2030年的人工智能：人工智能技術的發展前景與潛在影響

AIin2030

Extrapolatingcurrenttrends

ThisEpochAIreportwascommissionedbyGoogleDeepMind.Allpointsofviewsandconclusionsexpressedarethoseoftheauthorsanddonot

necessarilyreflectthepositionorendorsementofGoogleDeepMind.

AIin2030|EpochAI1

AIin2030|EpochAI2

TableofContents

Executivesummary3

Introduction6

Scalingandcapabilities19

Scale25

Compute28

Investment33

Data39

Hardware45

Energyandtheenvironment50

Interlude:Fromscaletocapabilities57

Capabilities66

Howcapabilitiesaredeployed67

Softwareengineering72

Mathematics77

Molecularbiology83

Weatherpredictions89

Discussionandconclusion92

Appendix:AI?spotentialtoreduceGHGemissions94

Appendix:benchmarkextrapolationdetails99

AIin2030|EpochAI3

Executivesummary

HowwilladvancedAIbedeveloped,andwhatwillitseffectsbeintheworldatlarge?WhatwillhappenifcurrenttrendsinscalingupAIdevelopment

persistallthewayto2030?Thisreportexamineswhatthisscale-upwouldinvolveintermsofcompute,investment,data,hardware,andenergy.Weexploretheroleofcomputeacrossinferenceandtraining,thepromiseofeconomicvaluethatwouldbenecessarytojustifysuchinvestment,and

potentialchallengesindataavailabilityandenergy.

BasedonthesepredictionsforhowAIwillbedeveloped,weturntopredictfutureAIcapabilities,andtheimpactstheywillhaveinscientificR&D.AIforscienceistheexplicitgoalofseveralleadingAIdevelopers,andislikelytobeamongthetopprioritiesforAIdeployment.ScientificR&Dprovidesa

valuablelensforunderstandingwhatadvancedAIwillachieve.

ComputescalinghasplayedakeyroleinAIdevelopment,andwilllikelycontinuetodoso.ComputefortrainingandinferencedrivesimprovementsinAIcapabilities,andmuchprogressinAIresearchhascomefrom

developinggeneral-purposemethodstoenabletheuseofmorecompute.

ThetrajectoryofAIdevelopmentcanbeforecastedbasedoncontinuedcomputescaling.ScalinghassignificantimplicationsacrossmanyareasofAIdevelopment:trainingandinferencecompute,investment,data,

hardware,andenergy.Whenwepredictthatcomputescalingwillcontinue,wecanthenexaminetheconsequenceswithineachofthese—andhow

theyneedtoscaleaccordinglytoallowcomputescalingtrendstocontinue.

Exponentialgrowthwilllikelycontinueto2030acrossallkeytrends.

Acrosstrainingandinferencecompute,investment,data,hardware,and

energy,wearguethatacontinuationofexistingtrendsisfeasible.We

exploreeachfactorindetail,showinghowgrowthcouldcontinueto2030,anddiscussingthemostcrediblereasonsforslowdownoracceleration

beforethen.WearguethemostcrediblereasonsforadeviationfromtrendarechangesinsocietalcoordinationofAIdevelopment(e.g.investor

sentimentortightregulation),supplybottlenecksforAIclusters(e.g.chips

AIin2030|EpochAI4

orenergy),orparadigmaticshiftsinAIproduction(e.g.substantialR&Dautomation).

Oncurrenttrends,thelargestAImodelsof2030willrequireinvestmentsofhundredsofbillionsofdollars,and1,000xthecomputeoftoday?s

largestmodels.InvestmentofthisscaleispotentiallyjustifiedifAIcan

automatesignificanttasksintheeconomy.Thepresenttrendof3xannual

AIlabrevenuegrowthwouldleadtorevenuesexceedinghundredsof

billionsofdollarsbefore2030.Findingdataforsuchtrainingrunsmaybe

challenging,butbetweensyntheticdataandmultimodaldata,thisshouldbesurmountable.Trainingrunsofthisscalewillrequiregigawattsofelectricalpower,approachingtheaveragedemandofentirelargecities.

