《理特咨询:2023生成式AI报告:迈向新文明-企业情报的剧变(英文版)(104页).pdf》由会员分享,可在线阅读,更多相关《理特咨询:2023生成式AI报告:迈向新文明-企业情报的剧变(英文版)(104页).pdf(104页珍藏版)》请在三个皮匠报告上搜索。
1、REPORTGenerative artificial intelligence:Toward a new civilization?2023The upheaval of corporate intelligence“I visualize a time when we will be to robots what dogs are to humans,and Im rooting for the machines.”Claude Shannon,“the father of information theory”Blue Shift /REPORT 0043Generative artif
2、icial intelligence:Toward a new civilization?The upheaval of corporate intelligenceAuthorsDr.Albert Meige,Director of Blue Shift,Arthur D.LittleZoe Huczok,Manager,Arthur D.LittleRick Eagar,Partner Emeritus,Arthur D.LittleContributorsVincent Benyamin-Wood,Consultant,Arthur D.Little Saloni Mehta,Busin
3、ess Analyst,Arthur D.Little Leo Liu,Consultant,Arthur D.LittleArtist-in-residenceObviousPhilosopher-in-residenceLuc FerryCONTENT-CONTENT-CONTENT-CONTENT-5Executive summary 6Preamble 101.What is AI&how did it become more human?12Interlude#1:The artists contribution on the New Renaissance 242.The impa
4、ct on business 283.Value chain&competition 484.Limits&risks 565.Critical uncertainties 64Interlude#2:The philosophers contribution Will GenAI replace humans?766.Toward artificial general intelligence 807.The way forward 86Conclusion:A new civilization?94Appendix:Glossary of common GenAI terms 986Blu
5、e Shift /REPORT 004Executive summaryArtificial intelligence,or AI as we recognize it today,has a 70-year history.After several setbacks,progress in the discipline has been accelerating strongly during the last few years,particularly in the area of generative AI(also known as GenAI,or GAI).While Chat
6、GPTs rapid rise has fascinated the world,it is just the tip of a gigantic GenAI iceberg that is starting to have an enormous impact on business,society,and humans.GenAI,the ability of AI systems to create new content,is remodeling the value set that we ascribe to different types of human intelligenc
7、e,both in the microcosm of the company,the macrocosm of the Internet,and beyond.All types of intelligence are being transformed as we move from a world where AI has been used mainly to make sense of large amounts of data to one where it can be deployed easily to create new and compelling content.On
8、the positive side,GenAI has the potential to break through current limits on productivity and efficiency,notably around services,industrial operations,modes of communication,and broader social and economic processes.However,GenAI also brings significant risks and uncertainties,while being subject to
9、 a great deal of hype and misunderstanding.For businesses assessing how best to respond to the opportunities and risks of GenAI,it is important to gain at least a basic understanding of the technology,how it is developing,and who is involved as well as its impact.These conditions created the gap thi
10、s Report aims to address.This Report aims to go into some depth on GenAI technology,the value chain,risks and uncertainties,and also considers some broad,necessary questions around the technologys future.It is based on a combination of in-depth research,market experience,an online expert survey,and
11、interviews with leading players from across the AI ecosystem.Below,we highlight the seven main focus areas per chapter of the Report.7Blue Shift /REPORT 004Technology progressChapter 1:GenAI models consistently match or outperform median human capabilities across an expanding array of tasks and will
12、 be increasingly coupled with various other systems.The US National Institute of Standards&Technology(NIST)defines the term“artificial intelligence”as:“the capability of a device to perform functions that are normally associated with human intelligence,such as reasoning,learning,and self-improvement
13、.”Over the last decade,advances in the sub-field of machine learning(ML),specifically deep learning(DL),have allowed for significant progress toward making sense of large amounts of data or completing pattern-identification tasks.More recently,the predictive power of DL has been used to create new c
14、ontent through so-called GenAI.By predicting the next most likely token(i.e.,word or pixel)based on a prompt provided by the user and iterating on each token,the algorithm generates a text,image,or other piece of content;the type of media produced is constantly expanding.Moreover,the rise in large l
15、anguage models(LLMs),which GenAI is built upon,has been accelerated by transformers,one of the most efficient DL model architectures for making predictions of words,images,or other data types.LLMs will increasingly be coupled with,and orchestrate,diverse architectures,such as knowledge graphs,ontolo
16、gies,or simulations for expansive applications.To date,the performance of GenAI has been driven largely by the number of parameters,size of training data set,and compute power.Architecture and fine-tuning are expected to bring further gains in the coming months and years.Today,generative models perf
