Capgemini-生成人工智能：人工智能的下一章（英）

GENERATIVEAI

Thenextchapterof

Artificial

Intelligence

FOREWARD

Whileweobserveahugeadoptionof

GenerativeAIacrossorganizationsand

industries-nearlyall(96%)executivesciteGenerativeAIasahottopicofdiscussionintheirrespectiveboardrooms1–thetopicisnotnew.

TheriseofGenerativeAIispartofan

overallevolutionofAI-fromMLandDLexplosiontoLargeModelsmaturity-

leadingtoanAIbecomingnowmore

powerful,scalable,andaccessible.

TheNext-generationAIishere,driving

radicalbusinesstransformation.From

contentproduction,workflowtoproductinnovation,itisrevolutionizingthewaywecreate,interactandcollaborate,completelyshiftingatthesametimethewaywelookatAIasawhole.

Asweobserveanunprecedententhusiasmaroundit-74%ofexecutivesbelievethe

benefitsofGenerativeAIoutweighthe

associatedrisks2-Ethicsismorethen

evercriticalfororganizationsto

successfullyandresponsiblyimplement

GenerativeAIacrosstheirdatavalue

chain,andthereforeshouldabsolutely

notbeseenasthefifthwheelonthe

wagon.

Theexpectedbenefitsarehugeandas

abusinessleader,understandhow

GenerativeAIistransformingtheway

yourorganizationoperateisamust.

ThisEverestGroupreportexplores

wheredoesthetruevalueof

GenerativeAIlie,consideringthe

potentialpitfallsandsharingthekey

areastoprioritize.Indefinitive,asfor

anyData&AItopic,theway

organizationsshouldapproach

GenerativeAIstartsbybuildingthe

rightfoundationsincludingastrong

testing&trustlayer.Nodoubtthe

derivedoutcomeswilloutweightthe

risksiftailoredtoorganizations

specificitiesandbuiltwithsecure,

privacyprotectingandreliablehigh-

scaleGenerativesolutions.

Ifyouwouldliketocontinuethe

discussionandknowmoreabouthowcan

helpcustomizingGenerativeAIforyour

ownpurpose,pleasereachout

MARKOOST

CustomGenerativeforEnterprise

GlobalLeader,Capgemini

mark.oost@

1,2Source:

CapgeminiResearchInstitute,HarnessingthevalueofGenerativeAI:Topusecasesacrossindustries

EverestGrop

GenerativeAI:theNextChapter

ofArtificialIntelligence

ThisdocumenthasbeenlicensedtoCapgemini

VishalGupta,VicePresidentPriyaBhalla,PracticeDirector

Copyright?2023,EverestGlobal,Inc.Allrightsreserved.

Contents

Introduction03

AI:thejourneysofar04

GenAI–what’swiththehype?05

ThetruevalueofgenAI08

Potentialpitfalls–genAIisnot

10

abedofroses

Requisitestobuildasturdy

12

genAIstack

Conclusionandthewayforward14

|ThisdocumenthasbeenlicensedtoCapgemini

3

Introduction

Sinceitsconceptualizationin1956,AIhasbeenaremarkabletechnology,revolutionizingindustries,andredefining

human-machineinteraction.Thetechnologyhaspushedboundariesanduncoverednewfrontiersinthedigitalspace.Onesuch

remarkablebreakthroughthathascapturedtheimaginationof

researchers,innovators,businesses,andindividualsalikeisgenAI.Usingcomplexneuralnetworks,genAImodelsdevelopnew

contentinvariousformsandmodalities,suchastext,images,audios,videos,codes,andmore.

Inthisreport,weexaminetheadventofAI,tracingitsoriginsandfascinatinginnovations,upuntiltheemergenceofgenAI.Wethenexplorethetechnology’scapabilities,challenges,andthe

transformativeopportunitiesitpresents.AstheapplicationsofgenAIcontinuetoexpandacrossindustriesandwithitsabilityto

generatehuman-likecontentandmimichumancreativity,it

becomescrucialtoexploretheprofoundimpactitcanhaveonoursociety,economy,andeverydaylives.