Continuedscalingwillleadtocontinuedprogressincapabilities.Oncea

taskbeginstoshowsubstantiveprogresswithscaling,performance

tendstopredictablyimprovewithfurtherscaling.ExistingAIbenchmarks,despitetheirlimitations,covermanycapabilitiesthatwouldbegenuinely

usefulifautomatedintherealworld.Thus,existingbenchmarkscaninformourpredictionsonAI?sfuturecapabilities.Thiswillbeanimperfectview,

shapedbytherepresentativenessofexistingbenchmarks,andlimitedto

wherewecanalreadymeasureprogress.Wediscussthesechallenges

furtherin

Interlude:fromscaletocapabilities

.Nevertheless,thisprovidesuswithacompellingbaselinepredictionforwhatAIwillbeabletodo.

Ataminimum,AIwillactasavaluabletoolforscientificR&D.AIsystemsalreadyexcelathelpingusersfindrelevantinformation,implementcode,andperformwell-definedpredictiontasksbasedoncopious

domain-specificdata.Allofthesecapabilitiesaresettocontinueimproving.

Forexample,AIwillbeabletoimplementcomplexscientificsoftware

fromnaturallanguage,assistmathematiciansformalisingproofsketches,andansweropen-endedquestionsaboutbiologyprotocols.Allofthese

examplesaretakenfromexistingAIbenchmarksshowingprogress,wheresimpleextrapolationsuggeststheywillbesolvedby2030.Moreover,AI

toolsfordomain-specificapplicationswillcontinuetoimprove.Forexample,AItoolsalreadyofferstate-of-the-artpredictionsforbiomolecule

AIin2030|EpochAI5

structure/interactionsandweatherforecasting,andinbothareas,progressissettocontinue.

AdvancedAIwilllikelyleadtoaflourishingofdesk-basedresearch,

whichwilllikelybenefitfromalloftheaboveadvances.In2030,therewillbemoresoftware,moremathematicalresults,moreearly-stagemolecularbiologyresearch,moremethodologicaladvancesinfieldssuchasweatherprediction.Areassuchassoftwareandmathematicshavefewer

experimentalbottlenecks,andareparticularlylikelytobenefitfromAIprogress.

Forexperimentalfields,deploymenttimelinesarecontingenton

hard-to-predictsociotechnicalchoices.Basedoncurrentdrugapproval

pipelines,thedrugsthatwillbeapprovedviaclinicaltrialsby2030are

alreadyintheR&Dpipelinetoday.AImightbecontributingtothedrug

developmentpipelineby2030,butwithinthecurrentregulatoryframework,itisunlikelythatcontributionsfromAIwillleadtoapprovedproducts

availableinthemarket.

TheresultisaworldwithincreasinglyabundantAI-mediateddigital

services,knowledge,andanalysis.By2030,itislikelythatanything

physicalinscientificR&Dwillhaveadvancedproportionallylessthan

anythingdigital.However,ifthesepredictionscometopass,therewillbecorrespondinglystrongincentives(andadditionalresources)toacceleratethroughthesebottlenecks.TheseeffortsmayalsobenefitfromAI,butareoutsidethescopeofthisreport.

Introduction

ComputescalingisthekeytoAIprogress.UsingmorecomputefortrainingandinferenceisfundamentallywhatallowsAIcapabilitiestoadvance.Othercrucialfactorssuchasalgorithmicinnovationsanddataareimportant

primarilyinrelationtoenablingcomputescaling.Wewillarguethismorethoroughlylater,butfornow,considertheimplicationsifthisistrue.

Whatcancomputescalingpredict?

AssumingthatcomputescalingdrivesAIprogress,wecanpredictthenearfutureofAIdevelopmentbyextrapolatingrecenttrendsincomputescaling,andthenecessaryinputssuchasinvestment,data,electricalpower,etc.

Wearguethatthebaselineforforecastingthesethingsshouldbetrend

extrapolation:examinehowtheyhavegrownrecently,investigatethe

causes,andassumethatrecentgrowthwillcontinueunlessthereissomeobviousreasontopreventthis.Thisapproachisacommonbaselinein

forecasting(Armstrong2001),andhasbeenappliedinseveralareasofAIforecasting(AmodeiandHernandez2018;Sevillaetal.2024).

