds.IntheAgeofAGI,there'snoreasonforlabortocontinuetoplaysuchaspecialrole.IfAIcansolveallcomputationalproblemsthatourbrainscansolve,whyshouldhumanlaborhaveaspecialstatusinoureconomicsystem?Laborwouldbeaperfectsubstituteformachinessothetwowouldbeequallyscarce.Asmoreandmoremachinesareaccumulated,thevalueoflaborandbyextensionwageswilldiminish.Thiswillfundamentallychallengeourpresentsystemof

incomedistribution.

However,thedivergenceofoutputgrowthandthemarginalproductoflaboralsosuggeststhe

contoursofasolution.WhatisneededtocreatesharedprosperityistotakeasmallsliveroftherapidlyrisingoutputfromAGIandgiveittotheworkerswhoselabormarketvalueisdiminished.ThiscreatesthepotentialforaParetoimprovement,althoughsuchasolutionmayfacesignificantpoliticaleconomychallenges(see,e.g.,BellandKorinek,2023).

4AmorecomprehensiveproductionfunctionunderAGIcoulddistinguishlaborintoitscomponentsphysicalandcognitivelaborandobservethatAIandrobotsrespectivelysubstituteforeachofthetwocomponents.See,e.g.,Growiecetal(2024).

9

2.AGIscenarios

Toillustratehowtheforcesdescribedaboveunfoldinacomprehensivemacroeconomicmodel,Iprovideanoverviewofthefourmainexamplesfrom"ScenariosfortheTransitiontoAGI"byKorinekandSuh(2024).Ouranalysisisgroundedinatask-basedframeworkinthespiritofZeira(1999)andAcemogluandRestrepo(2018).Ourmaininnovationistorepresentworkasconsistingoftasks

varyingincomputationalcomplexitythatcanbeautomatedonlywhenAIreachesasufficientlevelofcapabilities.ThisapproachallowsustomodelhowexponentialadvancesinAIprogressively

automatemoretasksand,ifthereisamaximumleveloftaskcomplexityandAIsurpassesit,tocapturetheadventofAGI.

Weconsiderfourdistinctscenarios,eachbasedondifferentassumptionsaboutthedistributionoftaskcomplexityandtherateoftechnologicalprogress.Inallscenarios,weassumethatan

automationindex,representingthemaximumcomplexityofautomatabletasks,growsovertimeatanexogenousrate,reflectingreflectstheobservedtrendsincomputingpowerandAIcapabilitiesdiscussedbefore.Thekeydifferentiatorbetweenscenariosisthedistributionoftaskcomplexity.

1.Business-As-UsualScenario:ThisscenarioassumesthattaskcomplexityfollowsaPareto

distributionwithaninfiniterighttail,andthefractionofnon-automatedtasksdeclinesata

constantrate.Thisimpliesthattherearealwaystasksthatonlyhumanscanperformorthatwechoosenottoautomate.Underourassumptions,outputgrowssteadilyatabout2percentperyear.Forourparameterization,boththereturnstocapitalandthewagebillriseindefinitely,

approximatelyintandem.

Moregenerally,thereisaracebetweenautomationandcapitalaccumulation–automationtendstopulldownwageswhereastheaccumulationofcapitalthatiscomplementarytotheremainingtasksforlaborincreaseswages.Ifautomationproceedstooquickly,wagesmaydecline,eveniffullautomationisneverreached.

2.BaselineAGIScenario:Inthisscenario,thetaskcomplexityofhumanworkfollowsa

distributionwithafiniteupperbound,andfullautomationisreachedwithin20years,inlinewithHinton’supperrangeforwhenAGIwillbereached.Theeconomicimplicationsarestark:wagesgrowduringtheinitialperiodbutcollapsebefore