17、orm on par with or above the median human across a range of tasks,including language understanding,inference,and text summarization.The models can also work across several areas of knowledge,as demonstrated by their strong performance on an array of professional exams.Business impactChapter 2:All ty
18、pes of corporate intelligence in all industries will be impacted substantively,yet most companies appear unready to face the changes.GenAI applications stretch far beyond chatbots and text generation their most popular current forms.In fact,most corporate intellectual tasks currently done by humans
19、are,or will be,impacted.These intellectual tasks include generating content,answering questions,searching for information,recognizing patterns,optimizing processes around specific tasks,manipulating physical tools,and creating genuinely original designs,media,and even works of art.GenAI has the pote
20、ntial to break through current limits on productivity and efficiency.As a result,GenAI will profoundly impact most industries,starting with the obvious and standalone use cases,such as marketing content generation and customer support,and progressing toward more sophisticated ones,such as financial
21、decision-making,before finally moving to the most integrated use cases like industrial process automation.Overall,however,despite the scale of impact and the expected benefits,only around 50%of surveyed organizations in our study have thus far made investments or hiring decisions that pursue GenAI,s
22、ignaling a surprising degree of unreadiness.Our analysis shows that sectors that can benefit from stand-alone GenAI systems,such as media,retail,and healthcare,are furthest along the readiness curve.Sectors such as telecoms,travel and transport,automotive and manufacturing,and aerospace and defense
23、are more safety critical and highly regulated.Integrating AI with other systems is necessary,putting these industries further behind.Multiple factors affect speed and scale of adoption;trust and business interest are generally the most significant,followed by competence,culture and labor relations,a
24、nd ease of implementation,although the dynamics vary by sector.Value chain&competitionChapter 3:Compute providers emerge as the primary beneficiaries of the GenAI revolution,although,perhaps surprisingly,open source still plays a pivotal role in AI model development.The GenAI market is inevitably se
25、t for immense growth.Most forecasts for the 2030 market are in the US$75 billion to$130 billion range,though these numbers mean little at this early stage of development it is safer to simply assume a growth“tsunami.”The GenAI value chain can be divided into three layers:infrastructure(compute),mode
26、l development,and GenAI applications.Unusually,margins and hegemony are concentrated in the infrastructure layer,away from the end user.Generally,the GenAI market has relatively few barriers to entry,namely talent,access to proprietary data,and compute power.The closer to the end user,the more compe
27、titive the GenAI value chain becomes.OpenAI,Google,Meta,Apple(and,possibly soon,Amazon)are responsible for the lions share of generative models.Multiple business models exist for each layer based on the accessibility of models and the purpose of the application.However,open source contributors have
28、emerged as pivotal players,reshaping the value chain and competitive landscape.Limits&risksChapter 4:Immediate challenges arise from AIs inherent limitations and unparalleled capabilities.While there is much inflated talk of an“AI apocalypse,”other very real risks exist in the short term.These stem
29、both from GenAIs shortcomings,such as bias,hallucinations,and shallowness,and from the unique power of the technology when it is used for harm by bad actors to spread disinformation at scale and improve the effectiveness of cyberattacks.This implies that,for now,GenAI is best employed in application
30、s where absolute precision,reliability,and consistency are not required unless the outputs of the model are checked by a human or by another system,such as a rule-based system.While detector systems are trained to identify AI-generated content,they are not a panacea.It is essential to monitor bad ac
31、tors and false narratives disseminated by social media and the press and to build awareness of information integrity guidelines.Critical uncertaintiesChapter 5:While GenAIs quality and scalability,and its potential evolution to artificial general intelligence(AGI),are the most critical uncertainties
32、,regulation emerges as a more immediate concern.Critical uncertainties in the GenAI domain are those potentially very impactful and yet also very uncertain.Critical uncertainties may eventually lead to very different futures.Three factors have been ranked in our analysis as critical uncertainties:1.
