Capgemini-明天的對話#9-生成型人工智能Gener(AI)ting the future 2024

RESEARCHINSTITUTE

Gener(AI)tingthefuture

conver

sa

tions

FORTOMORROW

QuarterlyreviewN°92024

ve

sa

tns

conver

sa

tions

FORTOMORROWQuarterlyreview

GenerativeAItodayisatthethresholdofanoutburstofcreativeexuberance.Thecover

forthiseditionofConversationsforTomorrowrepresentsthemyriadpossibilitiesthat

ariseattheintersectionoflightandshadow–nottoodissimilartothepossibilitiesthat

generativeAIcreates.Atthesametime,italsohighlightsfleetingpatternsanddark

spotswhichonemustbearinmindastheymoveforward.Thecontentanddesignof

thisissuereflecttheopportunities,thechallenges,andtherisksthatgenerativeAIis

nowthrowingupinfrontoforganizations.

2CapgeminiResearchInstituteGener(AI)tingthefuture

AtCapgemini,we

helporganizations

preparefortomorrow

Foreword

bydistillingthe

>

uniqueinsightsandperspectivesofleadersfromglobalbusiness,academia,thestartupcommunity,andwidersociety.

Gener(AI)tingtheFuture

InConversationsforTomorrow,the

CapgeminiResearchInstituteidentifies

thestrategicimperativesforthefutureofbusinessandthesocietyitserves.Inthisnintheditionofthejournal,amongotherareas,weexplore:

?TherapidriseofgenerativeAI(GenAI)

?Therateatwhichorganizationsacrossindustriesareadoptingthetechnology

?Theusecasesthatitenables

?Itsimpactonsustainabilityjourneys

?Howitislikelytochangeworkandtheworkforce

?Howandwhyitneedstoberegulated

OurannualresearchintothestateofGen

AIshowsthatorganizationsareembracinggenerativeAI,thisisreflectedbyanuptickininvestmentlevels.Thevastmajority(80%)oforganizationsinoursurveyincreasedtheir

investmentin2023;20%maintainedtheirinvestmentlevels;andnoorganizationhas

decreaseditsinvestmentinGenAIfromlastyear.

Inthepastyear,everyorganizational

domain,fromsalesandmarketingtoIT,

operations,R&D,finance,andlogistics,hasseenanincreaseintherateofadoption.

Earlyadoptersareseeingbenefitsfromimprovedoperationalefficiencyto

enhancedcustomerexperience.Moreover,generativeAIadoptionamongemployeesisrobustinmostorganizations,withthe

majorityallowingitsuse.

con

tions

ButadoptershavealsohadtoacknowledgeAI’ssignificantcarbonfootprint.Over

one-thirdoforganizationsinourresearcharealreadytrackingtheirGenAIcarbonemissions.

Organizationsgloballyarerapidly

embeddingGenAIacrossfunctions,witharippleeffectforwidersociety.Inthis

editionofConversationsforTomorrowwefocusonthisAI-generatedfuture.

Wewouldliketothankalltheleadersandexpertswhohaveenrichedthiseditionof

thejournalwiththeirinsights.Bysharing

theperspectivesofsuchadiverserangeofaccomplishedindividuals,weaimtopresentacomprehensiveoverviewofGenAIanditscontributiontogeneratinganewfuture.

?TheCEOofoneofthehottestAIstartups

?Oneofthemost-respectedAIscientistsglobally,whoisalsoaboardmemberatAmazon

?AformermemberoftheEuropean

ParliamentwhoplayedakeyroleintheEUAIAct

?AleadingprofessoratStanford

?SeniorexecutivesfromAdobe,

Salesforce,OECD,Telefónica,andItaúUnibanco

?Capgemini’sownsubject-matterexperts

Pullingtogethersuchawiderangeofviewswasanextremelyinstructiveexercisefor

us.Wehopeyouenjoyreadingthiseditionasmuchasweenjoyedputtingittogetherforyou.

