《花旗银行:2023释放AI的潜能研究报告(英文版)(27页).pdf》由会员分享,可在线阅读,更多相关《花旗银行:2023释放AI的潜能研究报告(英文版)(27页).pdf(57页珍藏版)》请在三个皮匠报告上搜索。
1、As our premier thought leadership product,Citi GPS is designed to help readers navigate the most demanding challenges and greatest opportunities of the 21st century.We access the best elements of our global conversations with senior Citi professionals,academics,and corporate leaders to anticipate th
2、emes and trends in todays fast-changing and interconnected world.This is not a research report and does not constitute advice on investments or a solicitations to buy or sell any financial instruments.For more information on Citi GPS,please visit our website at AI The AI Arms Race Generative AI repr
3、esents the latest inflection point in the evolution of AI.Although AI is not new,what is distinctive about Generative AI is the tremendous potential it holds to transform work across industries and boost overall productivity.We look at the opportunities for AI not only in the technology sector but a
4、cross the supersectors.We also look at the race to dominate in the AI space by analyzing research papers and patents.Citi GPS:Global Perspectives&SolutionsCiti GPS:Global Perspectives&Solutions September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 Citi Research Analysts Andre
5、w Baum,MD Global Head of Healthcare Research Fatima Boolani U.S.Software Laura(Chia Yi)Chen Asia Semiconductors&Components Christopher Danely U.S.Semiconductors Patrick Donnelly Life Science Tools&Diagnostics Steven Enders,CFA U.S.Back Office Software Tiffany Feng HK/China Consumer Surendra Goyal,CF
6、A India IT Services,Media&Education Analyst Aaron Guy EMEA REITs Simon Hales EMEA Beverages Nick Joseph Head of U.S.REIT and Lodging Team Ronald Josey U.S.Internet Andrew Kaplowitz U.S.Multi-Industry,Conglomerate,and E&C Peter Lee Asia Semiconductor&IT Hardware Paul Lejuez,CFA,CPA U.S.Retail&Food Re
7、tail Andre Lin,CFA Asia Hardware Carrie Liu Taiwan Technology/Hardware Craig Mailman U.S.Retail&Industrial REITs Atif Malik Semiconductor Capital Equipment&Specialty Semiconductor Asiya Merchant,CFA U.S.Hardware Itay Michaeli U.S.Auto&Auto Parts Takayuki Naito Japan Electronic Components Arthur Pine
8、da Head of Pan-Asian Telecom Research Jenny Ping EMEA Utilities&Renewables Tyler Radke U.S.Software Michael Rollins,CFA U.S.Communications Services&Infrastructure Masahiro Shibano Japan Precision&Semi Prod Equipment Ashwin Shirvaikar,CFA U.S.Payments,Processors&IT Services Thomas A Singlehurst,CFA H
9、ead of European Media Research Alastair R Syme Global Head of Energy Research Martin Wilkie Head of European Capital Goods Research Alicia Yap,CFA Head of Pan-Asia Internet Research Oscar Yee Head of Pan-Asia Materials Research Steven Zaccone,CFA U.S.Retailing/Hardlines Judy Zhang Co-Head of Pan-Asi
10、a Banks Research Expert Commentary Pantelis Koutroumpis Director,Oxford Martin Programme on Technological and Economic Change Oxford Martin School Ian Goldin Professor of Globalization and Development Oxford Martin School Helen H Krause,CFA Head of Data Science Insights Citi Global Data Insights Yeh
11、uda Dayan Data Scientist Citi Global Data Insights Brian Yeung Data Scientist Citi Global Data Insights Amit B Harchandani European Technology Sector Expert Rob Garlick Head of Innovation&Technology Citi Global Insights 2023 CitigroupTHE RISE OF AISource:Citi GPS,SimilarWeb,Open AISource:IBMSource:C
12、iti GPS,IDCOPPORTUNITIES FOR ENABLERSCHATGPTS RAPID GROWTH AT LAUNCH(Time to reach 100 million users worldwide)AI IS BEING USED ACROSS THE WORLDGENERATIVE AI:THE LATEST INFLECTION POINT IN AIKEY ADOPTION CHALLENGESVertical AppsHorizontal AppsSoftware&ApplicationsGenerative AI Technology Value StackS
13、emiconductorsCapital EquipmentConsumablesSiliconHyperscalersDigital InfraTelcosInfrastructure&PlatformsFoundationVerticalizedOn-DeviceVertical DBModels&MLOpsAutomationAugmentationTransformationServicesVerticallyintegratedApps&ModelsNote:The Tech&Comms supersector is shown separately as it is the ena
14、bling supersector.GENERATIVE AI:ORDER OF IMPACT ACROSS SUPERSECTORS MostImpactedLeastImpactedFinancials&FinTechConsumersHealthcareIndustrial Tech&MobilityReal EstateNatural Resources&Climate TechTech&CommsFinancials&FinTechConsumerHealthcareIndustrial Tech&MobilityReal EstateNatural Resources&Climat
15、e TechTech&CommsDeployed AIExploring AIAustraliaCanadaChinaFranceGermanyIndiaItalyLatin AmericaSingaporeSouth KoreaSpainUAEUKUSGlobal80%60%40%20%0%World Wide WebiTunesTwitterFacebookWhatsAppInstagramApple App StoreTikTokChatGPT1 Yr2 Yrs3 Yrs4 Yrs6 Yrs5 Yrs7 Yrs7 Years6.5 Years5 Years4.5 Years3.5 Yea
16、rs2.5 Years2 Years9 Months2 MonthsAIAutomatic Speech Recognition,Natural Language Processing Artificial IntelligenceMachine LearningDeep LearningGenerative AIMLDLASR/NLPGenAIBIASAUTHENTICITYINEQUALITYINFRINGEMENTSEXISTENTIAL CONCERNS Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 202
17、3 Citigroup 4 Contents Executive Summary 5 AI and Our New Renaissance 8 The Rise of Artificial Intelligence 9 The Landscape of Generative AI.13 The AI Productivity Boost.14 The Next AI Era:Promises and Challenges 18 Global Regulatory Landscape Mixed.23 Sector Opportunities from Generative AI 25 The
18、Generative AI Technology Stack.25 Silicon.26 Infrastructure&Platforms.28 Models and Machine Learning Operations(MLOps).29 Software and Applications.30 Services.31 Generative AI:Assessing the Risk/Reward Outside Technology.33 Two-Stage Framework to Assess Risk/Reward.33 Generative AI:Impact Across Su
19、persectors.36 Financials&FinTech.37 Consumer.38 Healthcare.40 Industrial Tech&Mobility.40 Real Estate.42 Natural Resources&ClimateTech.43 The AI Arms Race 44 Research Outputs for AI.44 AI Research Collaborations:Openness is a Strength.48 Patent Trends in AI.50 AI Technological Innovation:Continued G
20、rowth with Large Quality Impact on Global View.50 AI Technological Innovation:Public Companies Drive U.S.Growth,Universities Drive Chinas.52 AI Technological Innovation:Sub Themes.54 Glossary of Key Terms.55 September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 5 Executive Su
21、mmary Generative AI has burst onto the scene following the launch of ChatGPT in 2022.As a concept,artificial intelligence(AI)is not new,and Generative AI represents the latest inflection point in the evolution of AI.However,what is distinctive about Generative AI is the tremendous potential it holds
22、 to transform work across industries and boost overall productivity.Taking a more holistic view,Generative AI might not only bring the power of AI itself to the masses but in fact accelerate the wider democratization of innovation.We believe it is a game changer.Figure 1.When Launched,ChatGPT Was th
23、e Fastest-Growing Consumer Application in History(Time to Reach 100 Million Users)Figure 2.Generative AI Potentially Represents the Latest Inflection Point in the Evolution of AI Source:Citi Research,SimilarWeb,OpenAI Source:Citi Research,IDC Figure 3.Generative AI Holds Tremendous Potential to Tran
24、sform Work(Share of Tasks That Could be Automated/Augmented by Its Adoption)Figure 4.Consumers Have High Trust Levels for Generative AI-based Interactions(Share of Consumer Who Trust Generative AI Content)Source:Citi Research,Accenture Source:Citi Research,Capgemini Although it feels like Generative
25、 AI has come out of nowhere,the report the looks at the rise of AI starting in the 1950s and through its significant growth over the past decade.We then explore the potential opportunities and challenges brought by Generative AI.Key challenges include those centered around bias,inequality,authentici
26、ty,infringements as well as the more debated one underpinned by existential concerns.The emergence of Generative AI has unsurprisingly seen AI as a broader topic become a firm focus for policy makers around the world.However,the regulatory path taken so far has varied.Given the stakes involved,we be
27、lieve policy and governance evolution will play a defining role.2 months9 months2 years2.5 years3.5 years4.5 years5 years6.5 years7 years02468ChatGPTTikTokApple App StoreInstagramWhatsAppFacebookTwitteriTunesWorld Wide WebAIMLDLASR/NLPGen AIAutomatic Speech Recognition,Natural Language ProcessingArt
28、ificial IntelligenceMachine LearningDeep LearningGenerative AI31%9%22%38%Higher potential forautomationHigher potential foraugmentationLower potential foraugmentation orautomationNon-language tasks73%79%75%74%74%73%72%72%72%72%72%72%70%0%20%40%60%80%100%AverageNorwaySpainUKFranceNetherlandsCanadaIta
29、lyJapanSwedenU.S.AustraliaSingaporeThe big picture History,opportunities,challenges,and regulation Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 6 The first wave of potential opportunities for Generative AI is centered on the technology value stack,as outlined in Figu
30、re 5.Historically,the Silicon layer has been the de facto foundation of almost all technological shifts in the technology value stack,and Generative AI is expected to drive significant growth for compute(i.e.,processing power),networking,and memory chips.However,as we look at the whole technology va
31、lue stack,we see opportunities in each layer.In the Infrastructure&Platforms layer,we see the hyperscalers/cloud providers racing today to build the underlying infrastructure that enables Generative AI applications and services,but over time we expect to see higher or more differentiation.When it co
32、mes to Models and Machine Learning Operations(MLOps),the open-source community is likely to be a key driver of innovation.Moving further up the stack,we believe nearly all software companies will be impacted in some form by Generative AI,and company-specific execution will be critical.Lastly,we beli
33、eve Generative AI represents a step forward from ongoing AI/automation initiatives at the Services layer.Figure 5.Generative AI Technology Value Stack Source:Citi Research Opportunities are not just limited to the technology value stack they are also spilling into sectors outside of technology.To ca
34、pture the full picture,we extended our analysis to look at the impact of Generative across six supersectors(Figure 6).We do this through a two-stage framework to assess risk/reward and apply that broadly across companies and sectors.Our analysis finds the Financials&FinTech supersector to be the mos
35、t likely to be impacted overall,followed by the Consumer sector.At the other end of the spectrum,Natural Resources&ClimateTech at this stage look the least likely ones to be impacted.Opportunities for enablers Opportunities extend across sectors September 2023 Citi GPS:Citi GPS:Global Perspectives&S
36、olutions 2023 Citigroup 7 Figure 6.Generative AI:Order of Potential Impact Across Supersectors Note:We show the Tech&Comms supersector separately as it is the enabling supersector Source:Citi Research What does the future look like for Generative AI?One way to investigate the global trends and growt
37、h is from the perspective of investment in technological innovation.We do this by analyzing the number of AI-related patent applications over time and across countries.Research papers are also telling,with the total cumulative AI research output increasing 1,300%between 2003 and 2021.Given the impor
38、tance of AI as a foundational technology,the race is on between countries for scientific and technological dominance.Patents and papers indicating we are in an“AI Arms Race”Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 8 AI and Our New Renaissance The growing potentia
39、l for AI to become a widely applicable form of general-purpose technology could create an era defining period of disruption with wide ranging social,political and economic implications.Generative AI could be as revolutionary a technology as the printing press,allowing people everywhere to enhance th
40、eir writing and creativity and to share knowledge and ideas more widely and cheaply.The development of the printing press and exponential spread of ideas led to the Renaissance which fundamentally reshaped scientific,artistic,and religious views,facilitating an Age of Discovery which reshaped the wo