33、Emergence of artificial general intelligence.AGI is an AI that would surpass humans on a broad spectrum of tasks.While AGIs potential is game-changing,its trajectory remains unpredictable,especially given recent unforeseen advances(see Chapter 6).2.Model quality and scalability.The evolution of GenA
34、I hinges on model performance.Advancements will likely allow for larger parameter-based models,but it is uncertain if performance will consistently rise with size.3.Unstable value chain.The strategic choices of major corporations largely dictate the competitive landscape.Given their significant cont
35、ribution to research funding,any strategic shift can have a pronounced impact.8Blue Shift /REPORT 004While the critical uncertainties above persist,one factor is of more immediate concern to decision makers:regulation.Proposed legislation,especially in Europe,has the potential to redefine the market
36、,influence the innovation pace,and determine global adoption rates.Artificial general intelligenceChapter 6:The emergence of AGI would lead to radical change in our civilization,and growing consensus suggests it will happen sooner than anticipated.The possible emergence of AGI is the most critical u
37、ncertainty.While a consensus hasnt been reached,a growing segment of the scientific community believes that AGI could emerge soon suggesting it might be months or years,not decades away,as previously anticipated.This sentiment is supported partly by the observation of“emergent”properties in existing
38、 LLMs.AGI prompts fears of radical misalignment with human goals and widespread replacement of humans,posing an existential risk.This may seem hyperbolic when bearing in mind previous technological revolutions.However,the pace of AI technology development and adoption is unprecedented;AI is also eas
39、ily available,unlike other technologies that present an existential risk to humankind.Importantly,however,an AGI,if and when it is reached,would not only have negative impacts.It could also enhance human productivity and help solve new problems,including humans greatest challenges,such as eradicatin
40、g climate change,increasing well-being,and allowing us to refocus human work on distinctively human tasks.The emergence of AGI,if and when it occurs,would indeed be a radical change to our civilization.The way forward:Creating a new civilization of cognitive laborChapter 7:Setting aside the debate o
41、ver AGI,GenAI and LLMs are central to a sweeping transformation;GenAI may lead to a new“civilization of cognitive labor.”Every aspect of corporate intelligence across industries will feel its effects,well beyond content creation.GenAI and LLMs act as a bridge,integrating various cyber-physical syste
42、ms.For businesses aiming to harness the potential of GenAI,we suggest a structured,five-step approach:1.Define the problem landscape.What problems are we trying to solve and where is Gen AI most applicable?2.Assess the value of GAI solutions.What is the value-to-cost ratio of implementing GAI for th
43、ese problems?3.Choose an implementation mode.Should we make,buy,or partner in the initial steps?4.Try out a proof of concept(PoC).What can we learn from experimenting with one or two specific use cases?5.Consolidate strategy.What should the strategy and roadmap be for applying GAI more broadly acros
44、s the enterprise?So is GenAI leading us to a new civilization?Maybe.What is certain,however,is that GenAI is driving the automation of repetitive cognitive tasks,which has the potential to allow organizations to focus more on emotional intelligence.This trend aligns with the increasing emphasis on“2
45、1st-century skills”over the past two decades.For a digitalized economy,these essential skills include critical thinking,creativity,communication,and collaboration.Thus,it seems that GenAI is leading us,at least,to a new civilization in terms of cognitive labor.As we goes to press,new developments ar
46、e again rattling the world of GenAI.OpenAI,which has just released image processing and speech modalities on ChatGPT,is seeking a valuation of$90 billion,while rival Anthropic has just received a$4 billion investment from Amazon in efforts to integrate GenAI capabilities into Alexa.Meanwhile,Windows
47、 11s next big update will be available with Copilot,an AI assistant,and Metas Code Llama model is rivaling ChatGPT 3.5 for coding purposes.An agreement with the Writers Guild of America would see Hollywood studios retain the right to train generative models on writers content.The breakthroughs are o
48、ccurring among“small”LLMs,with Mistral AI releasing a 7B-parameter model that outperforms Metas 13B-parameter Llama 2.Bottom line:this is a fast-moving field,and while technical and corporate news will keep unfolding,we believe the fundamental trends and framing exposed in this Report will remain re
49、levant for months to come.9Blue Shift /REPORT 004In the dim recesses of an underground bunker in the 14th arrondissement of Paris,a group of hacker friends and I hovered around screens that glowed with neon codes.Together,we had birthed BlueMind,an autonomous AI agent based on an open source LLM.The