Thereportrecognizesthatwhilethistechnologyholdsgreat

promise,italsopossessesinherentriskssuchasrisingconcerns

aboutdataprivacy,identitytheft,andmisinformation.Moreover,

accountabilityforitsconsequencesbecomesapressingconcernasAI-generatedcontentbecomesincreasinglyindistinguishablefromhuman-generatedcontent.Byaddressingtherisksandchallengeshead-onandadoptingindustrybestpractices,enterprisescan

unlockthetruepotentialofgenAIwhileethicallyandresponsiblyintegratingthisgroundbreakingtechnology.

|ThisdocumenthasbeenlicensedtoCapgemini

Spreadofdataandcompute

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

4

AI:thejourneysofar

JohnMcCarthyfirstusedthetermartificialintelligencein1956,butAImadeitsfirstappearancemuchearlierina1927filmtitledMetropolis,featuringahuman’srobotdouble.Thisinitialfictionalidea

sparkedaseriesofadvancesthathaveledtothemostadvanceddevelopmentinthehistoryofAItoday,genAI,whichcancreatefictionalcharactersandstoriesofitsown.Notably,atechnology

intendedtoenhancehumancapabilitiescannowpotentiallytakeoveramultitudeoftasksthathumansperformedtraditionally.ButAIdidnotreachthisstageinasprint;ithasbeenalongandchallenging

journeyinvolvingsignificantinvestments,numerousunsuccessfultrials,andbreakthroughadvances.Exhibit1providesanoverviewofthevariousstagesofAIdevelopment,innovation,andadoption

overtime.

EXHIBIT1

Milestonesinthejourney

Source:EverestGroup(2023)

Experiment

JohnMcCarthycoins

thetermAIatthe

DartmouthCollegeSummer

AIConference(1956)

Rapidenterprise

adoption;remarkablebreakthroughsin

deeplearningandgenAI

Advancesinprocessing

powerandcomputational

capabilities,enablingmore

complexAIalgorithms

Early2000

Enablement

Birthofcloudcomputing

withthelaunchofAWS,

GCP,andAzure

Enrichment

OpenAIlaunchesChatGPT(2022)

RiseofMLandexplorationoffoundationalAItechniques

2015andbeyond

1900s

Timeline

Thetermartificialintelligencebecamethebuzzwordofthetimeafteritsfirstappearance.However,thedevelopmentofAIdidnotreallybeginuntilthelate1960s,asthenecessarycomputingpoweranddatawerenotyetavailable,and,hence,mostoftheworkwasaroundthemathematicsofAI.Duringthe

1960sand1970s,AItechniquessuchasML,NLP,andcomputervisionwereestablished,whichlaidasolidfoundationforAItomakeinroadsintoourdailylives,pavingthewayforitswidespreadadoption.

|ThisdocumenthasbeenlicensedtoCapgemini

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

5

Theearly2000swasaperiodofsignificantprogressforAI.Athrivingecosystememergedthat

supportedAIinfrastructure.Advancesincomputingpower,storage,andnetworkingtechnologies

facilitatedtheprocessingofvastamountsofdatafortrainingAImodels.Thebirthofcloudcomputingin2006wasacatalystforincreasingAIdevelopment.Significantimprovementsinhardware,suchasthedevelopmentofpowerfulprocessors,GPUs,andspecializedchips,weremadeforAIworkloads.The

periodalsomarkedtheappearanceofathrivingecosystemofAIstart-ups.

AIdevelopmentandadoptionfast-trackedfrom2015.Notableimprovementsinalgorithmsandsoftware

tools,suchasthelaunchofopensourceAIsoftwareTensorFlowandPyTorch,madeiteasierfor

developerstobuildandscaleAIapplications.Thesedevelopments,coupledwithaccesstodynamiccomputingpowerthankstothecloud,enabledenterprisestoaccelerateAIadoption.