Gener(AI)tingthefutureCapgeminiResearchInstitute

3

ve

sa

tns

Contents

4CapgeminiResearchInstituteGener(AI)tingthefuture

P.14

THECEOCORNER

P.16

ArthurMensch

CEO

MistralAI

ve

sa

tns

P.16

AimanEzzat

CEO

Capgemini

5

Gener(AI)tingthefutureCapgeminiResearchInstitute

ve

sa

tns

P.26

EXECUTIVECONVERSATIONSWITH…

P.64

ClaraShih

CEO,SalesforceAI

P.28

ErikBrynjolfsson

Professor,Stanford

P.76

AndrewNg

CEO,LandingAI

P.36

AudreyPlonk

DeputyDirector,DirectorateforScience,TechnologyandInnovation(STI),OECD

P.86

ChemaAlonso

ChiefDigitalOfficer,Telefónica

P.44

Drago?Tudorache

FormerMEP,EUAIAct

co-rapporteur,EuropeanParliament

P.94

RicardoGuerra

ChiefInformationOfficer,ItaúUnibanco

P.54

ScottBelsky

ChiefStrategyOfficer,Adobe

Contents

6CapgeminiResearchInstituteGener(AI)tingthefuture

ve

sa

tns

P.104

PERSPECTIVESFROMCAPGEMINI

GENERATIVEAIFORMANAGEMENT

P.106

ElisaFarri

and

GabrieleRosani

fromTheManagementLabbyCapgeminiInvent

GENERATIVEAI:THEARTOFTHEPOSSIBLE

P.116

RobertEngels

HeadofGenerativeAILab,Capgemini

OPERATIONALAIIS

CHANGINGHOWWELOOKATDATA

P.124

AnneLaureThibaud

ExecutiveVicePresident,Data,AI&AnalyticsGroupOffer

Leader,Capgemini

and

SteveJones

ExecutiveVicePresident,DataDrivenBusiness&GenAI,

Capgemini

P.130

INSIGHTS

FROMTHECAPGEMINIRESEARCHINSTITUTE

P.131

GenAIinenterprise

TheriseofgenerativeAIinvestmentsinorganizations

P.138

GenAIinsoftwareengineering

HowsoftwareengineeringisbeingshapedbygenerativeAI

7

Gener(AI)tingthefutureCapgeminiResearchInstitute

ve

sa

tns

ExecutiveSummary

ExecutiveSummary

GenerativeAI(GenAI)ismakingrapid

inroadsintoorganizationalstructures,

transformingthemrapidlyfromtheinside

out.Asbusinessesacrossindustriesbegin

toimplementGenAI,severalkeythemes

emergethathighlightitspotentialtoimpactorganizations,workforce,andsociety.

GenAIhasthepotentialtoreimaginethefutureoftheworkforce

GenAIhaspotentialtounlockhuman

creativity,allowingemployeestofocusoncomplexstrategictasks.ClaraShih,CEO,

SalesforceAI,comments:“AIwillallow

workerstomoveawayfromrepetitivetaskstofocusondoingwhathumansdobest,

whichisbuildingrelationships,unlockingcreativity,makingconnections,and

addressinghigher-orderproblems.”

ScottBelsky,ChiefStrategyOfficerand

ExecutiveVicePresident,Designand

EmergingProducts,Adobe,adds:“Wearemostcreativelyconfidentwhenwearefiveyearsold.Weloseourcreativeconfidenceaswegetolderbecauseoftheskillsgap,exposuretocriticism,andjustthelackofaccesstocreativetools.GenerativeAIis

fundamentallychangingthis.”

AndrewNg,CEO,LandingAI,emphasizestheboosttoproductivity:“Formanyjobs,AIwillonlyautomateoraugment20-30%oftasks.So,there'sahugeproductivityboost,but

peoplearestillrequiredfortheremaining70%oftherole.”

Wearemostcreativelyconfidentwhenwe

arefiveyearsold.

Weloseourcreativeconfidenceaswe

getolderbecause

oftheskillsgap,

exposuretocriticism,andjustthelackof

accesstocreativetools.GenerativeAIisfundamentallychangingthis.”

ScottBelsky

ChiefStrategyOfficer,Adobe

8CapgeminiResearchInstituteGener(AI)tingthefuture

ExecutiveSummary

ve

sa

tns

Whenevera

traditionalactivitygetsreplacedor

augmentedwith

onebasedonbits,itusuallybrings

significantenergyandenvironmentalbenefits.”