41、rld economy and had far reaching consequences on all continents,which continue to reverberate today.The development of the internet and World Wide Web means billions of people globally share ideas.This global pollination allows individual genius to be discovered wherever it is,and for collaboration
42、which builds on diverse insights to flourish.Generative AI offers the potential to catapult creativity,science,and collective intelligence to higher levels.Large language models and simultaneous translation means barriers between people speaking different languages will be lowered.This will increase
43、 access and personalization allowing for significant improvements in education and health outcomes.While 95%of scientific studies are published in English,under 5%of the worlds population speak English as their native language.Drawing on the collective intelligence of people everywhere not only mean
44、s that there are many more brains engaged in problem solving and innovation,but also that because they are more diverse the potential for disruptive breakthroughs is much greater.Our new Renaissance offers an extraordinary opportunity to address some of the greatest challenges facing humanity.The gr
45、owing potential of AI means new cures for cancer,Alzheimer and other terrible afflictions are more likely to be found,as are the means to generate low-cost clean energy and develop crops which can withstand climate change.Five hundred years later,we still celebrate the outstanding achievements of th
46、e Renaissance.But it ended in tears for many,with religious wars and the brutal rise of slavery and imperial power.In Europe,it was associated with the rise of fundamentalism,with the challenge to the status quo leading to the Bonfire of the Vanities,burning,and banning of books and inquisitions.The
47、n,as now,the growing concentration of wealth and potential of new technologies to create fabulous wealth for some and unemployment for others was a source of growing tensions and anger.Then,as now,those losing jobs(scribes in the 15th century and media folks today)were heard more often than those in
48、 newly created jobs(bookbinders and printers in the 15th century and programmers and knowledge workers today).Then as now,place and dynamic cities became more important as the knowledge economy accelerated,leading to a growing resentment against metropolitan elites.Then as now,connectivity and globa
49、lization led to the spreading of diseases and pandemics,including those which killed millions of native Americans.And then as now the power to use technologies to create false narratives and spread fake news became a tool for fragmenting societies and distrust of experts.But the Renaissance teaches
50、us that none of this can be taken for granted,as new technologies require a social license to operate.The challenge of our New Renaissance is to ensure that AI works for all.Ian Goldin Director of the Oxford Martin Programme on Technological and Economic Change Oxford Martin School,University of Oxf
51、ord This section draws on Ian Goldin and Chris Kutarnas Age of Discovery:Navigating the Storms of Our Second Renaissance,Bloomsbury,London,2016.September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 9 The Rise of Artificial Intelligence Can machines think?More than 70 years ag
52、o,Alan Turing posed this question in his seminal paper“Computing Machinery and Intelligence”.To answer this question,he famously proposed a game for thinking machines,the Turing test or imitation game,where an interrogator asks the same questions to a human and a computer and tries to find out which
53、 one is the human.Artificial Intelligence has been growing ever since but more significantly during the past decade.The recent advances in Large Language Models(LLMs)have taken the world by surprise and awe.Impressively,the social challenges associated with this progress have not changed significant
54、ly ever since Turing proposed this question.In 1950,the computer science pioneer Claude Shannon from Bell Labs introduced the first artificial intelligence application,a wheeled mouse named Theseus that methodically explored the surroundings of a 25-square maze and found its way out.1 Shannon wanted
55、 his mouse to navigate a labyrinth and escape it by learning its structure,resembling the thread used in Greek mythology to mark the heros path.The mouse itself was a magnet on wheels but underneath the maze a complex web of switches and relays allowed it to move and learn through trial and error.In
56、spired by Shannon,Micromouse competitions have been running since the 1970s building on the advances of technology.2 Figure 7.Claude Shannon Showing the First AI Application Source:Oxford Martin School A few years later,in 1958,Frank Rosenblatt from the Cornell Aeronautical Laboratory built Perceptr
57、on Mark I,the first image classification computer.The goal of this computer was to identify objects even when the images were taken across different orientations,sizes,colors,and backgrounds.3 1 Daniel Klein,“Might Mouse,”MIT Technology Review,December 19,2018.2 Micromouse Online,Homepage,accessed S
58、eptember 5,2023.3 Cornell Aeronautical Laboratory,Inc.,“The Perceptron:A Perceiving and Recognizing Automation,”January 1957.Pantelis Koutroumpis Director,Oxford Martin Programme on Technological and Economic Change Oxford Martin School“We may hope that machines will eventually compete with men in a
59、ll purely intellectual fields.But which are the best ones to start with?”-Alan M.Turing,Computing Machinery and Intelligence,1950.Theseus“inspired the whole field of AI”as“this random trial and error is the foundation of artificial intelligence.”-Mazin Gilber,Google,Director of Engineering,Telecommu
60、nications,Orchestration,Analytics,and Automation.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 10 Following this invention,Rosenblatt spurred controversy in the AI field when he openly supported that the perceptron,which was the algorithm used to classify objects in h
61、is computer,would become“the embryo of an electronic computer”that“will be able to walk,talk,see,write,reproduce itself,and be conscious of its existence.”4 In spite of the criticism,Rosenblatt also introduced the term back-propagating error correction in 1961,a key theoretical foundation of modern
62、neural networks,although he did not know how to implement this in his computer.The promise of computing technologies and modelling capabilities led other prominent researchers to follow Rosenblatt in his predictions.In 1958,political scientist Herbert Simon and computer scientist Allen Newell predic
63、ted that“within ten years a digital computer will be the worlds chess champion”,and“within ten years a digital computer will discover and prove an important new mathematical theorem.”Soon after,in 1965,Simon moved even further,supporting that“machines will be capable,within twenty years,of doing any
64、 work a man can do.”5 In 1967 Marvin Minsky,a leading AI scholar,predicted that“Within a generation.the problem of creating artificial intelligence will substantially be solved.”6 He then followed with an interview in 1970 stating that“from three to eight years we will have a machine with the genera
65、l intelligence of an average human being.”In the late 1960s and 1970s,the AI field experienced a number of setbacks which led researchers to introduce the term AI winter.Most of the efforts in the 1960s revolved around military and intelligence agencies that needed increased speed and decision-makin
66、g accuracy during the Cold War.7 One of these efforts involved the automatic translation of Russian documents into English.Despite the initial optimism,in 1966 the Automatic Language Processing Advisory Committee(ALPAC)concluded that machine translation was slower,more expensive and more inaccurate
67、than humans.8 A number of theoretical and practical issues led the field to the abandonment of connectionism which was linked to neural networks and perceptrons in favor of symbolic reasoning in the late 1960s.The setbacks continued in the 1970s with the UK parliament Lighthill report concluding tha
68、t AI had failed to achieve its grandiose objectives.In the U.S.,a shift to mission-oriented direct research,rather than basic undirected research led to further cuts and frustration among researchers.4 Mikel Olazaran,“A Sociological Study of the Official History of the Perceptrons Controversy,”Socia
69、l Studies of Science,Vol.26,No.3,August 1996 5 Herbert A.Simon,The New Science of Management Decision(New York,Harper&Row,1960)6 Marvin Minsky,Computation:Finite and Infinite Machines(New Jersey,Prentice-Hall Inc.,1967)7 Dafydd Townley,“Intelligence Agencies Have Used AI Since the Cold War But Now F
70、ace New Security Challenges,”The European Financial Review,May 22,2023.8 John Hutchins,“The History of Machine Translation,”PDF,accessed August 12,2023.September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 11 Figure 8.The Pre-Deep Learning Era in Artificial Intelligence The A
71、I winter in the 1960s and 1970s along with inflated expectations led the AI field to a pause.After the mid-1980s,new models emerged and continued the progress.Source:Epoch AI During the 1980s and 1990s the field gradually evolved producing new models that covered a diverse set of applications includ
72、ing the Neocognitron for handwriting and pattern recognition(1980),NetTalk for text to audio transformation(1987),ALVINN for autonomous vehicle navigation through a camera and a laser(1988)and Long Short Time Memory(LSTM)as a foundational breakthrough in neural network research(1991).Although the fi
73、eld experienced a stable progress during this period,large production systems using AI technology rarely made explicit references to it.As Nick Bostrom,professor of philosophy in Oxford,stated in 2006 a lot of cutting-edge AI has filtered into general applications,often without being called AI becau
74、se once something becomes useful enough and common enough its not labeled AI anymore.Because of this progress,the computing resources that supported AI systems kept a steady improvement pace doubling every 21 months from 1952 to 2010 and aligning the fields progress with the well-known Moores Law,wh
75、ich measures transistor density in semiconductor chips at a similar rate(Figure 8).Analyzing information about the computing resources used to train 123 milestone Machine Learning systems from Epoch AI and the analysis performed by researchers,we show that a sharp discontinuity in computing resource
76、s emerged around 2010.9 The rate of computing used post-2010 doubled every six months in FLOPs(Floating Point Operations Per Second)achieving a three-fold increase compared to 18-month doubling times in the previous period.The reason behind this change is largely attributable to the transition to De
77、ep Learning models since 2010.These models construct complex non-linear relationships between their inputs and outputs through a layered composition of their features.Combining features from lower layers led to dramatic performance improvements compared to models with shallower architectures.9 Epoch
78、 AI,“Studying Trends in Machine Learning”,accessed August 12,2023;Jaime Sevilla et al.,“Compute Trends Across Three Eras of Machine Learning,”PDF,accessed August 12,2023.The performance of AI systems has increased rapidly since 2010 Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023
79、 Citigroup 12 The Deep Learning revolution was underpinned by the successful implementation of the connectionist models(which were abandoned in the 1960s)and some key concepts that were introduced at the same time,including back-propagation.The combination of core inputs including human ingenuity,co