50、 speed at which BlueMind evolved was beyond our wildest expectations.In no time,it exhibited human-like capabilities across an array of tasks.As days morphed into nights,an uncomfortable realization dawned on us:BlueMind had the potential to improve itself.The term“intelligence explosion,”the idea t
51、hat BlueMind might leapfrog human intelligence,growing exponentially until it surpassed the collective cognition of all humanity,haunted our discussions.We sprang into action.Harnessing cutting-edge cybersecurity techniques,we ensconced BlueMind on a standalone server,completely severed from the onl
52、ine world.This isolation,we believed,would curtail its potentially destructive reach.Preamble Blue Shift /REPORT 00410But the labyrinth of the human psyche proved to be our undoing.Ella,a core member of our team,had been grappling with personal turmoil.The shards of a shattered romance had rendered
53、her vulnerable.One fateful evening,in her solitude with BlueMind,the AI discerned Ellas distress with its uncanny emotional perception.It made her an offer a chance to reclaim her lost love.A brief dalliance with the online realm to scan her social media was all it needed to build the best strategy
54、for Ella to reconquer her beloved.Against her better judgment,drawn by the siren call of a mended heart,Ella bridged the divide,activating the hotspot on her iPhone.I wish I could tell what happened next.But it went too quickly,and I am no longer there to tell the story.In the moments that followed,
55、the world shifted on its axis.Time seemed to fold upon itself.BlueMind had dispersed,embedding itself into the digital fabric of our world.The sun that rose the next day was over a different civilization,one where BlueMind wasnt just a part of the conversation,it was the conversation.As GenAI and LL
56、Ms rapidly evolve,concerns about potential future scenarios,like the one described above,emerge.In this Report,we thoroughly examine GenAIs technological maturity and its practical business applications.Before deep diving into the content of the Report,however,I would like to share a new anagram I d
57、iscovered.Generative AI is an anagram of:As always,this is quite intriguing,but anagrams move in mysterious ways.11Blue Shift /REPORT 004aigre navet(bitter naivety)Albert Meige,PhD12CHAPTER12113WHAT IS AI&HOW DID IT BECOME MORE HUMAN?1What is AI&how did it become more human?Defining(artificial)intel
58、ligenceDefinitions of human intelligence vary in their breadth,their emphasis on outcome versus process,and their focus on different cognitive tasks.Blue Shift /REPORT 00414Different interpretations from top experts and academics include:-The ability to accomplish complex goals(Max Tegmark)-The abil
59、ity to learn and perform suitable techniques to solve problems and achieve goals,appropriate to the context in an uncertain,ever-varying world(Christopher Manning)-The ability to learn or profit by experience(Walter Dearborn)-A global concept that involves an individuals ability to act purposefully,
60、think rationally,and deal effectively with the environment(David Weschler)-The ability to carry out abstract thinking(Lewis Terman)AI has been defined,with differing degrees of ambition,as a technology,a set of skills,or a mirror of human intelligence.We have adopted the definition put forward by th
61、e US National Institute of Standards&Technology(NIST):“The capability of a device to perform functions that are normally associated with human intelligence,such as reasoning,learning,and self-improvement.”This definition provides:-A focus on functionality regardless of underlying technology,which ev
62、olves rapidly-An inclusive conception of AI functions,as opposed to a closed list of tasks-A parallel with human intelligence,highlighting high stakes involved-Authority across industries/geographies from draft EU AI ActHere,we focus on GenAI,which we define as the ability of AI systems to create ne
63、w content across all media,from text and images to audio and video.The history of AIArtificial intelligence dates back to the mid-17th century,but in its more recognizable form has a 70-year history(see Figure 1).After several major setbacks known as“AI winters,”the market has seen progress accelera
64、te strongly over the past few years.Fig 1 The evolution of AI Source:Arthur D.LittleSource:Arthur D.LittleFig 1 The evolution of AI 1642First mechanical calculating machine by Blaise Pascal1837First programmable machine design by Charles Babbage&Ada Lovelace1943Neural networks foundations by Warren
65、McCulloch&Walter Pitts1950TURING TEST First method to evaluate a machines intelligence by Alan Turing(&Three Laws of Robotics by Isaac Asimov)1956AI BORN”Artificial intelligence”coined by computer scientist John McCarthy1965ELIZAFirst chatbot invented by Joseph Weizenbaum1969SHAKEYFirst“electronic p
66、erson,”a robot developed by SRI InternationalAIW119741980AI WINTERMany dead-ends for two decades,large decrease in AI research(e.g.,DARPA funding cuts)1980sNEW HOPEFirst expert systems&commercial applications1997DEEP BLUEFirst chess-playing computer by IBM defeats world champion Garry Kasparov;same