Aslongasinvestmentkeepsgrowing,computecankeepscalingonits

currentexponentialtrenduntil2030.1Then,becauseAIprogressisfairly

predictablefromscaling,wecanpredictAIcapabilities.Predictionrequiresexistingprogressonarelevantbenchmark.Fortunately,manyrelevant

benchmarksalreadyprovideevidenceacrosseconomicallyvaluable

domains,scientificR&Dandotherwise.Andthesepredictionsofimprovedcapabilitiessuggestthatinvestmentincomputeislikelytocontinue

growing,becausesuchAIcapabilitieswouldhavelargeeconomicvalue.

ThisallowsustopredicttheinputstoAIdevelopment.Inaworldwherewe"justkeepscaling",howmuchcomputeisusedin2030?Howmuchis

investedinAIclusterstoachievethatcompute?Howmuchelectricalpower

1Whyshouldthetrendbeexponentialgrowth,ratherthansomeotherform?Thispropertyariseswhengrowthisproportionaltocurrentvalue.Thispatternisseenacrossmany

phenomenaintechnologicalandeconomicprogress–forexampleeconomicgrowth,investments,microchipadvances,etc.

AIin2030|EpochAI6

isneededtosupplythem?Howmuchdatawouldthereneedtobeforthecomputetobeproductivelyused?ThisalsoallowsustopredictthetasksthatAIislikelytobeabletodoataminimum.WhatsortsofcapabilitieswillAIhaveby2030?

Whycomputeratherthanalgorithmsordata?

Therearetwocommonobjectionstoascaling-focusedviewofAIprogress:algorithmicinnovationsanddata.Wearguethatalthoughtheycomplicate

thepicture,theyremaincompatiblewithit.

Algorithmicinnovationsplayavitalrole,buttheyarecloselypairedwith

computescaling.ToparaphrasetheBitterLesson,themostimportantandeffectivealgorithmicinnovationsaregeneral-purposemethodsthatenablecomputescaling.2,3Moreover,thereissomeevidencethatalgorithmic

innovationsrelyoncomputescalingfortheirdevelopment.Thissuggests

thatweshouldanticipatealgorithmicprogress,butenabledby,andfocusedon,computescaling.Nevertheless,thisisakeyuncertainty.Capabilities

couldimprovefasterthanpredictedhere,ifcomputeisnotabottleneck.

DataisessentialforAItraining,andthequalityofdatasetscansignificantlyinfluenceresults.However,therearetworeasonstothinkthatcomputeismoreofarate-limitinginput.First,computeismoreofabottleneckinthe

currentparadigmofAItraining,atleastforgeneral-purposeLLMs.We

couldscaleupforatleastafewmoreyearsusingexistingpublictextdataandothermodalities(

Data

).Second,itappearsincreasinglylikelythat

inferencescalingwillmaketrainingmorecompute-intensive,effectively

usingcomputetogeneratedataforreasoningtraining(

Datawon?trunout

by2030,althoughhuman-generatedtextmight

).Specificdatabottleneckscanbeimportantwithinparticularapplications,andwediscussthese

furtherin

CapabilitiesinscientificR&D

.Hence,wemustconsiderdata

2TheBitterLessoninbrief:“Thebiggestlessonthatcanbereadfrom70yearsofAIresearchisthatgeneralmethodsthatleveragecomputationareultimatelythemosteffective,andbyalargemargin?(Sutton2019).

3Examplesofgeneral-purposecompute-leveragingAIinnovationsfromthepastfewdecadesincludeConvolutionalNeuralNetworks,GPUacceleration,theTransformerarchitecture,andLargeLanguageModelpretraining.

AIin2030|EpochAI7

availabilitywhenweinvestigatescaling,butthisremainscompatiblewithascaling-focusedview.

Whatcan?tcomputescalingpredict?

Whatthisdoesn'tallowispredictingwhenwehavegeneralintelligence,i.e.AIthatcanperformanycognitivetaskatthelevelofaskilledhuman.Thisquestionsuffersfromtwomassiveuncertainties:gapsinAIbenchmarks

andourunderstandingofthem,andgapsincurrentAIcapabilitiesthatmightnotbefilledinthenextfiveyears.