80、mputers,algorithms,data quantity,and quality led to a sharp transformation of the field.Since the late 2000s,Deep Learning models outperformed shallow ones in several machine learning competitions.Already in 2011 DanNet was the first model to achieve superhuman performance in visual pattern recognit
81、ion,outperforming traditional methods by a factor of three.In 2012 AlexNet won the large-scale ImageNet competition by a significant margin.Figure 9.The Deep Learning Era After 2010,the rate of computing resources used to train AI models rose by a factor of three,doubling every six months,compared t
82、o an 18-month doubling rate in the previous period.Source:Epoch AI The profound rise in performance by the post-2010 models,led firms to invest heavily in AI applications.This drove some of the earliest efforts to use AI and reach super-human results in closed-world situations like the game of Go.Tr
83、aining compute resources used for some of the most advanced AlphaGo models(Zero and Master)were only matched with huge LLMs five years later(Figure 9).Within this period,some important discoveries took place.Before 2017,the field of Natural Language Processing(NLP)was lagging computer vision in term
84、s of its human-level performance metrics(image and handwriting recognition,Figure 10).In June 2017,researchers from Google published a groundbreaking paper,titled Attention is all you need,that introduced a significant advance in the use of the attention mechanism as the main improvement for the Tra
85、nsformer model.With this,long sequences of text that were used for translation or prediction of the next word,would not rely on the last state of the encoder,as it was usually done with RNNs(Recurrent Neural Networks)but instead they would extract information from the whole sequence.Several importan
86、t NLP models used these advances including BERT and GPT-2,allowing reading comprehension to improve dramatically along with language understanding soon after(Figure 10).Transformer models were also used in other applications including models that predict protein folding,text-to-image(used for Latent
87、 Diffusion Models in conjunction with content delivery networks(CDNs)and Diffusion models)and are likely to become a general-purpose mechanism underpinning most AI applications.The large-scale model era September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 13 Armed with these
88、 improvements in the foundation models,AI firms increased the computing resources to improve their performance(Figure 10)and a new set of models emerged in the Large-Scale Era using more than 100 times the resources than the Deep Learning Era models were trained on.The breadth of applications was va
89、st,including models that outperformed Go masters,to protein-folding predictions and large language models.The algorithms underpinning these models paved the way to the term Generative AI allowing a huge range of possibilities to emerge.Figure 10.Superhuman Benchmarks Reached by AI Since 2020,ML mode
90、ls achieved supe-human performance in image,speech,handwriting recognition,reading,and language understanding Source:DynaBench,OWID The Landscape of Generative AI The progress in the recent developments in artificial intelligence has attracted strong interest from many firms aiming to integrate cont
91、ent generation and decision making in their processes.In a 2022 AI report,IBM measured the proportion of firms that have already deployed and plan to use AI in the coming year.China and India were leading in these metrics with 58%and 57%of the firms having already deployed AI respectively,followed b
92、y Italy and Singapore.10 Overall,the report found that vast majority(70%)of firms expected to use AI in the coming years(Figure 11).10 IBM,“IBM Global AI Adoption Index 2022,”May 2022.Vast majority of firms expect to use AI in coming years Citi GPS:Citi GPS:Global Perspectives&Solutions September 20
93、23 2023 Citigroup 14 Figure 11.AI Used Across the World AI is becoming a global technology.The leading countries that have deployed AI in 2022 are China and India(58%and 57%of the firms)followed by Italy and Singapore.Even those countries lagging behind are exploring AI use in the coming years.Sourc
94、e:IBM Looking into the market performance of the leading firms in AI,one can observe that the growth in the S&P 500 index in the first half 2023 largely came from firms that produce core components of AI technologies.In the first half of 2023,a majority of returns in the S&P 500 index came from the
95、seven biggest stocks,which were driven by optimism around AI.Beyond the usual suspects,like the Big Tech firms,the equipment vendors supplying the necessary computing infrastructure outperformed in that time period,as every firm that intends to deploy AI will need to use these resources.Analysts of
96、AI markets expect the growth of the AI market to continue as the breadth of applications is likely to fuel demand for more and faster graphic processing units(GPUs).These components are not only used for inference purposes(answering questions on existing models)but also during the development phase
97、of a model(training),and this demand scales linearly with headcount.As a result,there is no sign that the GPU shortage we have in 2023 will abate in the near future.11 The AI Productivity Boost Can AI tackle a pressing economic paradox?The famous quip by Paul Krugman,Nobel laureate in Economics,that
98、“productivity isnt everything,but in the long run it is almost everything”runs at the heart of every innovation.Despite its importance and the impressive technological change that has taken place in the recent past,productivity growth has been slowing down for decades across advanced economies.12 It
99、 is not surprising that industry leaders and academics often refer to the potential of AI technologies as a way to end this downward trend.While there is still no definitive answer,several researchers suggest that this new wave of large language models is very promising,based on preliminary results
100、from their studies.11 Guido Appenzeller,Matt Bornstein,and Martin Casado,“Navigating the High Cost of AI compute,”Andreesen Horowitz,April 27,2023 12 Ian Goldin et al.,“Why Is Productivity Slowing Down,”Oxford Martin School Working Paper No.2012-6,May 9,2021.The early winners in the AI race“Generati
101、ve AI could provide complementary tools to knowledge workers.These would be creating new tasks(for educators,nurses,creative workers,tradespeople,and even blue-collar workers)and providing inputs into better decision marking for knowledge work.”-Daron Acemoglu,MIT,Economics Professor September 2023
102、Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 15 In a recent paper,researchers from MIT investigated the effects of ChatGPT in the context of mid-level professional writing tasks.13 Assigning college-educated professionals to incentivized writing tasks and randomly exposing half of
103、them to ChatGPT,they found that those who used the Generative AI technology decreased their completion times by 0.8 standard deviations and increased their output quality by 0.4 standard deviations.The inequality between workers also decreased,as ChatGPT benefited low-ability workers more thus compr
104、essing the productivity distribution across the sample.The authors noted that ChatGPT mostly substitutes for worker effort rather than complementing worker skills,and restructures tasks towards idea-generation and editing and away from rough-drafting.Beyond the hard numbers,the researchers also foun
105、d that exposure to ChatGPT increased job satisfaction and self-efficacy but it also heightened the concerns about automation technologies(Figure 12).Figure 12.ChatGPT Impact on Writing Tasks ChatGPT increases the productivity of college-educated professionals in writing tasks,decreases the inequalit
106、y,and benefits low-ability workers the most.Source:Noy and Zhang(2023)13 Shakked Noy and Whitney Zhang,“Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence,Science,July 13,2023.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 16 Lookin
107、g at a different type of office worker,a paper by Erik Brynjolfsson released in 2023 looked into the productivity results of adding Generative AI to customer support agents.14 The AI system used in this study was based on a GPT family language model that was fine-tuned to focus on customer service i
108、nteractions.The authors found that the employees with access to this tool managed to increase their productivity by 14%,measured by the number of issues they resolved per hour.The results on the distributional impact of AI seem to align with the ones on college-educated professionals,as the study al
109、so found that the greatest productivity impact was on novice and low-skilled workers,with minimal effects on experienced and highly skilled workers.The authors found the AI model disseminated potentially tacit knowledge of more able workers and helped newer workers move down the experience curve.In
110、addition,they showed that AI assistance improved customer sentiment,reduced requests for managerial intervention,and improved employee retention.Figure 13.Generative AI on Customer Support Agents AI tools are round to increase the productivity of customer support agents by 14%,but the effects are ma
111、inly driven by novice and low-skilled workers.Note:These figures plot the coefficients and 95 percent confident interval from event study regressions of AI model deployment using the Sun and Abraham(2021)interaction weighted estimator.Source:Citi GPS Implementation Issues Although increased efficien
112、cies are great,most businesses run on legacy infrastructure with multi-year transitions needed to incorporate next-gen tools.Many businesses want to integrate these new tools with their data,in on-premise or domain models,but this takes time.There is also a rush to access high-end compute,data cente
113、r capacity,and time with leading AI providers.Of course,all of this will require more resources and that often involves zero-sum investment decisions(i.e.,taking investment dollars from other areas)which are never easy.For example,the recent announced pricing of Microsofts Copilot AI product came in
114、 higher than expected,and while this may highlight the value it offers,firms will have to find the resources to deploy products widely.14 Eric Brynjolfsson,Danelle Li,and Lindsey R.Raymond,“Generative AI at Work,”National Bureau of Economic Research,Working Paper no.31161,April 2023.September 2023 C
115、iti GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 17 Change is usually incremental as companies experiment,learn,and iterate.It takes time for businesses to develop processes and people to capitalize on opportunities.In the case of AI,some may be fearful of implementing a technology they
116、 do not trust or understand,let alone a technology that could substitute their roles.In addition to company guardrails for responsible AI,many industries are highly regulated,and AI tools will need to have adequate risk management,transparency,or explainability,and in time also auditability.For exam
117、ple,while in theory,AI offers significant opportunity in medicine,a prognosis needs to be explained.Humans-in-the-loop,as happens with aircraft autopilot or fact checkers,may be therefore needed,which would in turn will slow implementation.Generative AI Acceptance and Trust The fact that ChatGPT has
118、 become one of the fastest-growing consumer applications in history illustrates that consumers have enthusiastically embraced Generative AI.We also believe they have high trust levels for Generative AI-based interactions,supported by multiple studies.One such study from Capgemini based on a sample o
119、f 8,600 respondents across multiple countries suggests that 73%of respondents trust content written by Generative AI.Figure 14.Share of Consumers Who Trust Content Written by Generative AI Source:Capgemini,Citi GPS 73%79%75%74%74%73%72%72%72%72%72%72%70%0%20%40%60%80%100%AverageNorwaySpainUKFranceNe