ErikBrynjolfsson

Stanford

Thisshiftisfosteringacultureofcontinuouslearningandadaptability.ErikBrynjolfsson,ProfessorattheStanfordInstitutefor

Human-CenteredAIandDirectorofthe

StanfordDigitalEconomyLab,discussestheimportanceofworkforceskillenhancement:

“AIrequiressignificantchangesinthe

economytocreatefullimpact,particularlyintermsoforganizationandskillsofthe

workforce.Identifyingwhichskillsare

important,followedbyself-learningand

trainingprograms,arerequiredtopreparetheworkforce.Secondly,businesseswill

needtorestructureandadapttocapitalizeonnewtechnologies.”

GenAIalsohasasignificantimpactonmanagersandleadership.

Inthearticle,GenerativeAIfor

Management,ElisaFarri,Vice-PresidentatCapgeminiInvent,andGabrieleRosani,DirectoratCapgeminiInvent,comment:

“AI'scapabilitytocollaborateonacognitiveandemotionallevel,offeringinsightsandcontributingtocomplexdecision-making

processes,isanareathatmanymanagershaveyettofullyrealizeorintegrateintotheirstrategicthinking.”

Theyadd:“Executivesneedtocultivatetheabilitytoadopttheco-thinkingmindset.

Whetherinindividualtasksorteam

endeavors,masteringthisshiftwillbecomeavitalcompetitiveadvantage.”

Gener(AI)tingthefutureCapgeminiResearchInstitute

9

ve

sa

tns

ExecutiveSummary

InGenAIdeployment,ethicalconsiderationsareparamount

GenAIhasemergedasatransformativeinnovation.However,itspotential

formisuseemphasizestheneedfor

organizationstoupholdstrongethicalstandards.

AimanEzzat,CEO,Capgemini,stressestheimportanceofsafeuseofGen

AI:“Organizationsshouldestablish

employeeguidelinesforsafeuseofGenAIandvalidatingoutputstoeliminate

bias.”

ArthurMensch,CEO,MistralAI,

underscorestheimportanceofmanaging

AI-drivenproductswithinethical

boundaries:“WhenanorganizationismakinganAI-drivenproduct,itmust

considerthedecisionsandoutputsthesystemwillmake.Thesedecisionsandtheseoutputsshouldbeconstrainedtorespectthecompany'srole.”

AudreyPlonk,DeputyDirectoroftheOECDDirectorateforScience,Technology,andInnovation,is

responsiblefortheOECD’sdigital

policyportfolio.Sheelaboratesondata

concerns:“Dataprivacyconsiderations

areakeyaspectthatorganizationsand

individualsareexploringextensively.

Thereisalotofworktobedoneto

improvetransparencyanddetermine

whichsourcesofdatashouldbeusedtotrainAImodels.Itisessentialtoputtheappropriatesafeguardsinplace,includingdataprotectionconsiderations.”

SharingCapgeminiperspectives,AnneLaureThibaud,ExecutiveVicePresident,Data

AIandAnalyticsGroupOfferLeader,and

SteveJones,ExecutiveVicePresident,Data-drivenBusinessandGenAI,remark:“The

assumptionisthatGenAIcannotbetrustedinthesamewayasahumanemployeeandgiventheopportunity,willactoutsideits

boundaries.Organizationsarecompelledtobuildalltheinformationabouttheirculture,mission,andguardrailsintotheAItheyusetoretaincontrolofit.”

10CapgeminiResearchInstituteGener(AI)tingthefuture

ThepathtoresponsibleAI

TosecurethefutureofAI,comprehensivesafeguardsandcollaborativeeffortsare

essential.UnifiedactionsfrompolicymakersandorganizationsarecrucialtoensuretheresponsibleuseofGenAI.

DragosTudorache,formermemberofthe

EuropeanParliamentandRapporteurontheEUAIact,says:“Mostcompaniesworking

withAIalreadyhadgeneralprinciplesor

codesofconductorself-regulationinplace.TherewereguidelinesoutlinedbyUNESCO,OECD,andevenbytheEuropeanParliament.Butwerealizedthesemeasureswere

insufficienttomitigatetheveryrealrisks,suchasdiscriminationbias.”

Organizationsshouldestablishemployeeguidelinesforsafe

useofGenAIand

validatingoutputstoeliminatebias.”