Currentbenchmarksmightnotadequatelyrepresentthemost

difficult-for-AItasksthathumansdo.AndnotallbenchmarksshowAI

progressyet:therearetaskswhereAIdoesn'tyetshowmuchimprovementwithscalingsofar,forexample"autonomouslyproveanewsubstantive

mathematicaltheorem".ItisfundamentallyuncertainwhereAIwillhave

reachedexpert-levelperformancebythetimeitsolvesexistingbenchmarks(thatshowprogress).ItisalsofundamentallyuncertainwhenAIwillsolve

allexistingbenchmarks,becauseitonlyshowsprogressonsomeofthem.Nevertheless,itisfairlycertainthatAIwillsolvemanychallenging

benchmarksby2030,andthesehaveclearimplicationsforusefultasksthatAIwillbeabletoperform.

Meanwhile,gapsinthecapabilitiesofpresent-daygeneral-purposeAI

systemsshedsomelightonthecapabilitiesthatAIcouldfailtoachieveby

2030.AImodelsexcelatidentifyingrelevantinformationfromalarge

trainingcorpus,butfrequentlyveerintoillogicalhallucinations.Theyare

brilliantatingestinglargeamountsofdataandidentifyingunderlying

patterns,yetfailtoreliablyapplyreasoningstepsthatwouldseemnaturaltoahuman.Reliabilityandrobustnessareproblemsmorebroadly,although

theyhaveatleastshownincrementalimprovementwithscaling.AIexcelsatsolvingclosed-endedoptimisationproblemssuchasgames,yetstrugglestoperformconsequentialactionsintherealworldwithagency.Itcan

performshallowprocessingoflongcontentmuchfasterthanahuman,butitstrugglestousethislong-contextinformationforsolvingchallenging

problems.ThereareenoughgapsincurrentAIcapabilitiesthatitishardtoevenbecertainwhichofthemareoverlapping–perhapslong-context

AIin2030|EpochAI8

comprehensionisrelatedtorobustnessofreasoning,orperhapstheyareentirelyseparateproblems.TheselimitationsarerelatedtothechallengeswithdesigningandinterpretingAIbenchmarks:adequatelybenchmarkingtheselimitationsisalsoanopenproblem.

ItisuncertainwhichoftheseAIlimitationswillimproveby2030,andby

howmuch.Itisuncertainwhetherthesewillimproveby"justscaling"

existingsystemswithsmallmodifications,butalso,itisunclearhowmuchcomputetheywouldneed.Considertheexampleofreasoningmodels:

pre-existingsystemsalreadyusedinferencescaling,butreinforcement

learning(RL)madethisfarmoreeffective,yieldingbreakthroughresultsinseveralbenchmarks.Doesthischallengetheviewthatscalingisthedriverofprogress?Arguinginfavourofscaling-drivenprogress,many

researcherspredictedaheadoftimethatbetterinferencescalingwouldbenecessary,andthisarrivedaftermodelsscaledupsufficientlyforreasoningRLtowork.Furthermore,trainingscalingholdsforRL:usingmoreRL

trainingcomputeimprovesthecapabilitiesthatreasoningmodelsachieve.Ontheotherhand,thisemphasisesthechallengesofpredictionfrom

existingresults.TheareaswhereAIstrugglestodaycansometimesseebreakthroughalgorithmicprogress,andthisisinherentlyhardtopredict.

Eveninascaling-firstview,itisunclearhowmuchmorescalingneedstohappentoreachAGI.Itisalsounclearwhetherthiswillrequiresignificantalgorithmicadvances.Itisuncertainwhethersuchalgorithmicadvances,ifneeded,mightstillhappenby2030.Thesearethebiggestchallengesto

scaling-focusedpredictionsforAI,andparticularlywhentryingtopredictbeyondcontinuingprogressonalready-progressingbenchmarks.

Despitethesesignificantchallenges,scaling-focusedpredictionsarestill

useful.Wecanpredictaminimalbaseline:tasksthatweexpectAIto

continueimprovingatwithfurtherscaling.Wecanthenexaminehowthe

resultingcapabilitieswouldaffectrealworktasks.Wecanreflectonfurthertasksthatarenotcoveredbythebaseline,andtheimplicationsfor

automationifAIdidbecomecapableofthese.Andthen,finally,wecan

followthroughtoreasonaboutthebroaderimplicationsthiswouldhave

withinpeople'swork.ThisallowsustobridgetwocompetingviewsofAI:AIasapowerfultool,andAIasavirtualworker.

AIin2030|EpochAI9

WhatdoesscalingpredictaboutAldevelopment?