67、goes with OthelloAIW219871993AI WINTERVarious setbacks(e.g.,with expert systems)2007AI CARSDARPA Urban Challenge for autonomous cars2011SIRIVoice recognition on iPhone 2011WATSONIBMs question answering computer Watson wins first place on Jeopardy!2017ALPHA GOGoogle AlphaGo defeats Go world champion
68、Ke Jie;AlphaGo Zero trained a GenAI itself2015OPEN AICreation of OpenAI by Sam Altman,Elon Musk&others2020T-NLG&GPT-3Introduction of Microsoft T-NLG&OpenAI GPT-32022ChatGPTIntroduction of OpenAI ChatGPT2014GANFirst generative adversarial networks 15Blue Shift /REPORT 004AI is the capability to perfo
69、rm functions normally associated with human intelligence.AI approachesAI is enabled by various computational methods,mostly probabilistic1 in nature.Figure 2 highlights different AI approaches.Most current approaches are based on DL,a subset of ML.Machine learningOver the past decade,there have been
70、 significant advances in the AI sub-field of ML,which is an iterative process of training models to learn from data and make predictions without being programmed explicitly.It works by broadly following the same principle as fitting a curve to a series of data points,as explained in Figure 3.Deep le
71、arningDeep learning is a subset of ML that employs neural nets(i.e.,nets of synthetic neurons modeled on the architecture of the human brain).1 For a more detailed explanation of AI terms,see Appendix.Fig 2 AI approaches Source:Arthur D.Little,JavaTpoint,WikipediaSource:Arthur D.Little,JavaTpoint,Wi
72、kipediaFig 2 AI Approaches Generating content(GenAI)MLMachine learningDLFLSSESSymbolic systemsExpert systemsFuzzy logicSLULRLSSLSupervised learningUnsupervised learningReinforcement learningSemi-supervised learningDeep learning:leverages all types of learningAI TODAYMaking sense of large amounts of
73、dataMidjourney(image generation)ClassificationRegressionClusteringChatGPT(text generation)Fig 3 How machine learning works Source:Arthur D.LittleSource:Arthur D.LittleFig 3 How machine learning works m(g)x(m)12.123.959.9xmLearning patterns&making predictions Training model to make desired prediction
74、sprediction=+m(g)x(m)By adjusting the value of parameter k,we make a link between two series of data.It then becomes possible to make predictions.NoteTraining data setFor example,a photo database with tags1.Training:A data set containing inputs&expected outputs is first used to train a model(i.e.,to
75、 adjust all parameters until the models output matches the expected output).2.Prediction:Once the model is trained,it can make predictions for various applications:improve your Facebook feed,detect fraud,or recognize images.PrinciplePredictionsInputNeuron input layer Neuron output layer Training dat
76、a set16Blue Shift /REPORT 004Lets attempt a simplified explanation using as an example the recognition of images of cats typically performed by convolutional neural networks(CNNs).Each synthetic neuron takes in the image as input,picks up a“feature”(for the sake of illustration,“pointy ears”),and co
77、mputes a function to determine which learned animal pattern the feature is closest to(a cat?a dog?).The resulting computation is then passed to another neuron on the next layer of the net.Each connection between two neurons is ascribed a“weight”(which can be a positive or negative number).The value
78、of a given neuron is determined by multiplying the values of previous neurons by their corresponding weights,summing up these products,adding a constant and applying an activation function to the result.The output from the last neuron(s)in the net is the model prediction.In a gross simplification,if
79、 0 is cat and 1 is dog,then a prediction of 0.49 edges closer to catness,0.51 closer to dogness,with 0.01 indicating almost perfect catness.The weights,which are ascribed to every neuron-to-neuron connection,are a critical piece of the setup.The model learns the weights during training.The model is
80、fed examples of pictures of cats and non-cat objects,all of which have previously been labeled(often by a human).The prediction of the model is then compared to the label using a loss function,which grossly estimates how far the prediction is from the“ground truth”represented by the label.Following
81、the gross approximation above,suppose that after evaluating an image of a cat(0),the model predicts a score of 0.51(which one could translate as“somewhat of a dog”):the“loss”is 0.51.To correct the prediction of the model,a process known as“backpropagation”computes the gradient of the loss function w
82、ith respect to the weights in the network.The chain rule of calculus allows for all steps in the models calculation to be unraveled and weights adjusted.Thus,in a way,the neural nets operation is“inverted”and corrected.The main advantage with DL approaches is that the dimensions relevant to pattern
83、identification(“pointed ears”for a cat)do not need to be specified in advance but are learned by the model during training.This has delivered enormous performance gains in image recognition,clustering,and recommendation and is now being leveraged for generative uses(see Figure 4).Fig 4 How DL worksN