AimanEzzat

CEO,Capgemini

OntheneedtodrafttheEUAIAct,headds:

“Weneededtoputstrongersafeguardsin

placethatcommandrespectand,ultimately,helpsocietytotrustintheinteractionwiththistechnology,hencethedecisionto

formulatethepolicy.”

ExecutiveSummary

RicardoGuerra,CIO,ItaúUnibanco,highlightstheresponsibilityof

organizationstouseAIethically:

“Organizationshavetotakealotofthe

responsibilityforuseandgovernanceofAI

andothertechnologies.Butgovernmentsmuststillstayinformedandtryto

implementsupportiveregulation.”

Steeringtowardsa

sustainablefuturewithGenAI

GenAIwillplayapivotalroleinaddressingclimatechange.Additionally,businesses

mustadoptsustainablepracticesthatalignwithenvironmentalgoals.

Onclimateengineering,AndrewNg

elaborates:“Giventheworld'scollective

inabilitytomoveCO2emissionsinthewayweknowitneedsto,Ithinkitispasttimetotakeclimateengineeringmoreseriously.IthinkAI,specificallylargeAIfoundation

modelsofclimate,havealargeroletoplayinthat.”

RicardoGuerratalksaboutsustainabledatacenters:“We'relearningwhento

usedifferentsolutionsandemphasize

investinginsustainabledatacentersandgreentechnologies.We'realsoclosely

monitoringthemarket,prioritizing

providersthatoffergreensolutions.”

ErikBrynjolfssonsays:“Wheneveratraditionalactivitygetsreplacedoraugmentedwithonebasedonbits,itusuallybringssignificantenergyandenvironmentalbenefits.”

ve

sa

tns

Gener(AI)tingthefutureCapgeminiResearchInstitute

11

ve

sa

tns

ExecutiveSummary

Theriseofopen-sourceandsmalllanguage

models

Smalllanguagemodels(SLMs)arehighlycost-effective,resource-efficient,and

haveminimalenvironmentalimpact.

Open-sourcemodelspromoteinnovation,enhancecollaboration,andensuregreatertransparencyindevelopmentandusage.

AimanEzzatsays:“Thereareclear

advantagestobothopenandclosed

approaches.Opennessboostsinnovationanddrivescollaboration.Openmodelsalso

alloweveryonetoscrutinizethemodelforpotentialsourcesofbias,demystifyingthe‘blackbox’natureofAImodels.”

ArthurMenschcomments:“Smallermodelsalsomeantheapplicationsarelesscostlytorunand,moreimportantly,ifyouhave

amodelthatis100timessmaller,youcancallit100timesmoreforthesamecost,bringingalittlemoreintelligencetoyourapplicationwitheachcall.”

ClaraShihadds:“ThefutureofAIwillbea

combinationofbothlargeandsmallmodelsbecauseofclimateimpact,aswellasforcostandperformancereasons.”

WhenanorganizationismakinganAI-drivenproduct,itmust

considerthedecisionsandoutputsthesystemwillmake.These

decisionsandthese

outputsshouldbe

constrainedtorespectthecompany'srole.”

ArthurMensch

CEO,MistralAI

12CapgeminiResearchInstituteGener(AI)tingthefuture

ExecutiveSummary

ve

sa

tns

ThefutureofAIwill

beacombinationof

bothlargeandsmallmodelsbecauseof

climateimpact,as

wellasforcostand

performancereasons."

ClaraShih

CEO,SalesforceAI

GettingGenAIright

requiressound

technicalstrategyandacultureofinnovation

EffectiveGenAIadoptionrequiresawell-coordinatedstrategicapproachwitha

strongtechnicalfoundation,supportfromleadership,andanorganizationalcultureofinnovation.

ChemaAlonso,ChiefDigitalOfficer,

Telefónica,suggests:“Youneedtohavea

robusttechnicalstrategybasedoncloud

andsounddata,andtherestwillfallinto

place.Secondly,youneedtohavestrong

supportfromtopmanagement.Finally,

youneedsufficientbudget.Onceyouhavethat,youneedtomakesurethatyourwholeorganizationisverywelltrainedonGenAI–whatcanandcannotbedone.”

RicardoGuerrasays:“AdoptingGenAI

requiresacultureofinnovation.WithGenAI,weneedtoengagethebusinessand

designteamsactively,astheymustidentifyopportunitiesbeyondmeretechadoption.”