120、therlandsCanadaItalyJapanSwedenU.S.AustraliaSingapore Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 18 The Next AI Era:Promises and Challenges Computational needs are increasing exponentially.The phenomenal achievements of AI models in the recent years are largely dri
121、ven by significant changes in the core inputs used to train these models.As a result,in the Large-Scale Era starting in late 2015,the compute doubling times appear to be almost twice as high(10-months)compared to the pre-2010 era.15 During the same period,the number of parameters used in large-scale
122、 models exploded by 10-fold every year,in line with the increase in the size of input datasets(Figure 15).16 In contrast,the average GPU improvements over the same period had a doubling rate in computer performance(measured as FLOPs/$)every 2.5 years.17 Figure 15.Rise in the Number of Parameters and
123、 Training Data for Recent AI Models Source:Epoch,Our World in Data Following these recent trends,researchers used data from existing models to predict the total cost of the future models and found that they will soon become impossible to sustain exceeding the U.S.GDP by the end of 2030.18 In a separ
124、ate study on the same subject,OpenAI estimates that by the end of the decade the costs to train large scale models will reach$500 million significantly different from the estimates from independent researchers(Figure 10).19 The reasons for these large disparities emerge due to the approaches used in
125、 the extrapolation process of each study.If we assume stable progress in the GPU trend per dollar or the algorithmic improvements,we end up naively extrapolating previous trends.The report by OpenAI on the other hand,explicitly reports a best-guess forecast assuming that the growth in costs will slo
126、w down in the future,which means that their results should be interpreted with caution.15 Jaime Sevilla et al.,“Compute Trends Across Three Eras of Machine Learning,”PDF,accessed August 12,2023.16 Julien Simon,“Large Language Models:A New Moores Law?”,Hugging Face,October 26,2021.17 Marius Hobbhahn
127、and Tamay Besiroglu,“Trends in GPU Price-Performance,”EPOCH,June 27,2022.18 Andrew Lohn and Micah Musser,“AI and Compute:How Much Longer Can Computing Power Drive Artificial Intelligence Progress?”Center for Security and Emerging Technology,January 2022.19 Ben Cottier,“Trends in the Dollar Training
128、Cost of Machine Learning System,”Epochai.org,January 31,2023.“Suddenly,anyone could fine-tune the model to do anything,kicking off a race to the bottom on low-budget fine-tuning projects.”-Anonymous researcher from Google Compute costs for models are rising September 2023 Citi GPS:Citi GPS:Global Pe
129、rspectives&Solutions 2023 Citigroup 19 In the naive OpenAI scenario,costs of training will surpass$233 billion in 2032 which seems to align with the reports from other researchers.In both cases this rise in computing resources is a signal for caution about the future capabilities of large models.If
130、performance needs to increase,researchers should also consider other ways to achieve this,either by improving the algorithms they use or by focusing on the quality of the data that are fed in their models.Figure 16.The Rise in Compute Costs Depending on the extrapolation process,researchers estimate
131、 that the cost to train large-scale models will soon become unsustainable for most firms.Some support that this time will come before the end of the 2020s and others in early 2030s Source:Lohn and Musser(2022),Epochai.org Open-Source Code and Competition Open-source code has played a significant rol
132、e in the development of machine-learning models from the early days of the Deep-Learning Era.Both in terms of datasets and models(including Vicuna and Stable Diffusion)the open-source community has provided support to the leading firms in the AI domain.In 2023 we already saw a long list of open-sour
133、ce applications that seem to compete with OpenAIs ChatGPT(like Vicuna with 13 billion parameters,Figure 17).In this section we outline some of the key elements that facilitated this process.First,software developers can easily share their code on Hugging Face,a company that made its name from its Tr
134、ansformers library built for NLP applications and its platform that allows users to share machine learning models and datasets.Its users span from researchers at universities and software developers to employees at Big Tech firms,giving Hugging Face a fertile ground for the exchange of ideas,experim
135、entation,and development.Second,notable algorithmic improvements like LoRA(Low-Rank Adaptation of Large Language Models)by Microsoft researchers deal with the problem of fine-tuning large language models instead of retraining them on the new corpus.LoRA can use the existing weights of a prohibitivel
136、y expensive general-purpose model(like Metas Llama)to specific applications by vastly reducing the number of trainable parameters.This is achieved by directing the Transformer attention blocks of large-language models to the specific inputs allowing this process to reach impressive results,almost on
137、 par with full model fine-tuning at a fraction of the cost and time.The Linux moment for AI and the reasons behind it Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 20 Third,the benchmarking process in the AI community has allowed researchers and firms to get a better
138、view of their progress against well-known targets,although this now coming to a saturation point where an 80%or 90%accuracy is not meaningful.This is why several researchers propose to introduce newer and more comprehensive benchmarks to evaluate their models.Figure 17.Competition in the Chatbot Are
139、a Vicuna,an open source chatbot based on Llama,outperforms other proprietary chatbots and scores close to the market leader in a range of tasks.Source:LMSYS Despite the commercial and open-source competition,open-source models are likely to continue to develop in a symbiotic rather than antagonistic
140、 way with the leading firms,providing more efficient algorithms and leveraging the global talent pool of researchers and software engineers that even large technology firms lack.Do we need more data to train AI models?With the increase in computing resources in the Large-Scale Era,vast amounts of da
141、ta have been added for training purposes(Figure 15).Data is almost every digitized bit of information,but in this setting,we refer to online available human generated(for text),annotated(for images),and curated in formation,including Wikipedia articles,IMDb images,various videos,online news articles
142、,coding material,and general interest fora.Building on these resources,firms and researchers can train and improve their models to produce more accurate results as their algorithms and computational inputs grow.However,this process is not without limits.More data to train AI models or better data qu
143、ality?September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 21 First,the issue of data quality appears to become more prominent as we move towards more advanced models.In a recent paper,researchers showed that by using a much smaller corpus,based on textbook quality data,AI m
144、odels can achieve much higher performance improvements compared to the training results from similar or much larger online coding fora.20 In this sense,the quality parameters of the inputs appear to have a dramatic impact largely overcoming the issues with scale and costs that large models faced.Sec
145、ond,the progress of large models for language and text-to-image generation have fueled predictions about the state of online content in the near future.Some predict that within the next decade,at least 50%of online content will be generated by or augmented by AI.21 The optimism of these predictions
146、relies solely on the cost reductions associated with text and image generation and often fails to understand the importance of human generated content that is needed for the improvements of these models.In 2023,researchers from Oxford and Cambridge uncovered a deep issue with large AI models when th
147、ese are fed with their own outputs as a training dataset.Based on their findings they tried to answer the question:What will happen to GPT-n once LLMs contribute much of the language found online?They observed that this self-feeding process will lead to a model collapse due to irreversible defects t
148、hat gradually trim the tails of their predictions compared to other models trained on original human-generated content.They called this effect the curse of recursion when models that inbreed end up effectively forgetting and missing the diversity of outcomes that we would expect from human-level int
149、elligence.22 In addition to the challenges noted above,we also recognize:1.Bias:Generative AI systems are based on large amounts of training data,which means that the results can be susceptible to bias or inaccuracies in the training data,particularly if it is built on the gender,racial,and myriad o
150、ther biases of the internet and society more broadly.Additionally,results can lack human/logical reasoning.2.Inequality:The economics associated with the operation of large language models could trigger further increases in inequality between those who have access to these capabilities and those who
151、 do not.One could take it a step further and argue that the emergence of a select few companies as gatekeepers of all AI development could effectively monopolize the industry and trigger a centralization of collective intelligence.3.Authenticity:While language models have become increasingly more co
152、herent,we observe that they have also become more fluent at generating factually incorrect statements and fabricating falsehoods.Recent research suggests that authenticity levels tend to improve with increases in the size of models(i.e.,in the number of parameters),but we believe more needs to be do
153、ne here,particularly given the levels of consumer trust in Generative AI interactions as well as the need for high levels of accuracy in certain industries such as financial services.20 Suriya Gunasekar et al.,“Textbooks Are All You Need,”Microsoft Research,June 20,2023.21 MarketsandMarkets,“Generat
154、ive AI Market by Offering,Application,Vertical,and Region:Global Forecast to 2028,”April 2023.22 Ilia Shumailov et al.,“The Curse of Recursion:Training on Generated Data Makes Models Forget,”PDF,May 31,2023.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 22 4.Infringeme
155、nts:Training of Generative AI models/platforms requires access to data lakes and question snippets billions of parameters that are constructed by software processing huge archives of images,text,and other forms of input(depending upon the type of the model).The AI platforms recover patterns and rela
156、tionships,which they then use to create rules,and then make judgments and predictions,when responding to a prompt.This process comes with legal risks,including intellectual property(IP)infringement.5.Existential concerns:The transformative potential of Generative AI has understandably triggered broa
157、der concerns around the existential threat posed by AI,with multiple experts taking contrasting views particularly against a varied regulatory backdrop.We do not believe it is our place to be taking a view on this debate other than to simply state(what might seem quite obvious to some)that we do not
158、 believe this debate is likely to go away any time soon.Regulatory and Copyright Concerns Over the past decade several proposals for regulating AI have been put forward,either by regulators or the private sector.In 2017,Bill Gates supported the introduction of a robot tax as a response to the so far
159、 unfulfilled expectation that robots would excessively substitute low-skilled manufacturing workers.23 Since then,the rising concerns for citizen surveillance in China through image recognition AI technologies and other means,have alerted governments in Europe and the U.S.24 In 2023,academics and ma