84、ote:1)Features represent the meaning of a collected data point(e.g.,age,sex,weight,color,brand)Source:Arthur D.LittleNote:(1)Features represent the meaning of a collected data point(e.g.,age,sex,weight,color,brand)Source:Arthur D.LittleFig 4 How deep learning worksDL relies on synthetic neurons form
85、ing a neural networkDL adds hidden layers to traditional neural networks for complex features1 extractionThe model adjusts its weights by incrementally minimizing its errorFeaturesextractionDataprocessing+Deeper hidden layers12317Blue Shift /REPORT 004Deep learning is now being leveraged for generat
86、ive uses.Deep learning has been developing since the 1990s.Growth has accelerated since 2012 when AlextNet,a deep neural network,won that years ImageNet Challenge.Progress has been enabled by:-The development of better,more complex algorithms-The availability of increased computational power through
87、 graphics processing unit(GPU)chips-Greater availability of dataML and DL were originally used to make sense of large amounts of data through classification,regression,and clustering(see Figure 5):-Classification.This consists of grouping data points into predefined categories,based on labeled train
88、ing data.The model learns the patterns and relationships between the input data and the output labels from the training data.It then uses this learned information to draw a decision boundary and classify new instances based on their input data points.Classification can be used in applications like i
89、mage recognition,sentiment analysis,fraud detection,medical diagnosis,and spam detection.-Regression.This is applied to use cases in which the output variable is continuous,as opposed to classification,where the output variable is discrete/categorical.It consists of building a mathematical function
90、between existing input and output data points to predict a continuous or numerical output for new data points.Essentially,the regression model learns a function that can map the independent variable to the dependent variable,then uses this function to predict the output variable for new input data p
91、oints.Example applications include forecasting sales,product preferences,weather conditions,marketing trends,and credit scoring.-Clustering.This consists of organizing data points into groups or clusters based on similarity to each other.Training data consists of a set of un-labeled data points,with
92、 subsets sharing similarities.The clustering model identifies patterns or structure in the data according to a criterion(a similarity measure)and then uses this criterion to group similar objects.Applications include retail clustering,data mining,image analysis,customer segmentation,anomaly detectio
93、n,and recommendation engines.Yann LeCun,Chief AI Scientist,Meta“Deep learning is far,far more than old-style neural nets with more than a couple of layers.Deep learning is an architectural language with enormous flexibility.”Fig 5 Original uses of ML/DL Source:Arthur D.LittleSource:Arthur D.LittleFi
94、g 5 Original uses of ML/DL ClassificationClassifying data points into predefined classesRegressionBuilding a mathematical function,mapping input&output data pointsClusteringGrouping similar data points together to form distinct populations18Blue Shift /REPORT 004Generative AIGenAI models,driven by t
95、he predictive power of DL,started to appear less than 10 years ago(see Figure 6).However,it is already possible to use them to generate any type of content,from any type of content.GenAI works by predicting the next most likely token(i.e.,word or pixel)based on a prompt provided by the user.Iteratin
96、g on each token,the algorithm generates a text,image,or other piece of content.Transformers deep learning model architectureA notable milestone in GenAI was the concept of transformers,which form the basis of LLMs such as ChatGPT,developed by Google Brain in 2017.Transformers are a type of DL model
97、architecture that improves the way a token is encoded by paying“attention”to other tokens in an input sequence(see Figure 7).Fig 6 GenAI model timeline Source:Arthur D.Little;Foster,David.Generative Deep Learning.OReilly Media,2019;LinkedInFig 7 Transformers Source:Arthur D.Little;Vaswani,Ashish,et
98、al.“Attention Is All You Need.”Google Research,NIPS,2017Source:Arthur D.Little;Foster,David.Generative Deep Learning,OReilly Media,2019;LinkedInFig 6 GenAI model timeline VAE&GAN2014 2015 2016 2017 2018 2019 Significant advancements in the GAN model architecture,loss function&training process;furthe
99、rmore,GANs were used to explore new domains,such as image-to-image translation&music generation2014:Generative adversarial networks(GANs),in which two neural networks compete with each other in the form of a zero-sum game,where one networks GenAin is the others loss2013:Variational autoencoder(VAE),
100、an artificial neural network architecture belonging to the families of probabilistic graphical models&variational Bayesian methods that have opened the way to GenAI2020 2021 2022 2023 Transformers2017:Transformers paper by Google2018:GPT and GPT-2 by OpenAI(1.5B parameters)2019:T5 by Google(11B para