Weexamineseveralkeyinputs:compute,investment,data,hardware,andenergyandtheenvironment.Anyofthesecouldunderminecontinued

progress–forexample,whatifcomputescalingstopsbeingeffective?4Whatifwerunoutofdata?Someoftheseargumentsarestrongerthanothers,butweseenosinglecompellingargumenttopreventcurrent

progresscontinuingto2030.Weexploretheimplicationsofthisacrosseachofthekeyinputs:surveyinghowfartheymightscale,andhowtheymightbederailed.

lnbrief,wepredictthat,oncurrenttrends,leadingAlmodelsin2030willbetrainedwith1,000xthecomputeoftoday’sleadingmodels.Theclusters

usedfortrainingsuchmodelswouldrequireinvestmentoftwohundred

billiondollars,closeto1%ofpresent-dayUnitedStatesGDP.Trainingand

deploymentwillrequiregigawattsofelectricalpowerforthelargestmodels,andtotalAldatacentrepowercouldeasilygrowto2+%ofglobalelectricitydemand,similartothelevelofdemandfromelectricvehicles(around2%by2030[lEA2025d])orthelnternet(2_3%in2025[Roziteetal.2023]).

KEYFINDINGSFORAIDEVELOPMENTTRENDS

Compute:Trainingcomputehasincreased4_5xperyearsince2010,andislikelytocontinuegrowingatasimilarpace.By2030,oncurrenttrends,thelargestAlmodelsarelikelytobetrainedwith1,000xthecomputeusedin

today’sleadingmodels.Scalingupinferencecomputewillbeanother

importantsourceofcontinuingAlimprovement.Thisisunlikelytointerferewithscalingoftrainingcompute,andforagivenmodel,itslifetime

inferencecomputewillprobablybecomparabletoitstrainingcompute.

Investment:Toenabletrainingatthisscale,thenecessaryAlhardware

wouldcosthundredsofbillionsofdollarsoncurrenttrends.Theamortisedcostofdevelopingindividualmodelswouldbebillionsofdollars.These

4Wediscussfurtherin

Scalingisnot“hittingawall?,althoughitisgettingharder

thattherearemanydifferentsensesinwhichgrowinginvestmentscouldbecomeuncorrelatedwithfurtherAlcapabilitiesimprovements.

Alin2030|EpochAl10

projectionsalignwithcurrentAIinvestmentsandvaluations,aswellas

capitalexpenditureplansfromAIclusterdevelopers.FrontierAIlabs

alreadyearnbillionsfromchatbots,withrevenuesgrowing3xperyearinthelastcoupleofyears.IfAIcansignificantlyraisenetproductivityacrosstheeconomy,itwillbeworthtrillionsofdollars.Thiswouldjustify

substantialinvestmentsinitsdevelopment.WediscusslaterhowAIcouldachievesuchnetproductivitygains–thisdependsonboththecapabilitiesachieved,andbeingabletodeploythemcost-effectively.

Data:Datasetsforgeneral-purposeAItrainingrecentlygrewat2.7xper

year,butfurtherdatasetgrowthcouldchangesignificantly.Ashifttowardsmultimodalandsyntheticdatamaybenecessaryashigh-quality

human-generatedtextdatabecomesscarce.Recenttrendsinreasoning

trainingsuggestthatgrowinghuman-provideddataatamuchslowerratecouldneverthelessenablecomputescalingtocontinueviasyntheticdataforreasoningtraining.IfAIcapabilitiescontinuetoimprove,thenparticularsourcesofspecialistdatawillbecomeincreasinglyvaluable:essentially,

datatoenabletrainingonhigh-valueproblems.

Hardware:TotalinstalledcapacityforleadingAIchipsislikelytocontinuegrowing2.3xperyear,drivenbyproducingmorechipswithbetter

performance.5LargeAIclusters,inlinewithcurrenttrends,arealready

beingplannedanddevelopedforthelargestAIdevelopers.However,itis

likelythatlargeAIworkloadswillbeincreasinglydistributedacrossmultipledatacentrestoeasethedemandforelectricalpower.