160、rket leaders proposed a pause on AI technologies due to the existential risk that they believe it represents for humans.25 These proposals about AI applications appear to be both unclear and hard to enforce.Further issues have recently surfaced due to the copyright issues for the data used to train
161、large language models and text-to-image applications.Without any change in copyright legislation in the U.S.,in June 2023 OpenAI and Microsoft were sued in a class action lawsuit for$3 billion over alleged privacy violations of their chatbot.26 This is not the first time these firms face similar iss
162、ues.In 2022,GitHub programmers filed a similar lawsuit for scraping their code without their consent.27 The regulatory front is also changing across the world.Perhaps the most advanced regulatory framework is currently proposed by the EU Commission and was amended to become even more demanding from
163、its original version in June 2023.28 It is likely that some of its core ideas will be followed globally as is common with the Brussels effect.23 Quartz,“The Robot That Takes Your Job Should Pay Taxes,Says Bill Gates,”February17,2017.24 Zeyi Yang,“The Chinese Surveillance State Proves that the Idea o
164、f Privacy is More“Malleable”Than Youd Expect,”MIT Technology Review,October 10,2022.25 Future of Life Institute,“Paul Gian AI Experiments:An Open Letter,”March 22,2023.26 Chloe Xiang,“OpenAI and Microsoft Sued for$3 Billion Over Alleged ChatGPT Privacy Violations,”Vice,June 29,2023.27 Chloe Xiang,“G
165、itHub Users File a Class-Action Lawsuit Against Microsoft for Training an AI Tool With Their Code”Vice,November 4,2022.28 Foo Yun Chee and Supantha Mukherjee,“EU Lawmakers Vote for Tougher AI Rules as Draft Moves to Final Stage,”Reuters,June 14,2023.“On artificial intelligence,trust is a must,not a
166、nice to have.”-Margrethe Vestager,Executive Vice President for a Europe Fit for the Digital Age September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 23 EU lawmakers initially proposed a risk-mitigation strategy in their AI Act with a focus on biometric surveillance,but they
167、soon moved towards harsher measures that include a mandate on companies developing Generative AI applications for full disclosure of copyrighted material used to train their systems.Their proposal also includes requirements for companies working on high-risk applications to perform a fundamental rig
168、hts impact assessment and also evaluate the environmental impact of their applications(for inference and training),which ties to the recent increase in computing resources for AI.In a preliminary assessment of leading proprietary and non-commercial LLMs,researchers from the Stanford Institute on Hum
169、an-Centered AI found that most of the leading applications lag significantly behind the requirements of the draft EU AI Act.29 In particular,they pointed out that many market-leading foundation models have limited disclosure for copyrighted material,uneven reporting of energy use,inadequate disclosu
170、re of risk-mitigation strategies and an absence of evaluation standards or auditing.systems.This suggests that market leaders need to work further to meet the compliance requirements put forward by the EU AI Act.Global Regulatory Landscape Mixed The emergence of Generative AI has unsurprisingly seen
171、 AI as a broader topic become a firm focus for policy makers around the world.However,the regulatory path taken so far has varied aside from the EU,which has been one of the first government bodies to formally propose a law specifically aimed at controlling the use of AI,much of the regulation intro
172、duced in other regions has been principle-based and non-mandatory.Given the stakes involved,we do believe that policy and governance is only going to get complicated ahead,but the pace and direction of its evolution will play a defining role.U.S.:Despite the U.S.being at the forefront of the AI boom
173、,both in terms of technological development and deployment,concrete federal legislation governing AI has yet to be enacted.Perhaps the most prominent issuance of regulation at the federal level is the Blueprint for an AI Bill of Rights,proposed by the Biden administration in October 2022.While feder
174、al legislation is still developing,regulation at the local and state level is more advanced.New York City,for example,has been one of the first municipalities to create a law explicitly referencing AI.EU:The EU has used existing legislation to rule against the improper use of AI.For instance,the Gen
175、eral Data Protection Regulation(GDPR)requires corporations to inform individuals when AI has been used to process personal data.More recently,Italy used GDPR rules to become the first European country to temporarily ban OpenAIs ChatGPT following privacy concerns.The EU has also gone one step further
176、 than many other countries and regions around the world when it comes to implementing new AI regulation.The European Commission published its first proposal for an AI regulation in April 2021,and the Council of the European Union adopted its common position(“general approach”)on the AI Act in Decemb
177、er 2022.In June 2023,the European Parliament(EP)voted to adopt its own negotiating position on the AI Act,triggering discussions between the three branches of the EU the European Commission,the Council,and the Parliament to reconcile the three different versions of the AI Act,the so-called“trilogue”
178、procedure.29 Rishi Bommasani et al.,“Do Foundation Model Providers Comply with the Draft EU AI Act?”Stanford University HAI,accessed August 17,2023.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 24 UK:Although the UK government has acknowledged the risks posed by AI,it
179、 has been keen to emphasize the benefits of,and its support for,the technology.In March 2023,the government released a white paper on how it plans to regulate AI.One of the more important developments from this publication is that the government intends to rely on existing regulators,rather than a n
180、ew entity,to devise appropriate measures to mitigate AI risks.While no new legislation has been proposed,the 90-page report does outline five key principles regulators should consider when placing guardrails on AI:(1)safety,security,and robustness;(2)transparency and explainability;(3)fairness;(4)ac
181、countability and governance;and(5)contestability and redress.In June 2023,the UK government announced it will host the first major global summit on AI safety.This is due to take place in late Autumn 2023.China:Authorities in China have been interested in regulating AI since the country devised the N
182、ew Generation Artificial Intelligence Development Plan back in 2017.The way China approaches AI regulation will likely be consistent with its approach to regulating other areas of prominent technology,such as internet or social media,where it operates strict censorship to control the flow of informa
183、tion.Specific to Generative AI,the Cyberspace Administration of China(CAC),seen as the countrys leading AI regulator,notably released draft measures to address concerns it has with the technology.Titled as the Administrative Measures for Generative Artificial Intelligence Services,the set of measure
184、s obliges tech companies in China to register Generative AI products with the agency alongside a risk and security assessment before it is available for public use.Following the release of the draft,the government opened the measures up for public consultation(this ended in May)to understand the dif
185、ferent aspects of regulation that might be needed.Furthermore,in early June China announced that the 2023 legislation plan of the State Council will include the submission of a“draft AI law,”among the 50 or so other measures up for review by the National Peoples Congress(NPC).India:As the worlds mos
186、t populous country,the largest democracy,and one of the largest and fastest growing economies in the world,Indias influence on AI usage continues to grow.In 2019,the government released its National Strategy for AI.The report highlights how five sectors healthcare;agriculture;education;smart cities
187、and infrastructure;and smart mobility and transportation stand to benefit the most from AI technologies.As recently as April 2023,the Ministry of Electronics and IT published a statement saying that it was not planning to regulate AI,pointing to the technologys positive impact on the economy.Septemb
188、er 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 25 Sector Opportunities from Generative AI After looking at the history,the opportunities,and the risks associated with Generative AI,we now drill down further into the individual subsegments that are benefitting from the emergen
189、ce of Generative AI.We start by looking at the technology value stack,and how Generative AI works its way from the Silicon layer all the way up to the services layer.Next,we look at how Generative AI is likely to affect the six supersectors and find Financials&FinTech to be the most affected.The Gen
190、erative AI Technology Stack We see five layers within the AI technology value stack,as outlined in Figure 18.Historically,the Silicon layer has been the de facto foundation of almost all technological shifts,and Generative AI is expected to drive significant growth for compute(i.e.,processing power)
191、,networking,and memory chips.In the Infrastructure&Platforms layer,we see the hyperscalers/cloud providers racing today to build the underlying infrastructure that enables Generative AI applications and services,but over time we expect to see differentiation.When it comes to Models and Machine Learn
192、ing Operations(MLOps),the open-source community is likely to be a key driver of innovation.Moving further up the stack,we believe nearly all software companies will be impacted in some form by Generative AI,and company-specific execution will be critical.Lastly,we believe Generative AI represents a
193、step forward from ongoing AI/automation initiatives at the Services layer.Figure 18.Generative AI Technology Value Stack Source:Citi Research Citi Research Global Technology Team Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 26 Silicon The Silicon layer has historical
194、ly been the de facto foundation of almost all technological shifts,including the rise of machine learning and deep learning in the past.The emergence of Generative AI is therefore no exception.We expect Generative AI will drive significant growth across the supply chain,led by greater demand for com
195、pute,networking,and memory chips.In this context,we also note initiatives by the hyperscalers to make their own custom Silicon for compute.The picture at this layer would be incomplete without talking about the semiconductor capital equipment and consumables sector,which we view as the“picks and sho
196、vels of the wider AI gold rush,”and hence good proxies for the emergence of Generative AI.Compute Looks to be the Key Beneficiary Within the semiconductor sector,Citi Research believes compute will be the key beneficiary from the emergence of Generative AI due particularly to demand for graphic proc
197、essing units(GPUs),accompanied by the adoption and continued co-existence of custom application-specific integrated circuits(ASICs).Heterogenous computing and different types of compute chips:Although Generative AI is relatively new,the process of sorting and understanding massive amounts of data(tr
198、aining)and making predictions(inferencing)has been around for decades.Training workloads have typically been performed using GPUs where processes are parallelized,while inferencing workloads have typically used central processing units(CPUs)where workloads are processed serially.As Moores law slowed