BothenterprisesandconsumersbecameincreasinglycomfortablewithAIinthepastdecade.Infact,

AIissodeeplyembeddedinourdailylivesthatitisalmostimpossibletoimagineaworldwithoutittoday.TheuseofAIhasalsosignificantlyscaledacrossenterprises.AccordingtoEverestGroup’s2023AIsurvey,96%ofenterpriseshavesuccessfullyimplementedAIinoneormoreoftheir

operations.

Forthelongesttime,AIcouldperformrepetitivetasks,suchasrecognizingpatternsoridentifyingobjects.ThatchangedwithOpenAI’slaunchofChatGPTonNovember30,2022.ChatGPTisanAI-poweredchatbottrainedonlargedatasetsofunlabeledtexttogeneratehuman-likeoutput.Weexamineitscapabilitiesnext.

GenAI–what’swiththehype?

Theworld’slargestandmostvaluableenterprisesareeithertalkingaboutgenAIorhavebeguntolaythefoundationsforitsimplementation.OpenAIresearcherIanGoodfellowiscreditedforcoiningthe

termgenerativeAIin2014.EverestGroupdefinesgenAIisafieldofAIthatcancreate,manipulate,andsynthesizenewcontentthatdidnotexistbeforeinvariousformsandmodalities.

ThankstoChatGPT,whichhasdemocratizedtheuseofAIandfundamentallychangedtheway

consumerssearchcontent,thecategoryistrendingmorethantheoverallgenAImarket.ThechatbothasputAI–whichwasearlierprivytotechnologycreators–intoconsumers’hands.

However,onemustbecarefulaboutthesynonymoususeofthetermsChatGPTandgenAI.WhilegenAIisafieldofAIwithgenerativecapabilities,ChatGPTisagenAIapplication.Beforeweprobethe

commercialandapplicationfacetsofgenAI,itisvitaltounderstandthedifferencesbetweenAIasweknowittoday(alsoknownasdecisionAI)andthedisrupternext-generationgenAI.

?

|ThisdocumenthasbeenlicensedtoCapgemini

6

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

Exhibit2showshowgenAImodelsdifferfromtraditionalMLmodelsandliststheirinput/outputfeatures:

EXHIBIT2

DecisionAIvs.genAI:acomparativeview

Source:EverestGroup(2023)

GenAI

Parameter

DecisionAI

Training

Canbetrainedonsmallerdatasetswith

Needlargedatasetswithan

parameters

fewerparameters

exponentiallyhighnumberoftrainingparameters

Trainingtime

andcost

Relativelycheaptotrainanddeploy

Relativelyquicktotrain

Hightraininganddeploymentcosts

Highcostofacquiringlargequalitydatasets

Significantlylongertrainingtime

Computeand

Canbetrainedandrunonstandard

Needspecializedhardwaresuchas

infrastructure

computinginfrastructure

GPUsandTPUs

Capability

Providepredictionsorclassificationsbasedonexistingdata

PerformspecificAIapplicationsonwhichtheyaretrained

Generativecapability–imageandvideosynthesis,textgeneration,speech

synthesis,codegeneration,etc.

General-purposemodelscapableofperformingmultipleAItasks

ThefundamentaldifferencebetweentraditionalMLmodelsandgenAImodelsisthenumberof

parameterstheyaretrainedon.Suchtraininghasbecomepossibleduetoincreasedavailabilityofqualitytrainingdataandhardwarecapacity,whichwerethetwobiggestconstraintsinthisfield.Forexample,trainingalargeimagemodelrequiresadatasetofmillionsofhigh-qualitylabeledimages,whiletraininganMLclassifiertorecognizespecificobjectsinimagesmayrequireadatasetof

thousandsoflabeledexamples.