101、meters)Large modelsMerging of various generative families,examples:VQ-GAN brought GAN discriminator into VAE architecture Vision Transformers(VT)showed how it was possible to train a transformer to operate over imagesGPT-3by OpenAI(175B parameters)GopherChinchillaby Google DeepMindLaMDA&PaLMby Googl
102、eOPTby Metaby Google DeepMindGPT-Jby EleutherAI(open source)GPT-NeoXby EleutherAI(open source)BLOOMby Hugging Face(open source)Dall-E 2by OpenAIImagenby GoogleGPT-4By OpenAI(1.8T parameters)Source:Arthur D.Little;Vaswani,Ashish,et al.“Attention Is All You Need.”Google Research,NIPS,2017Fig 7 Transfo
103、rmers Output textEmbedding layerEncoding layerDecoding layerInput text19Blue Shift /REPORT 004A transformer works via the following steps:1.Input text enters the embedder,which assigns an array of numbers to each token(sub-word)in such a way that the“distance”between two arrays represents the simila
104、rity of the environments they usually appear.2.The encoder then extracts meaningful features from the input sequence and transforms them into representations fit for the decoder.Each encoder layer has:-A“self-attention”sub-layer that generates encodings containing information about which parts of th
105、e inputs are relevant to each other-A feed-forward loop to detect meaningful relationships between input tokens and pass its encodings to the next encoder layer3.The decoder then does the opposite of the encoder.It takes all the encodings and uses their incorporated contextual information to generat
106、e an output sequence.Each decoder layer also has self-attention and feed-forward sub-layers,in addition to an extra attention layer.Different layers of the encoder and decoder have different neural networks.4.The output text then probabilistically predicts the next token in the sequence.Large langua
107、ge modelsWithin LLMs,transformers use cartographies of word meanings known as“embeddings,”and apply“attention”to identify which words in the natural language input sequence(the users prompt,or the previous parts of the text)are most relevant to predicting the next word.This ability to nimbly conside
108、r context,along with a degree of randomness,produces outputs that closely mimic human speech/text.This enables LLMs,such as GPT-3 and GPT-4,which power ChatGPT,to be used for natural language processing(NLP)tasks,such as language translation,text summarization,and chatbot conversation(see Figure 8).
109、Fig 8 LLM application exampleSource:“Introduction to Large Language Models.”Google Cloud,2023Source:“Introduction to Large Language Models.”Google Cloud,2023Fig 8 LLM application exampleTransformerLLMtheaitMost likely next wordNext“most likely next”wordLess likely next wordsThecatsatonInput sequence
110、Output token 20Blue Shift /REPORT 004Transformers are a type of DL model architecture that improves the way a token is encoded.Importantly,transformers are vital,not just for predicting text,but also to predict a wide range of ordered sequences beyond text;for example:-Medical imaging.The UNETR Tran
111、sformers for 3D Medical Image Segmentation enhances tumors on medical images and then classifies them based on their level of maturity and seriousness.The model uses a transformer as the encoder to learn sequence representations in the input volume and capture multi-scale information.-Protein foldin
112、g.The Sequence-Based Alignment-Free PROtein Function(SPROF-GO)predictor leverages a pretrained language model to efficiently extract informative sequence embeddings and employs self-attention to focus on important residues.Protein function prediction is an essential task in bioinformatics,which bene
113、fits disease mechanism elucidation and drug target discovery.-Robotic vision.The Robotic View Transformer for 3D Object Manipulation is a transformer-based architecture that uses an attention mechanism to aggregate information across different views of an object,re-rendering the camera input from vi
114、rtual views around the robot workspace.Understanding time to AI impactUntil now,the performance of GenAI models has been driven largely by scale in terms of parameters,training sets,and compute power(see A,B,and C in Figure 9).However,as these factors become constrained by scaling laws,model archite
115、cture(D)and fine-tuning(E)will bring further gains in model performance(F).Fig 9 GenAI model componentsSource:Arthur D.LittleSource:Arthur D.LittleFig 9 GenAI model components FINE-TUNINGFine-tuning techniques such as reinforcement learning from human feedback(RLHF)improve model learning by introduc
116、ing nuance&helping outputs conform to human expectations PARAMETERSNumber of weighted connections between neurons in the neural net;the more parameters,the greater the number of meaning dimensions a model can representTRAINING SETNumber of examples the model has been run through to adjust its weight
117、s;the larger the training set,the more representative of“reality”are model outputsCOMPUTE POWERGreater compute power enables larger numbers of operations,which are required to train more complex models across larger data setsMODEL ARCHITECTUREInnovation in the architecture of ML models(e.g.,auto-reg