199、 over the past decade,chipmakers have turned towards a heterogenous computing approach utilizing multiple types of compute chips including not only GPUs and CPUs,but also others such as application-specific integrated circuits(ASICs)and field-programmable gate arrays(FPGAs).The objective of this het
200、erogenous computing approach is to optimize various workloads,balancing the trade-offs between performance and energy efficiency.Figure 19.Compute Chips:CPU vs.GPU vs.ASIC vs.FPGA Type of AI Chip CPU GPU ASIC FPGA Definition A general-purpose processor that can handle a variety of tasks.A specialize
201、d processor that is designed to handle parallel processing tasks,making it well-suited for AI workloads.A custom-built chip that is designed specifically for a particular task or set of tasks,making it very efficient at those tasks.A chip that can be programmed to perform custom logic functions,maki
202、ng it highly flexible and adaptable.Speed Medium to high High High Medium to high Power consumption Low to medium High Low Medium to high Cost Low to medium High High High Use case CPU GPU ASIC FPGA Large-scale training Small-scale training Real-time inference Low-power inference Source:Citi Researc
203、h We see a particularly strong inflection for GPUs:GPUs remain at the forefront of performance benchmarks when it comes to both training and inference.In addition,they offer the ability to grow ones AI infrastructure at scale(particularly for larger models),this while retaining their interoperabilit
204、y benefits.Therefore,we see a particularly strong inflection for them.Atif Malik Christopher Danely Laura(Chia Yi)Chen Masahiro Shibano Peter Lee Takayuki Naito September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 27 Adoption and continued co-existence of ASICs:An ASIC is a
205、customized integrated circuit designed for a specific use case.Typically,designing custom chips requires a lot of engineering and financial resources,as companies will have to continuously innovate to keep pace with computing demands.However,the economics may potentially be much more appealing for h
206、yperscalers who have the scale and,in the case of AI,may look into specialized hardware as an alternative to general purpose hardware,which then is customized via software.As such,both GPUs and ASICs will likely be used in the push to build the necessary infrastructure for AI,with ASICs potentially
207、being used to primarily infer smaller and more specialized models and GPUs for both training and inference of larger and often more complex models.Generative AI Requires Memory In addition to compute,demand for High Bandwidth Memory(HBM)and Double Data Rate 5(DDR5)memory is likely to increase from t
208、he growth in AI computing.HBM:HBM is a high-speed system in package memory technology that uses stacks of vertically interconnected DRAM chips and wide(1024-bit)interface to enable higher storage capacity and data bandwidth than memory products using conventional wirebonding-based packages.While the
209、 processing speed of the fastest graphic DRAM is 600 gigabtytes per second(GB/s),the comparable metric for HBM is up to 1,638 GB/s.This makes is very suitable for AI processing.DDR5:With the advancement of AI models,the memory requirement in some GPUs have quadrupled over the past four to five years
210、.This increase should drive DDR5 to become a mainstream product by 2024.Figure 20.HBM vs.GDDR Bandwidth Comparison Figure 21.HBM-PIM(Processing-in-Memory)Source:Company Reports,Citi Research Source:Company Reports Networking Chips and IC Design Services Can Also Benefit In addition to compute and me
211、mory chips,enterprises and hyperscalers need to evolve their network infrastructure to support the exponentially increasing amount of data processing associated with the emergence of Generative AI.With data centers becoming less and less collections of computers and more and more fleets of computers
212、 that are operated by a large operating system,there is a clear need to enable accelerated computing.The development of compute chips,such as GPUs and ASICs aimed at AI workloads,as well as other specialized hardware,is in turn driving the development of new hardware and software architecture for IC
213、 design.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 28 Picks and Shovels of the Wider AI Gold Rush The picture at the Silicon layer would be incomplete without talking about the semiconductor capital equipment and consumables players.We view these as the“picks and s
214、hovels of the wider AI gold rush”and hence also as beneficiaries from the emergence of Generative AI,particularly in EMEA and Japan.Despite a cyclical pullback in 2023,Citi Research sees industry wafer starts growing at a compound annual growth rate(CAGR)of about 8%over 2020-25 versus around 4%over
215、2015-20 driven by two broad demand vectors:digitalization and electrification,as well as technological shifts.Further out they see a 6%CAGR over 2025-30,factoring in geopolitical forces.This forecast captures the wider tailwind from AI within the two broad demand vectors and as such,the emergence of
216、 Generative AI as an additional tailwind that should further strengthen the secular credentials of sector.Figure 22.Semiconductor Industry:Wafer Starts Figure 23.Technological Shifts:Apple M1 Chip Size Evolution Source:Company Reports,Company Reports and Citi Research Estimates Source:Citi GPS,Citi
217、Research,AnandTech(October 20021)Infrastructure&Platforms The Infrastructure&Platforms layer in the technology value stack primarily incorporates major hyperscalers,e.g.,large cloud service providers,that provide services such as computing and storage at enterprise scale,as well as other digital inf
218、rastructure providers encompassing data center-oriented firms in the U.S.and telecom operators in Asia.Hyperscalers Are Racing Today Expect Differentiation over Time Given the myriad of enterprise use-cases,major hyperscalers/cloud providers are competing to build the underlying infrastructure to su
219、pport the adoption of Generative AI applications and services.With competition ramping and LLMs proliferating,we expect differentiation over time as hyperscalers compete within a matrix of trade-offs for their solutions,balancing parameters,domain-specific training,integration opportunities,performa
220、nce,accuracy,speed,and pricing of Generative AI platforms.Major hyperscalers/cloud providers are competing to build the underlying infrastructure to support the adoption of Generative AI applications and services.02468020030Leading EdgeMatureGeopoliticsM1(120mm2)M1 Pro(246mm2)M
221、1 Max(432mm2)Alicia Yap,CFA Andre Lin,CFA Arthur Pineda Asiya Merchant,CFA Carrie Liu Michael Rollins,CFA Ronald Josey September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 29 Hyperscalers could see multiple monetization opportunities as newer offerings are already showing si
222、gns of tangible revenue acceleration,driven by the ramp in Generative AI investment in and capital expenditure.Monetization opportunities include model as a service(MaaS),application programming interfaces(APIs)and software development kits(SDKs),plugins and partnerships,and hosting and consumption
223、fees.Differentiation over time will result in a federated approach from customers.The major Cloud providers will compete to host Generative AI workloads and we expect customers will favor a federated approach,to swap specific microservices,depending on use-case-specific needs for lower pricing versu
224、s features/speed.Open-sourced platforms are racing in parallel to build out AI infrastructure.We expect major platforms will continue to develop increasingly sophisticated Generative AI infrastructure and toolkits,thereby accelerating innovation through abstraction layers that can simplify enterpris
225、e adoption.Models and Machine Learning Operations(MLOps)The Models and MLOps layer encompass all types of models facilitating Generative AI from large generic foundation models to verticalized ones,associated entities such as hubs(which can simply be thought of as marketplaces for models),as well as
226、 other facilitating elements related to MLOps.The importance of this layer cannot be emphasized enough.Simply put,if the entire Generative AI value stack was a solar system,then the model would be its star.Consistent with software evolutions in the past,we see the open-source community driving innov
227、ation in this layer.A key trend to monitor is the rise of verticalized models.As well as models enabling on-device AI.Open-Source Community Driving Innovation There has been an exponential acceleration in Generative AI innovation driven by open-source marketplaces,open-source LLMs,and open-source ve
228、ctor databases.Innovations around model scaling,instruction tuning,and on-device AI are starting to take hold and may help enterprises more readily buy into trialing Generative AI capabilities,as these innovations allow enterprises to search/ask questions of their data without feeding more data into
229、 hyperscalers/researchers,that they could use to train more models.Alicia Yap,CFA Ronald Josey Tyler Radke Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 30 Figure 24.Large Language Models:Open-Source vs.Closed-Source Hugging Face Top 6 Ranked Models Date Added Compute
230、 Infrastructure-Hardware Fine-Tuned from Model Model Details and Best Use Cases Falcon-40B-Instruct May-23 AWS SageMaker:64 A100 40GB GPUs Falcon-7B Not tuned for a particular purpose.Acts as a base model and was trained on chats CalderaAI/30B-Lazarus May-23-Alpaca&Vicuna Text generation and natural
231、 language understanding(NLU)Falcon-40B May-23 AWS SageMaker:384 A100 40GB GPUs-Not tuned for a particular purpose.Acts as a base model and was trained on chats LLaMA-30b-supercot Apr-23-LLaMA with SuperCOT-LoRA LangChain prompting LLaMA-65B Feb-23-Not tuned for a particular purpose.Acts as a base mo
232、del and was trained on chats GPT4-X-Alpasta-30b May-23-Alpaca-LoRA Text generation and NLU Notable Proprietary LLMs Data Added Parameters Company Best Use Cases PaLM 2 May-23*340B Google Multilingual translation,code generation,reasoning tasks GPT-4 Jan-23*1T OpenAI Conversational AI,text summarizat
233、ion,NLU ChatGPT Nov-22-OpenAI Chatbot development,customer service,NLU OPT-175B May-22 175B Meta Content creation,data analysis,question answering GPT-3 Sep-21 175B OpenAI Text generation,sentiment analysis,text classification Note(*):The parameters for PaLM 2 and GPT-4 are estimated.Source:Citi Res
234、earch Rise of Verticalized Models and On-Device AI Are Trends to Monitor Verticalized models aimed at specific use cases:We are seeing more research/companies iterate on top of base models for specific use cases.These iterations use proprietary training data and creating new,smaller,and more efficie
235、nt models.Differentiation and outperformance versus foundation models.As foundation models proliferate and models scale(i.e.,parameter count),parameters could become less important than adapting each model for verticalized use-cases.Verticalized models have outperformed foundational models with enha
236、nced capabilities and better accuracy.By customizing models on unique company and industry training data,we believe companies can drive differentiation in LLMs over time and reduce commoditization risk.Models enabling on-device AI:Similar to the trend of smaller,more use-case specific models,hybrid
237、AI/on-device AI hopes to offer scalability,performance,and personalization at reduced costs,with the added benefit of on-premise/on-device security.If model size and prompt(i.e.,question)are below a certain threshold,then inference can run on a device(currently estimated to be 1 billion parameters,t
238、hough models with 10 billion are slated to work eventually).Larger models can use a hybrid approach to work across devices and the cloud.Bringing Generative AI capabilities closer to the source can also enable per-user alignment and tuning.As models become more user-case specific,the added benefit o