Therisingnumberoftrainingparametersisanindicatorofincreasingmodelcomplexityandthe

model’sabilitytoperformmoregeneralizedtasks.However,foundationgenAImodels,whichare

trainedfornopurpose-definedtasksaretrainedonanextremelyhighnumberoftrainingparametersandrequirespecializedresources.Todate,onlylargetechgiantshavebeendevelopingfoundation

genAImodelsduetotheircomplexityandresourcerequirements.However,inthisraceofdevelopingbettergenAImodels,qualityisprovingtobemoreimportantthanquantity.CustomgenAImodels,

whicharedesignedforspecifictasksandaretrainedonsmallerbuttargeteddatasets,aregaininghightractioninthemarket.

|ThisdocumenthasbeenlicensedtoCapgemini

Numberoftrainingparameters(inmillion)

Logarithmicbase10

Jurassic-1

1,00,000廠YaLM

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

7

Exhibit3showstheprominentlargemodelslaunchedovertimeandtheirtrainingparametersizes.

EXHIBIT3

EvolutionofMLmodels

Source:TheAIIndex2023AnnualReportbyStanfordUniversity1andEverestGroupanalysis

.FoundationgenAImodels

AImodels

Customgen

1,00,00,000

.WuDao2.0

10,00,000PaLMGPT-4

Megatron-Turing..●Minerva

PanGu-?±GopherBLOOM-GPT-

3

GPT-3-1.rLaMDA

2

T5-11B

OPT●BloombergGPT

HyperClovaChinchillaGLMAlphaCodeLLaMA

TuringNLG

10,000StarCoder

.DALL-ECodexGPT-NeoXESMFoldFlan-UL2LParti

T5-3BMegatronGPT-J。CogE3.0。DALLsic-Alpaca.DALL-E2

1

Grover-Mega

.ERNIE

1,000LaMDA1Wu-NeoStableDiffuil.iffusion2.0

GPT-2

1002019onwardTimeline

Note:Therepresentationisnotexhaustiveandcoverslargemodelsthathaveapubliclydisclosednumberoftrainingparameters

Interestingly,ittookMeta2US$4.05milliontodevelopits65B-parameterLargeLanguageModel(LLM),LLaMA,whichwastrainedusing2048NVIDIA3A100GPUs.Thecostexemplifiesthesignificant

resourcesrequiredtodevelopgenAImodels,particularlyintermsofextensivecomputation.While

advancesintechnologyandtheavailabilityofsuperiordatasetsmayhelpbringdowndevelopment

costs,thecostofdevelopinggenerativemodelsisexpectedtoremainconsiderablyhigherthanthatfortraditionalMLmodels.ConsideringOpenAI’sGPTmodel,whichwastrainedon10,000suchGPUs,

onecanonlyimaginethescaleandassociatedcostsinvolved.Asthebenefitsandpotentialofthis

technologycontinuetounravel,theinvestmentitselfholdsgreatpromisefortransformingindustriesandunlockingnewpossibilitiesintheAIrealm.

1

TheAIIndex2023AnnualReportbyStanfordUniversity

2

MetaResearch

3

NVIDIAGPUpricing

|ThisdocumenthasbeenlicensedtoCapgemini

M&EProfessionalservices

RCG

BFSITravelandtransportPharmaceuticalsand lifesciences Education Manufacturing HealthcareTelecommunications

Publicsector

Energyandutilities

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

8

ThetruevalueofgenAI

ThesuccessandadoptionofgenAIdependsonseveralcrucialfactors.Whilewebelievethetechnologywillmakeitsimpactoneveryindustryinthefuture,someindustriesarepositionedtoadoptthistechnology

fasterthanothers.EverestGroupsoughttounderstandthereadinessoftheseindustriesforgenAIadoptionbyanalyzingfourparameters–currentdataavailability,technicalreadiness,regulatoryandcompliancerequirements,andcriticalityofcontentacrossindustries.