InnovativestartupspointtothefutureofGenAI

Fromcontentcreationtomulti-agent

systems,alternativecomputing,andhybridAI,emergingtechstartupsarepushingtheboundariesofthenextgenerationofAI

applications.

SynthesiausesAItocreatecustomizablevideocontentfeaturingrealisticavatars,

allowingbusinessestogenerateengagingvideopresentations.Soundrawoffers

anAI-poweredplatformforgenerating

originalmusicwithouttheriskofcopyrightinfringement.

Intherealmofalternativecomputing,

Mythicdevelopsanalogchipsforfaster,

moreefficientAItaskssuchasmatrix

multiplications,whileGroqcreatesAI-

optimizedlanguageprocessingunits(LPUs)designedforrunningLLMs.

LiquidAIispioneeringthedevelopmentofhighlyefficient,task-specificmodelsusingliquidneuralnetworks,withapplications

suchasdronenavigation,showcasingGenAI’swiderangeofpossibilities.

Gener(AI)tingthefutureCapgeminiResearchInstitute

13

Cnve

sa

tns

TheCEOCorner

ArthurMensch

CEO

MistralAI

TheCEOCorner

indiscussionwith

AimanEzzat

CEO

Capgemini

14CapgeminiResearchInstituteGener(AI)tingthefuture

ve

sa

tns

Gener(AI)tingthefutureCapgeminiResearchInstitute15

ve

sa

tns

TheCEOCorner

ArthurMensch

CEO,MistralAI

ArthurMenschisaFrenchentrepreneurandscientist.

In2023,ArthurMensch,alongwithGuillaume

LampleandTimothéeLacroix,foundedMistralAIwiththemissionofmakinggenerativeAI

ubiquitousandpioneeringanewapproachtoAI-onethatismoreopen,portable,independent,andaccessibletoall.

Aftermorethan10yearsofacademicwork

focusedonthepossibilitiesofmachinelearninginthefieldofbrainimagingandonoptimizationofmachinelearning,hejoinedDeepMindParisin2020asaresearcher,wherehespentthreeyearsandplayedakeyroleinthedevelopmentand

deploymentofflagshipprojectsingenerativeAI.

AimanEzzat

CEO,Capgemini

Withmorethan20years’experience

atCapgemini,AimanEzzathasadeep

knowledgeoftheGroup’smainbusinesses.Hehasworkedinmanycountries,notablytheUKandtheUS,wherehelivedformorethan15years.

AimanwasappointedCEOinMay2020.

Priortothat,from2018to2020,heservedastheGroup’sCOOand,from2012to

2018,asCFO.AimanisalsoontheBoard

ofDirectorsofAirLiquideandisamemberoftheBusinessCouncilandtheEuropeanRoundTable(ERT)forIndustry.

16

CapgeminiResearchInstituteGener(AI)tingthefuture

TheCEOCorner

ve

sa

tns

Whatinspiredyoutoformanewplayer[inMistral]inthegenerativeAI(GenAI)space–andwhyinEurope?

—Arthur:Myco-foundersandIhavebeenworkinginthe

GenAIspaceforover10years,previouslyinlargeUS-basedorganizations.Whendevelopmentacceleratedatend-2022,itgaveusanopportunitytocreatesomeverystrongmodelsinashortperiodoftime.Wesecuredfunding,assembled

adedicatedteamandtheGPUs[graphicprocessingunits]requiredtotraintheLLMs[largelanguagemodels],andwerereadytogo.

WhyEurope?Europeisagreatplacetostartacompany.TheeducationsystemsinFrance,Poland,ortheUK,forexample,aregreatfortrainingAIscientists.Webroughtinrecent

PhDsfromParis;wewereabletogetthemostimportant

thingtogetstarted–theteam.AstheonlyplayerinEuropeinthefieldofconventionallanguagemodels,wehadsomestronggeographicalbusinessopportunities.

Weusebothanopen-source

modelandaportableplatformformodeldeployment."

ArthurMensch

Gener(AI)tingthefuture

CapgeminiResearchInstitute

17

ve

sa

tns

TheCEOCorner

Whatdoyouseeastheadvantagesoftheopen-sourcegenAImodel?