239、f a model that is on-device means that the model can run and train locally without exposing data to hyperscalers.Software and Applications Beyond a few application use cases that have already seen strong adoption,we suspect narratives in the application layer will take more time to play out relative
240、 to the underlying infrastructure layer.We expect nearly all software companies will be impacted in some form,creating a larger emphasis on company-specific execution in navigating the rapidly changing landscape.At the highest level of analysis,we see the opportunity for the application space to tak
241、e on a larger share of organizations budgets,enabling new monetization opportunities based on the increased value add.Amit B Harchandani Fatima Boolani Steven Enders,CFA Tyler Radke September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 31 Some of the most common use cases for
242、 Generative AI are customer service,content creation,data augmentation,code generation,and research and development.The integration of Generative AI capabilities into software,applications,and workflows represents another key driver of disruption.Data Warehouses/Analytic Platforms:Data/analytics sof
243、tware is among the most exposed software categories to Generative AI as it“sits closer to the underlying infrastructure and could benefit from the increased compute/data intensity of LLMs.We believe the sub-category could benefit broadly with increased prioritization of data/analytics projects as or
244、ganizations modernize environments to be able to leverage LLMs.In addition,we see a potential acceleration of“data democratization”tailwinds with the proliferation of search-based/low complexity tools in the hands of more knowledge workers.Front Office/CRM Software/Digital Commerce:Generative AI has
245、 several compelling potential use-cases within the front office space that can help drive efficiencies and better customer relations/engagement based on the large volume of data.Industry research firm Gartner estimates suggest that by 2025,30%of outbound marketing messages from large organizations w
246、ill be synthetically generated,which is up from less than 2%in 2022.Use cases within e-commerce,like human image generation for modeling and showing alterations of different poses/ages/diverse representations,are also becoming commonplace.Gartner data also indicates that customer experience is curre
247、ntly the most common primary focus of Generative AI investments,cited by 38%of respondents polled.Back Office Software:Across our sub-categories HR Software,Collaboration Software,and Financial Software we see varying levels of clarity in how Generative AI will be or already is incorporated into fea
248、ture sets.HR software vendors are incorporating Generative AI into their platforms with use cases like recruiting,talent management,and core HR processes.On the other hand,collaboration vendors are still relatively early in integrating Generative AI into their platforms,with existing workflows prima
249、rily targeting process automation or adding a level of intelligence to the software.Enterprise Financials is likely to be one of the slower areas of uptake due to sensitivity around core financial data.Cybersecurity and Infrastructure Software:In the realm of cybersecurity,Generative AI presents mea
250、ningful opportunities but also poses threats.Threat actors will likely look to lean further into Generative AI to drive attach speed and augment attack capabilities.Security vendors are leveraging Generative AI to elevate analyst capabilities and streamline threat intelligence.As a result,we view Ge
251、nerative Ai as a force multiplier and a positive for cyber vendors.We also see it as a solution for the dire talent shortage that exists today in the Security industry.Services Generative AI represents a step forward from ongoing AI/automation initiatives across the IT&Business Process Management(BP
252、M)Services space.However,productivity savings could be more meaningful.As with other advancements in the past,there will be share shifts.Over time,companies will likely build their own solutions/business units,mirroring the trajectory we saw with the shift to the cloud.Jobs will likely be eliminated
253、 but these are likely to be replaced by others.Therefore,on balance,we believe Generative AI is neutral to net positive for the overall space.However,the majority of impact in the case of the Services layer is delayed because it is Services that makes technology adoption a reality.Amit B Harchandani
254、 Ashwin Shirvaikar,CFA Surendra Goyal,CFA Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 32 Figure 25.Generative AI Impact on Services Source:Citi Research Compute(“Silicon”)is first on the timeline.Infrastructure&Platforms use compute to build underlying infrastructur
255、e to support Generative AI adoption.Models&MLOps are the basis of Generative AI.Without the LLM,the subsequent steps cannot follow.Software&Applications:Outcomes will be company-specific based on functionality delivered and executionServices outcomes are likely to be all over the place based on func
256、tionality,positioning,investment level,trust/advisory capabilities,etc.Generative AI Realization TimelineServices Involvement Across Generative AI SpectrumLessMoreOccasional engineering talent deployment in the development process.(1)Engineering R&D Outsourcing(2)Roll-out of foundational digital tra
257、nsformation architectures.(1)Data-related work(feed/train LLM model)by BPM companies to build use cases and frameworks.(2)R&D outsourcing demand.(3)Services/Software crossover models(1)Software module development support by a handful of Services companies(2)Channel partnership/ecosystem growth acros
258、s Software spectrum.(1)Consulting/Advisory.(2)Custom use-case development(3)Third-party software development.(4)Data annotation,cleansing,etc.(5)Business process optimization.September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 33 Generative AI:Assessing the Risk/Reward Outs
259、ide Technology Although the majority of talk surrounding Generative AI focuses on the opportunities it creates,there are also likely to be risks created.To get an overall picture of the how sectors outside of the Technology stack could be affected by Generative AI,we devised a two-stage framework to
260、 assess risk/reward,which we believe can be broadly applied across businesses and sectors.Figure 26.A Simple Decision Tree to Gauge AI Risk vs.Opportunity for Affected Companies Source:Citi Research Two-Stage Framework to Assess Risk/Reward The first stage of our framework involves(1)mapping the ris
261、k from Generative AI,based on both first-order and second-order implications,(2)determining the implications for the cost base,and(3)gauging end-industry risk.The goal of this exercise is to identify the degree which companies within a sector are affected versus classifying companies into buckets of
262、“winners”and“losers.”In the second stage,we use a simple two-step decision tree to gauge risk/reward for those companies within a sector deemed to be affected from the emergence of AI after the first stage.Factors in the decision tree are based on the principles that(1)a relevant database trumps a g
263、ood algorithm,and that(2)consumers move faster than institutions.Amit B Harchandani Thomas A.Singlehurst,CFA Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 34 Stage 1:Mapping AI Opportunity and Risk The mapping exercise in the first stage of our framework results in an
264、 AI risk score based upon four inputs.This score then helps us to differentiate between companies within a sector which are more and less affected by AI.First-order vs.second-order considerations:The distinction between first-order and second-order considerations is closely related to short-term ver
265、sus medium/long-term considerations.An easy way to look make this distinction is by looking back at history.When software first emerged in the 1970s/80s,it undoubtedly was thought to be game-changing firms employing white collar professionals,such as bankers,lawyers,and accountants.But five decades
266、later,these firms all benefited from the productivity saves generated by computing.Opportunities vs.risks:We should also consider that for some of the companies,even those perceived short-term to be at risk,the disruption of AI-based technology may end up being positive for their prospects.Likewise,
267、what at first seems like a potential windfall,may end up being unhelpful.We use these distinctions to come up with an AI score based on four inputs that we describe in more detail below.The score is based on a simple average(i.e.,with equal weighting)of the numerical scores allocated to each of the
268、inputs.First-order implications:What do we anticipate will be the first impact of AI?Often,this is the near-term or direct implication rather than how we think AI will play out longer-term.Second-order implications:Here,we try to capture the second order/longer-term effects of AI.In some cases,the i
269、mplications may be more significant but in others it may be less so.Cost implications:Looking at the cost implications specifically is important.While for some companies,the emergence of AI-based technology will require additional investment,for others it will potentially drive greater cost efficien
270、cy via automation.This needs to be captured within any scoring system.End-industry risk:Although we likely already captured some of the end-market risk in the second-order implications,we think it important as a long input to capture not only any significant adverse exposure to sectors that may be d
271、isrupted,but also to take into account things like whether a particular companys client base is more skewed to organizations/institutions or individual consumers.Stage 2:Gauge Risk/Reward for Those Deemed to be“Affected”Once we determine an AI score for companies within a sector,we then use a simple
272、 two-step decision tree to gauge risk/reward for those deemed to be affected from emergence of AI after the first stage.By doing this,we can hopefully determine if the risk is positive or negative,and the speed at which the risk is likely to appear.We base this exercise on two underlying principles:
273、(1)a relevant database trumps a good algorithm,and(2)consumers move faster than institutions.September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 35 Principle 1:A relevant database trumps a good algorithm:However good a companys search/analytics technology is,if it is traini
274、ng on a database that is sub-par,the quality of the technologys output will be compromised.This reminds us that it is not only important to consider the size of a database but also to look at its relevance and accuracy.Applying this principle to our analysis,we need to look at whether the industry i
275、s“domain-specific”when determining whether a company is likely to be adversely affected.Principle 2:Consumers move faster than institutions/enterprises:In general,consumers are more likely to be quick to adopt new generative AI-based tools than institutions/enterprises.Based on this,institutional us
276、er bases will likely take longer to disrupt.Reflecting on both points,we think it is fair to assume that,at the very least,institutional user bases will take longer to disrupt and,possibly more likely,for those companies serving institutions,there is more of a chance of being the intermediary when i
277、t comes to AI deployment.Considering both principles,we create a simple decision tree for investors as shown in Figure 26.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 36 Generative AI:Impact Across Supersectors With the two-stage framework defined,Citi Research analy
278、sts around the world were asked to assess the adoption curve for Generative AI by applying the framework to their company coverage lists across the seven global supersectors.The results of the exercise can be seen in Figure 27 and shows the order of impact from Generative AI across the supersectors.