OuranalysisshowsthatMediaandEntertainment(M&E),professionalservices,RetailandConsumerGoods(RCG),Banking,FinancialServices,andInsurance(BFSI),andtravelandtransportarewell

positionedtoadoptthetechnologybeforeothers.Forexample,RCGcompanieshavebeendatabanksforconsumer-andproduct-centricdataforyearsandhavebeenattheforefrontoftechnologyadoptionwith

robustdataandinfrastructurefoundationstobuildupon.Additionally,whileeveryindustryhasacertainlevelofsensitivityforregulationsandcompliance,RCGhasfewerregulationsthanotherindustries.

SeveralenterprisesacrossotherverticalshavealsostartedexperimentingwithgenAI.Forexample,the

travelbookingcompanyeDreamsOdigeohaspartneredwithGoogletoimplementitsgenAIcapabilitiestopersonalizecustomerinteractions,whileSiemenshaspartneredwithMicrosofttousegenAIforautomaticinspectionnotescreationonthefactoryfloor.

Exhibit4providesinsightsintothereadinessforadoptinggenAIbyindustry.

EXHIBIT4

IndustryadoptionofgenAISource:EverestGroupanalysis

Availabilityofqualitydata

Technology

readiness

Regulationand

compliance

Needfor

content

generation*

Timetoadoption

LowHigh

Currentmarketmovements**

Within1year

Within1year

Within1year

Within1year

Within1year

1-2years

1-2years

1-2years

1-2years

>2years

>2years

>2years

*Webelievecontentasavectorwillbeakeydecision-makerforindustriestoadoptgenAIintheshortterm;however,inthelongerterm,theimpactofthisparameterwillneutralizeacrossallsectors

**Marketmovementsaretrackedbasedonenterprises’publicannouncementsofadoptionofgenAIusecasesacrossindustries.However,theseimplementationsareinPoCstagesanddonotindicateproductiondeploymentofthetechnologyatthispointintime

|ThisdocumenthasbeenlicensedtoCapgemini

Keyindustryusecases

Writingassistant

Articlesummary

Syntheticvoice

Text-to-music

Image/Videocreationandenhancement

Game

development

AIavatars

AI-generatedmediaposts

Generatingproduct

descriptions

Customer

interactionbots

Orderprocessing

Personalization

Sentimentanalysis

Product

personalization

Newproductdesigning

Sketch-to-design

Report

summarization

Unstructureddatasummarization

Financialbots

Insurancebots

Syntheticdataforrisksimulation

Contractassistant

Underwriting

Claimsprocessing

Travelitinerarydesigning

Travelbots

Horizontalusecases

Customer

servicebots

Callnotescreation/

Summarization

Automatic

email

responsesforcustomer

queries

Enterprisesearch

Employeeassistancebots

CRMbots

Automaticemails

Automatic

slidegenerator

Policydraftcreation

Contractcreation

AI-generatedjob

descriptions

L&Dcontentcreation

Campaignandadvertisementcreation

Content

personalization

Mediapostsand

promotionalcontent

Financial

statementspreparation

Contractassistant

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

9

Whilemostindustriesareintheexperimentationphase,innovativeusecasesareemergingeveryday.

Exhibit5highlightsprominentgenAIusecasesthataregeneratinginterestamongenterprises:

EXHIBIT5

KeyindustryusecasesleveraginggenAI

Source:EverestGroup(2023)ILLUSTRATIVE

Travelandtransport

IndustriesM&E

RCG

BFSI

Professional

services

Contractassistant

Report

summarization

Researchassistant

Customerexperience

Employeeexperience

Sales&

marketing

Human

resources

IT

Financeandaccounting

Code

generation

Text-to-SQL

Synthetic

datasetsfor

modeltraining

Testcasesgeneration

ITdocumentcreation

Website

development

|ThisdocumenthasbeenlicensedtoCapgemini

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

10

Potentialpitfalls–genAIisnotabedofroses

Untilnow,theapplicationofAIsystemswasnotreadilyapparenttoendusers.Whileitdidhavea

transformationalimpactonhowcompaniesoperate,mostofitwastransferredtoendcustomersintheformofbenefitsanduserexperience.ChatGPTisgenAI’siPhonemoment,whichturnedatechnologyintothezeitgeist.ItprovidesitsusersafascinatingexperienceofengagingwithAIsystemsfirst-hand,alongwiththeabilitytohavemeaningfulconversationslikeneverbefore.