—Arthur:Weusebothanopen-sourcemodelandaportableplatformformodeldeployment.Evenourcommercialmodelsarelicensed.Thisallowsuserstocustomizethemodelstotheirneeds.

Itoffersportabilityandcomfort.Withamodelthatyoucandeployonanyplatform,onaprivatecloudoron-premiseorondedicatedinstanceson

thecloud,youcanusethetechnologywhereyourdatais.So,thisadaptstothedata-governanceconstraintsoftheenterprises,andourcustomersverymuchappreciatethisflexibility.

—Aiman:Thereareclearadvantagestobothopenandclosedapproaches.Opennessboostsinnovationanddrivescollaboration.Openmodelsalso

alloweveryonetoscrutinizethemodelforpotentialsourcesofbias,

demystifyingthe“blackbox”natureofAImodels.Therearealsochallenges.Customizingopenmodelsforaparticularindustryororganizationistricky,butusingopenmodelsoutoftheboxcanleadtosuboptimalperformance.Fine-tuninganyfoundationmodel,open-sourceorproprietary,isatime-

consuming,resource-intensiveprocessthatrequiressignificantfinancialinvestment.Hence,itisimportantforenterprisestoassessROIcarefullybeforepushingouttheGenAIboat.

"Openmodelsalsoallow

everyonetoscrutinize

themodelforpotential

sourcesofbias."

AimanEzzat

18

CapgeminiResearchInstituteGener(AI)tingthefuture

TheCEOCorner

ve

sa

tns

DoyouseeorganizationsusingageneralizedGenAImodelgoingforwardormanydifferentspecializedmodels?

—Arthur:Weseethefieldmovinginthesetwodirectionssimultaneously.Astronggeneralizedmodelgivesagoodplatformfortestingsolutions.Butthiscanbea

slowandcostlyprocess,offeringpoorROIforspecifictasks.YouwantyourLLM

toofferanintelligent,dynamicsolutionforaspecificissue,whetherthat’sparsingthelogsofanITsystemorparsingtheconversationbetweenacustomeranda

customeragent.

Fromascientificpointofview,smallermodelscansolvespecificissues,buttheymustbe

finelytuned.Wewanttobringsolutionstomarketthatdevelopthesmallestpossible

modeltosolveaspecificdefinedtask,whichwillallowforlow-latencyapplications.

Smallermodelsalsomeantheapplicationsarelesscostlytorunand,moreimportantly,ifyouhaveamodelthatis100timessmaller,youcancallit100timesmoreforthesamecost,bringingalittlemoreintelligenceto

yourapplicationwitheachcall.Wecallthis“compressedknowledge”.Wespecialize

modelsinordertomakedifferentiated

applicationsthatgofast,thatcallLLMoftenandthatarecost-controlled.

Ifyouhaveamodel

thatis100times

smaller,youcancallit100timesmoreforthesamecost,bringingalittlemoreintelligencetoyourapplication

witheachcall."

ArthurMensch

—Aiman:There’saveryclearmarketforbothgeneralizedandspecializedmodels.Ageneralizedmodelcanservethoseusecasesthatdon’trequireextensive

customization.Theseare“l(fā)ow-hangingfruit”thatrapidlydemonstratethepowerofGenAI.

Developingandtrainingspecializedmodelsforsomebasicusecasesmightevenbecounterproductiveintermsofcostandsustainability.Thatsaid,thereareusecasesthatbenefitfromspecializedmodels,forinstance,intermsofperformancecharacteristicsorindetectingandrespondingtospecificnuancesoftheindustryorusecase.Anyusecasethatrequireshighperformanceordeepdomain

expertisewilllikelycontinuetogodownthepathofspecializedmodels.Atthe

sametime,specializedmodelspotentiallyrequiresignificantresourcesintermsofmaintenanceandregularupdates,soorganizationsmightpreferageneralizedmodelforusecaseswithlessstringentrequirements.Iseeafuturewherebothtypesofmodelscoexistharmoniously.

Gener(AI)tingthefuture

CapgeminiResearchInstitute

19

ve

sa

tns

TheCEOCorner

Anyusecasethat

requireshigh

performanceordeepdomainexpertise

willlikelycontinuetogodownthepathofspecializedmodels."