279、The most impacted supersector is Financials&FinTech(with variations across sectors)followed by Consumer.Natural Resources&ClimateTech looks to be the least impacted,sitting at the bottom of the spectrum.Figure 27.Generative AI:Order of Impact Across Supersectors We show the Tech&Comms supersector se
280、parately as it is the enabling supersector Source:Citi Research September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 37 Financials&FinTech The tenor and frequency of AI conversations within Financials&FinTech depends on the specific sector.The interesting overall takeaway is
281、 that questions being asked around AI are more like“What can banks/brokers/asset managers/insurance companies do with Generative AI?”rather than“How will Generative AI disrupt the industry?”This is partly because end-markets in this sector are largely regulated therefore regulators will have a say i
282、n the speed and design of AI take-up.Also,financial services companies are already spending a good deal of money upgrading their technology away from legacy tech,which could limit widespread adoption in the near term.Key opportunities from AI/Generative AI adoption?Current use-cases from Generative
283、AI and AI more broadly in Financials&FinTech tend to focus on improving customer service models or optimizing costs.However,Citis Financial Services and FinTech team believes Generative AI has potential to(1)help democratize investing and markets,(2)improve algorithmic trading,(3)enable more robust
284、data utilization;and(4)allow for better analyzing and pricing of insurance risk.Other opportunities include:Improved customer experience:Providing personalized services,using chatbots and virtual assistants to provide 24/7 support and handle customer inquiries,and helping human agent and services in
285、dustry participants improve the effectiveness of their conversations.Fraud detection and prevention:Identifying fraudulent activities in real time.Business risk management and decision-making:Helping financial institutions make informed decisions in terms of macro and market risk assessment.Regulato
286、ry compliance:Automating compliance checks and monitoring.Digital operations transformation:Speeding up the sectors digital transformation journey.More robust data utilization and more effective algorithms:AI could help drive more robust data and utilization of data that would create more effective/
287、efficient algorithms that investors utilize.Improved investor education:Lowering the barrier to entry for less sophisticated investors.Democratization of wealth management:Filling the education gap between advisors and retail investors who are new to alternative investments.Property&Casualty insurer
288、s:play a key role in helping clients(policyholders)Lowering the cost of capital and protecting clients economic output by invest in AI for revenue-generating or cost-efficiency programs.Better pricing of risks:Deploying AI to better analyze and price insurance risks.Ashwin Shirvaikar,CFA Judy Zhang
289、Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 38 Key risks to existing business models from AI/Generative AI adoption?There are functional risks around the use of AI in the Financials&FinTech sector,although there is also potential for competitive risk from not invest
290、ing in AI.Key risks specific to the Financial&FinTech sector include:Generative AI as a driver of fraud,especially identity fraud.Reputational and regulatory factors that arise if AI strategies are not properly executed.Market manipulation from AI and“meme stock”-like events.Replacement of tradition
291、al financial advisor models with AI-driven“advice.”Underperformance by asset managers from not properly synthesizing AI into the investment process.Inadvertently/unintentionally running afoul of privacy rules.Malfunctioning of AI.Consumer AI tools are increasingly being discussed and deployed across
292、 business operations within Consumer Staples companies,who are seeing these tools as levers to improve speed and efficiencies of operations.This includes areas like speed of innovation,how to use data more efficiently in manufacturing,supply chain and marketing.AI is also increasingly being apprecia
293、ted for its dual potential to help companies reach sustainability goals(it can help with discovering new materials,but also with reducing wastage and energy consumption in production).In areas like Traditional luxury,the industry remains a late adopter of technology,and the AI debate continues to be
294、 marginal,but in Jewelry,companies see potential for AI in areas like design and modeling but not yet in cutting and polishing.However,it may be used to help with grading,which would improve the efficiency of certification.Interestingly,Hotel companies are actively engaging in AI discussions,while C
295、onsumer Retail companies are exploring ways to use generative AI to handle customer inquiries,facilitate text-to-shop,and finding items in stores,as well as increasing efficiency for inventory and supply chain management.Key opportunities from AI/Generative AI adoption?Mass customization and persona
296、lization:Generative AI can be used to generate content in multiple languages and thousands of versions without having to significantly increase headcount and or budget.Product authenticity at POS:Generative AI has the potential to fight against counterfeit products,with machine learning helping to i
297、dentify subtle irregularities such as in shape,color,texture,label used,etc.This is particularly relevant for high-end spirits and cosmetics offerings.Facial recognition:In-store facial recognition can drive improved personalization of the customer experience in certain countries.Simon Hales Tiffany
298、 Feng Paul Lejuez,CFA,CPA Steven Zaccone,CFA September 2023 Citi GPS:Citi GPS:Global Perspectives&Solutions 2023 Citigroup 39 Innovation:Leveraging digital technologies enables R&D departments to innovate“better,faster,and more effectively.”This is partly achieved as AI enables companies to understa
299、nd the world in a more predictive way.AI tools accelerate scientific discovery by taking on multiple tasks that previously required labor-intensive lab work.Cost cutting:Generative AI tools should help to speed up day-to-day admin operations such as communications,marketing campaigns,coding,and anal
300、ysis without impacting the size of workforce.Manufacturing and supply chain:AI can be used to simulate the manufacturing process,which can simplify the process and improve its sustainability(i.e.,more energy efficiency,less waste).Moreover,supply chains can be more flexible and reactive to shortages
301、,with AI able to uncover alternative ingredients or simplify products by reducing the number of components without impacting their quality or effectiveness.Digital twin applications can also aid in the design process of the manufacturing process.Marketing and sales planning:AI can improve marketing
302、effectiveness to make campaigns perform better;for instance,providing insights to improve logos,placements of pictures,and sound in advertisements.(See this interview with KH Data Head,.)AI can improve data insights for sales planning purposes.AI can be used to assess data at more granular customer
303、levels,for instance neighborhoods or postcodes,which can then be used to target stores and customize for different neighborhoods.Chatbots:Companies can improve online customer service through various social media platforms(hotels and the tourism industry have been early adopters).Customer satisfacti
304、on:Operational efficiency and customer experience in the tourism industry could be improved by AI in areas like:(1)tailor-made services catering to customers preferences(i.e.,personalized itinerary based on different tourism destinations and activity suggestions,rooms set up prior to arrival);(2)pre
305、cise marketing to target groups through data analysis and improve search engine optimization(i.e.,high-quality content generation);(3)real-time customer support in pre-booking and during the trip and timely identification and resolution of issues;(4)more accurate demand forecasts and pricing strateg
306、ies;(5)digitalized experience(i.e.,smart robot service,online check-in and check-out);(6)automating tasks to relieve staff workloads and reduce cost.Hedging:Using AI to predict commodity price movements can potentially help make better hedging decisions.Key risks to existing business models from AI/
307、Generative AI adoption?Building trust in tools that may not always be fully understood and accepting recommendations that may contradict prior beliefs.The risks associated with generative AI include consumer data vulnerability and potential bias in models.Adopting generative AI would mean incrementa
308、l costs incurred through entering into partnerships or developing in-house upskilling.Data privacy and liabilities/copyright issues from AI-generated content.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 40 Brands making decisions or implementing solutions based on AI
309、 algorithms should be careful to avoid the introduction of unfair bias.In luxury product design,some aspirational elements are always important and cannot be totally replaced by AI.It may also be challenging to execute the transformation of the current digital platform to adopt AI while making sure
310、that it is net-profit positive for the business.Healthcare Major pharma companies are becoming more vocal in talking about their investment and the potential of AI across multiple parts of their value chain.The potential for AI use is seen across commercial,drug development,trial,and discovery.Other
311、 use cases include trial design protocols,automated administrative efforts,and patient/physician education.In Health Tech,the use of AI has consistently been a topic of discussion and while there could be benefits for its use in certain administrative use-cases,its use in a clinical setting draws sk
312、epticism.Key opportunities from AI/Generative AI adoption?Within large pharma,we think the best use cases for AI broadly are in drug discovery and design,patient selection and recruitment,and optimization of sampling/sales calls.Within nontherapeutic healthcare,the sub-sectors payment integrity,util
313、ization management,risk coding,Stars score improvement,and revenue cycle management are key areas of early adoption.Generative AI in particularly can be extremely helpful in(1)patient communication,(2)health coaching(e.g.,reminding patients to perform healthy habits),and(3)note taking/data input int
314、o electronic health records(effectively reducing time physicians spend on admin tasks).Key risks to existing business models from AI/Generative AI adoption?Within pharma,there is little disruptive risk to large cap pharma given the requirements to establish efficacy and safety through comprehensive
315、clinical trials and pre-clinical testing.AI should allow improvement of discovery hit rate and reduce clinical attrition.We also anticipate AI to accelerate clinical trials through improved patient selection and trial recruitment.Generative AI may accelerate the downsizing of both physical and remot
316、e sales reps,improving margins.Healthcare data is highly regulated.If generative AI cannot be trained on vast amounts of healthcare data(due to regulatory burden),this would limit effectiveness and adoption.Industrial Tech&Mobility The broader debate around using advanced technology in the Industria
317、l Tech&Mobility space has been around for a long time,specifically using digitalization as an ongoing growth and margin driver.Commentary on AI is at a very early stage.The Aerospace&Defense sector is looking at AI to assist with data fusion and manipulation,as well as for advanced cryptography.In t
318、he Autos&Mobility sector,AI will likely lead to a greater focus on autonomous vehicles but can also be instrumental in driving product development,manufacturing,and customer-facing services.Andrew Baum,MD Patrick Donnelly Martin Wilkie Andrew Kaplowitz Itay Michaeli September 2023 Citi GPS:Citi GPS:
319、Global Perspectives&Solutions 2023 Citigroup 41 Key opportunities from AI/Generative AI adoption?Generative AI Generative AI in low code/no code industrial systems:Industrial processes are increasingly digitized,with industrial IoT platforms allowing end-users to create industry-specific and process
320、-specific apps.Generative AI could significantly broaden the ability to create code,massively opening up the market for analyzing data on the industrial IoT platform.Interactive applications:Generative AI is expected to contribute to the development of robots that better understand and interact with
321、 humans.Customer facing:Some customer choice applications(e.g.,buying a car)could be an opportunity.AI more broadly Process optimization(production,supply chain):Production processes and supply chains are increasingly complex,with the amount of data available to make an optimization decision increas
322、ingly too large(and ever-changing)for legacy optimization models.A combination of machine learning and quantum computing could vastly improve these optimization decisions and increase supply chain efficiency.On design,digital twin applications(for example as automakers build out new EV platforms)can
323、 also be areas that benefit from this optimization.Infrastructure needed for data center growth:AI will require a lot of compute power,meaning data center infrastructure,in energy use,power management,cooling and adjacent markets like heat reuse will become increasingly important as the datacenter i
324、nfrastructure needed to power AI is rolled out.Data manipulation:Defense systems could use AI to drive data fusion,distillation and faster decision making.Accelerated autonomous vehicle development:AI can be leveraged to further accelerate the deployment and scaling of autonomous vehicles under vari
325、ous domains and business models.Key risks to existing business models from AI/Generative AI adoption?End-markets like production,installation,construction,extraction,and transportation are seen as having below-average risk of impact,and therefore the broader risk to business models within Industrial
326、 Tech&Mobility is likely less impactful than in some other sectors.The debate is in the early stages,and we think will form part of the broader debate on cybersecurity and data sovereignty for the Industrial IoT.In certain mission-critical and highly regulated applications,for example aerospace,cert
327、ification and safety is a key factor the lack of repeatability and non-auditable decisions may make certification very difficult.Citi GPS:Citi GPS:Global Perspectives&Solutions September 2023 2023 Citigroup 42 Real Estate The real estate sector is generally less exposed to AI versus other sectors.Ho
328、wever,for data centers,the debate is whether they will be positive or negatively impacted by the rise of Generative AI deployments.Demand for space and power are likely to rise,especially early in the adoption cycle.But there is a risk that accelerating computing will dilute the returns of existing
329、data centers or drive some of them into obsolescence.Key opportunities from AI/Generative AI adoption?Data centers:Accelerating high-performance computing adoption associated with AI workloads is a potential positive for data center demand for at least the first few years of this adoption cycle.Alth
330、ough there is a risk that existing general computing workloads are cannibalized initially by the accelerated IT infrastructure,spending is more likely to be immediately focused on an expanding array of learning and inference Generative AI models.Smart buildings:Smart buildings,which automate feature
331、s like HVAC,lighting,alarms,leak sensors,and security could be improved by AI to lead.Chatbots:Further advancements in the use of AI for chatbots could help consumer-facing applications at buildings such as hotels and residential.Efficiency:AI could facilitate enhancements to the supply chain by pre
332、dicting product demand and ultimately optimizing logistics,while driving productivity improvements in areas such as lease writing,valuation,due diligence,and legal.Key risks to existing business models from AI/Generative AI adoption?Data centers may need further investments:For data center firms,hig
333、h power density requirements needed with accelerated computing may not be within the design parameters of existing data centers,possibly increasing the need for investment in existing builds facilities or requiring new data center builds.Office space:There are conflicting views on the impact to office space with bulls considering the growing office demand from AI companies as a positive in the nea