However,thesebroadconversationalabilitiesdon’tmovethegenAIneedleforwardinwaysthatare

meaningfulforbroaderindustryadoption.TheunderlyingissueswiththetechnologyhinderitsadoptionamongenterpriseslookingoutforimpactfulgenAIusecases.

Datapreparedness-isyourdatagenAI-ready?

Theoutputofagenerativemodelisatruereflectionofthedatathatisfedintoitduringitstraining.Mostlargemodelsaretrainedonunfiltereddatafromtheinternet(socialmediafeeds,publications,e-journals,etc.)andare,therefore,subjecttoinherentbiasesanderrors.

It’samistaketoberelyingonChatGPTforanythingimportantrightnow.

–SamAltman,ChiefExecutiveOfficer,OpenAI

”

Thecurrentlargemodelsaretrainedtopresenttheiruserswithanoutputfortheirqueryorprompt–nomatterhowrightorwrong.However,incontrasttowhatafewmightclaim,thesemodelshavenot

achievedtheabilitytoreasonyet.Manyatimes,insituationswherealargemodelhaslimitedorno

actualinformation,itfillsupanygapsbasedoninformationthatismostlikelytobecorrect.Thisopacity

regardingthesourcedatacancatastrophicallyaffecttheoutputqualityandthedecisionstakenthereafter.

Safetyfirst-howdoyousafeguardyourenterprisedata?

Earlierthisyear,AmazonwarneditsemployeesabouttherisksofsharingconfidentialinformationthroughChatGPT.Soonafter,Samsung’semployeesaccidentallyleakedcompanysecretsvia

ChatGPTandmadeheadlinesforseveralpressreleasesinearlyApril.Consequently,thecompanybannedtheuseofgenAIinternally.JPMorganhasalsobannedtheuseofChatGPT.

Modelsthatarepre-trainedonexternaldatacanpresenttheriskofexposingsensitiveorconfidentialdatatothirdparties.WhileemployeesmayusegenAIasaproductivitytool,atthebackend,the

platformcontinuouslylearnsfromthedatathatissharedwithit.Thiscanhavedisastrous

consequencesforenterprises,whichstandatthevergeofleakingtheirprivatedatatotheoutsideworld,includingtheircompetitors.

|ThisdocumenthasbeenlicensedtoCapgemini

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

11

Actionsbasedonpre-trainedfoundationgenAImodelsalsolackclearresponsibilityandaccountabilityfortheoutputgenerated.AsenterprisesbegintointegrategenerativecapabilitiesofAIintotheircore

operations,theymustaddresstheelephantintheroom–whoisresponsibleforthequalityandlegalityoftheoutputgeneratedbythese“intelligent”systems?Ifthingsgowrong,isitthemodelownertobe

blamedortheuser?

Costconsiderations-isgenAIaffordable?

Despitetheirexceptionalperformance,thepre-trainedfoundationmodelstodayarenotenterprise

ready.Theydonotaccuratelycapturethelanguageusedwithinanenterprise'sspecificindustryor

domain,andthuscanleadtosuboptimalperformance.Notably,thecustomizationofthesemodelsonenterprisedataisthebiggestenterpriseconcerntoday.TotrulyuncoverthepowerofgenAI,

enterprisesneedtofine-tunethesemodelswithlocalenterpriseknowledgeusingtechniquessuchas

transferlearning.IntherealmofgenAI,opensourcemodelsareemergingasacost-lightalternativeto

proprietaryfoundationmodels.However,fine-tuninganyfoundationmodel,beitopensourceor

proprietary,isatime-consumingandresource-intensiveprocessthatrequiressignificantfinancial

investment.Hence,itisimportantforenterprisestocarefullyassesstheRoIbeforetheypushthegenAIboatout.