Energyandtheenvironment:PowerdemandsforfrontierAI(bothtrainingandinference)arelikelytogrowataround2.1xperyear,andAIenergy

demandgenerallyisontracktogrowaround1.6xperyear.Inthiscase,AI

datacentreswouldgrowto1.2%ofglobalelectricitydemand.Dependingontheenergymixusedtopowerdatacentres,AIelectricityusecouldaccountfor0.03_0.3%ofglobalemissionsby2030.Althoughsignificant,thisis

muchsmallerthanprojectedemissionsfromcommercialflights(2.5%,[IEA2025a]).ThereisdemonstratedpotentialforAItoreduceemissionsinareas

52.3xperyeargrowthininstalledcomputingpowerisslowerthantheprojected4_5xperyeargrowthinfrontiertrainingcompute.Currently,individualfrontiertrainingrunsuse

about2.5%ofinstalledcapacitywhenoperating,sothereisroomfortrainingrunstogrowfasterthantotalcapacity.

AIin2030|EpochAI11

suchasenergyproduction,industrialprocessoptimisation,andtransport,butthisheavilydependsonsocietaldecisionsaboutdeploymentand

prioritisation.

WhatdoesscalingpredictaboutAIcapabilitiesandimpacts?

WhatcapabilitieswilltheAIsystemsof2030achieve,andwhatimpacts

mightithaveintheworld?Thisisanincrediblybroadquestion,andtomakeittractable,wenarrowourscopetoacriticalarea:automationofscientificR&D.AIforscientificR&DisexplicitlythegoalofseveralleadingAI

developers(Altman2023;Amodei2024;GoogleDeepmind,n.d.),and

occupiesanimportantpositionintheeconomyduetoitsabilitytoimproveproductivitymorebroadly.6WeexploreAI?spotentialforscientificR&D

acrossseveraldifferentareas:softwareengineering,mathematics,molecularbiology,andweatherprediction.

Aspreviouslydiscussed,ourpredictionsareanchoredonextrapolating

trendsinpresent-dayAIcapabilities.Therearetwomainreasonsthis

approachcouldbeoverlyaggressive.Thefirstreasonisthatifbenchmarksarenotrepresentativeofthecapabilitiestheyareintendedtomeasure.Weexaminethisfurtherwithineachindividualsectionof

Capabilitiesin

scientificR&D

.Inseveraldomains,suchassoftwareengineeringand

biology,thereisalreadysomeempiricalevidencesuggestingthat

benchmarkprogressiscorrelatedwithreal-worldprogress.Thesecond

reasonisthatbenchmarkprogresscouldbedeceptiveduetooverfitting.

Althoughthisisarealchallengeforcomparingmodelsatapointintime,webelieveitislessofaconcernforbroadlypredictingprogressacrosscomingyears.Benchmarksinthepastwerealsosubjecttooverfitting,but

nevertheless,solvingthemwenthand-in-handwithrelatedAIcapabilitiesprogress.Ifcurrentbenchmarksoverstateprogressduetooverfitting,then

6ThisisnottosaythatR&DwillnecessarilybethefirstormosteconomicallysignificantsetofactivitiestoseeAIautomation.TherearecredibleargumentsthatAIdevelopersface

largerincentives,andeasierchallenges,broadlyautomatingmany

tasksacrossthe

economy

.Nevertheless,AIdevelopers?explicitfocusonR&Dmotivatesustogiveitattentioninthiswork.

AIin2030|EpochAI12

ourextrapolationswillbeaggressive,butaslongasthereissomerealunderlyingprogress,theywillneverthelessbeinformative.

CapabilitiestrendssuggesttherewillbetremendousprogressinAIfor

scientificR&D,particularlyinareassuchassoftwareengineeringand

mathematics,whererealistictaskscanbetrainedonentirelyinsilico.To

offerconcreteexamples:by2030,existingbenchmarkprogresssuggestsAIwillbeabletoimplementcomplexscientificsoftwarefromnatural

language,assistmathematiciansformalisingproofsketches,andanswer

complexquestionsaboutbiologyprotocols.Wedescribethisfurtherbelow.

KEYFINDINGSFORAICAPABILITIESIN2030

Softwareengineering:Manyoftoday’sday-to-daytasksarelikelytobecomeautomatablebyAlagents.Existingbenchmarksbasedonwell-definedsoftwareissues,suchas

SWE-bench,areontracktobesolvedin2026.Currentprogressonsolvingdefined

hours-longscientificcodingandresearchengineeringproblems(RE_Bench)isslowe