AimanEzzat

WhatarethemostinnovativeusecasesthatyouareseeinginGenAI?

—Arthur:In?nancialseΓvices,foΓinstance,MistΓalhasbuiltmodelsthat

extΓactinfoΓmationfΓom?nancialΓepoΓtsandsummaΓizeitfoΓbankeΓsto

analyze.ThisharnessesthepowerofgenerativeAItoprocessalargeamount

oftextanddetectweaksi9nals,whichisveΓymuchthecoΓebusinessofbanks.TheotheΓsuccessfuldeploymentisincustomeΓseΓvices.

—Aiman:WehavebeenwoΓkin9onseveΓalinnovativecasesusin9GenAIacrossindustries.Inlifesciencesforinstance,wehavedevelopedwithgenerativeAIasolutiontodesignnewdrugmolecules.Thismethod

si9ni?cantlybooststhepΓocessof9eneΓatin9newstΓuctuΓes,o氏eΓin9ΓeseaΓcheΓsapotenttoolfoΓdesi9nin9moleculesaimedatspeci?c

biologicaltargets.ItillustratesAI'stransformativepotentialinacceleratingandΓe?nin9dΓu9discoveΓy,paΓticulaΓlyinthepΓeliminaΓyphases.

20

CapgeminiResearchInstituteGener(AI)tingthefuture

TheCEOCorner

ve

sa

tns

Giventheenergyrequiredtocreateandtrainthelargemodels,whatarethesustainabilityimplicationsforGenAI?

—Arthur:Mostofthecomputeandenergyresourcesrequiredtorunthesesystemsareusedatinferencetimeratherthanattrainingtime.Soyoutrainforacoupleofmonths,andwhenthemodelsaredeployedonmany,manyGPUs,thenthelargeenergyconsumptionismorelinkedtotheusagethantothetrainingitself.Therearetrade-offsbetweentheamountyouspend

ontrainingandthecompressionthatyoucanachieve.Ifyouinvestmoreintraining,youcanmakesmallermodels,achievingthesameperformanceasalargermodelwithlesscompute.Thesesmallermodelsconsumelessenergy

todeployatinferencetime.

AtMistral,wefocuson

compressingknowledgeand

makingmodelsthataresmallerthanthosethecompetition

produces.Limitingcarbon

emissionsisacausethatis

verydeartoourheartandthereasonwhywedeployedour

solutionsinEurope.InSweden,

inparticular,renewables

composeahighproportionofenergyconsumption.

Mostofthecomputeand

energyresourcesrequiredtorunthesesystemsareused

atinferencetime,ratherthanattrainingtime."

ArthurMensch

—Aiman:Ourresearchshowsthatmorethanthree-quartersoforganizationsare

consciousofenvironmentalconcernsaroundGenAI.Asaleaderintheeco-digital

revolution,weatCapgeminirecognizetheneedtoweightheimmensepotentialofGenAIagainstitscosttotheplanetandsociety.Wearecommittedtotakinga“sustainablebydesign”approachtodevelopingGenAIsolutionsthatharnesscutting-edgedata,AI,andclimatetechtomaximizebusinessoutcomesinasustainablemanner.

Mitigationstrategiesincludeoptimizingtheamountofdatarequiredtotrainthe

models,workingonsmaller,task-specificenergy-efficientmodelsthatemploymore

efficienttrainingandoperatingalgorithms,andpoweringtheAIinfrastructurewith

renewableenergyaswellasusingmoreenergy-efficientdatacenters.WealsopromotetransparencyinAIdevelopmentandoperationbymonitoringanddisclosingtheenergyconsumptionandcarbonfootprintsofGenAImodels.OurGenAIlifecycleanalysistoolshelporganizationstomitigateenvironmentalimpact.

Gener(AI)tingthefuture

CapgeminiResearchInstitute

21

ve

sa

tns

TheCEOCorner

HowshouldlargeorganizationsaddressethicalconsiderationsandpotentialbiasindeploymentofAImodels?

—Arthur:Whenanorganizationis

makinganAI-drivenproduct,ithasto

"Therare

talentthatwe

recommend

every

organization

lookforis

thesoftware

engineerwho

canalsododata

science."

ArthurMensch

considerthedecisionsandoutputsthesystemwillmake.So,thesedecisions

andtheseoutputs