Theconundrumofsustainability-cangenAIloweryoursustainabilityscore?

Sustainabilityhasbeenoneveryone’smindslatelyand,consequently,manyenterpriseshavelaid

downambitioussustainabilitygoalstoachieveinthecomingyears.Meanwhile,theexcitementcreatedbygenAIisunimaginable.Thegenerativemodelsaregettingbigger,butsoaretheircarbonfootprints.Trainingthesemodelsrequiresamassiveamountofcomputingpower,whichboilsdowntoincreasedenergyconsumption,furtheraggravatingtheongoingclimatecrisis.Anothersustainabilityissuewith

thesemodelsistheirpotentialtoperpetuatebiasandinequity.Thesemodelslearnfromlargedatasets,and,ifthosedatasetsarebiased,themodelmayproducebiasedoutputs.GenAImodelsalsoposea

largersocietalriskoftakingovercertainjobs.

Atabroaderlevel,AIsystemscarryhugepotentialtoovercomesomeofthemostpressing

sustainabilityissues.So,thequestionis,howcanweusethesesystemstobuildabettersociety?Howcanenterprisesandproviderssolvethisparadoxofabitter-sweetrelationshipbetweenAIand

sustainability?WillthecurrentgreenAIsystemsbeenoughtomanagethescaleofthesegigantic

systems?Whilethetechnologyispromising,evaluatingtheRoIofgenAIimplementationsandcarefullyweighingthemagainstthecostofsustainabilitywillbecriticalforbusinesses.

TotrulyuncoverthepowerofgenAI,enterprisesneedtofine-tunethesemodels withlocalenterpriseknowledgeusingtechniquessuchastransferlearning.

|ThisdocumenthasbeenlicensedtoCapgemini

GENERATIVEAI:THENEXTCHAPTEROFARTIFICIALINTELLIGENCE

12

RequisitestobuildasturdygenAIstack

WhileweweighinonthechallengeshinderingscaledgenAIadoption,onethingisbecomingevidentlyclear:oncecontextualized,thetechnologyhastremendouspotentialforenterprises.Unlikeother

technologies,genAIisheretostayforalongtimebecausecurrentAIsystemshavealreadysettherightstagewitharobustfoundationalinfrastructureandcomplementarytechnologies.EverestGroupbelievesthatenterprisesshouldconsidersixfactorstomovebeyondexperimentationwithgenAItocommercialadvantage.

Theneedforcustomization-howdoyoumakegenAImodelstalkyour

organization'slanguage?

TotrulyunlockgenAI’spower,enterprisesneedtomakethesemodelscompatiblewiththeir

organizations’language,securityandprivacyrequirements,andexistingsystemsandinfrastructure.

Onethingthatisevidentlybecomingclearisthatnoonecanwinthisracealoneandthatknowledge

willbeakeydifferentiatorinthisecosystem.Notably,RedditandStackOverflowhavesetforthplanstostartchargingforaccesstotheirAPIsforthosecrawlingtheirwebsitestogathertrainingdata.

Partnershipshavealwaysbeencriticalforthedevelopmentofanybreakthroughtechnology,andthisneedisfurtherdrivenwithgenAIpushingenterprisestore-thinktheircollaborationstrategieswiththeextendedecosystem.

TheChineseAIgiantAlibabarecentlyannouncedanewpartnerprogramtofindpartnersthatcanhelpbuildcustomgenAImodels.Manyfoundationmodelprovidersareeitherlookingforpartnersorhave