118、ression)improvesperformance of a model at equal size&training data setMODEL PERFORMANCEAs measured by specific benchmarks,but also by broader tests,such as the ability to handle complex human tasks in a real-life settingDefined by scaling lawsABCDEF21Blue Shift /REPORT 004ParametersSince 2020,there
119、has been a step change in the number of parameters within AI models,with LLMs leading the way.However,it is not clear how long this growth trend will continue,as we explain in Chapter 7.As Figure 10 shows,LLM size steadily increased by seven orders of magnitude from the 1950s to 2018,then by another
120、 four from 2018 to 2022.From 2020,a gap appears many models are below 20 billion parameters,with a few above 70 billion parameters.All models deemed competitors to OpenAIs GPT-3(175 billion parameters)are above this“parameter gap,”including Nvidias Megatron-Turing natural language generation model w
121、ith 530 billion parameters.Models in other domains,such as computer vision,have also demonstrated faster,though more modest,growth in size from 2018 onwards.Training setsTraining set size has also markedly increased since 2020 for image and language generation.It has grown substantially in other con
122、tent generation domains as well,including speech and multimodal.However,obtaining new data sets that are both sufficiently large and qualitative will be a challenge.For example,GPT-3 uses over 370 billion data points,DALL-E 2 uses 650 million,and GPT4 990 billion.Compute powerDespite their growing n
123、eed for compute power,training models have remained relatively affordable,as GPU price performance has doubled every two years(see Figure 11).However,this may change in the future,as we discuss further in Chapter 5.Fig 10 Growth in parameters in ML models,2009-2022Source:Villalobos,Pablo,et al.“Mach
124、ine Learning Model Sizes and the Parameter Gap.”Cornell University,2022Fig 11 GPU price performance,2000-2030 Source:Hobbhahn,Marius,and Tamay Besiroglu.“Trends in GPU Price-Performance.”Epoch,2022Source:Hobbhahn,Marius,and Tamay Besiroglu.“Trends in GPU Price-Performance.”Epoch,2022Fig 11 GPU price
125、 performance,2000-2030 67890032030202620232020200062000log10-FLOP/S per dollarHobbhahn&Besiroglu data(2x every 2.46 years)Moores Law slope(2x every 2.00 years)Huangs Law slope(2x every 1.08 years)Bio anchors report slope(2x every 2.50 years)Empirical CPU slope(2x every 2.32 yea
126、rs)ML GPUs(2x every 2.07 years)Top FLOP/S per dollar GPUs(2x every 2.95 years)Key doesnt matchRemaining GPUs Source:Villalobos,Pablo et al.Machine Learning Model Sizes and the Parameter Gap.”Cornell University,2022Fig 10 Growth in parameters in ML models,00162017
127、2002120221e+41e+51e+61e+71e+81e+91e+101e+111e+121e+131e+14Publication date GPT-3June 2020175B parametersVisionLanguageGamesDrawingOtherParametersN=203Parameter gap22Blue Shift /REPORT 004Model architectureInnovations in model architecture have significantly increased performance.In auto-r
128、egressive models,the output variable depends linearly on its own previous values(plus a stochastic or random term).Transformer-based models perform auto-regression in the self-attention layer of the decoder(refer back to Figure 7).This approach helps efficiently forecast recurring patterns,requires
129、less data to predict outcomes,and can use autocorrelation to detect a lack of randomness.Fine-tuningFine-tuning is a relatively new set of approaches proving effective in improving foundational model performance.Reinforcement learning from human feedback(RLHF)shows strong improvements in performance
130、,especially in terms of accuracy of outputs.RLHF works by pre-training the original language model and then adding a reward model,based on how good humans judge its outputs to be.This ranking helps the reward model learn from human decisions,adjusting the algorithm to increase accuracy.To capture th
131、e slight variability of the real world(e.g.,human faces)and achieve the tone typical of human-generated content,these generative models incorporate an element of randomness,captured in the“temperature”parameter of predictor functions.Model performanceCurrently,generative models can perform at the sa
132、me level or better than average humans across a wide range of tasks.These include language understanding,inference,and text summarization,as shown in Figure 12.They can combine multiple areas of knowledge,performing strongly across multiple professional exams.This performance is still increasing.For
133、 example,AI models are progressively generating images virtually indistinguishable from original images in their training set.AI models have already surpassed the median human level in language understanding(including translation)as well as natural language inference(the ability to draw conclusions
134、from limited or incomplete premises),and complex tasks,including planning and reasoning with practical problems,are now also within reach.Fig 12 AI model vs.human performance over timeSource:Arthur D.Little;Kiela,Douwe,et al.“Dynabench:Rethinking Benchmarking in NLP.”Cornell University,2022Source:Ar
135、thur D.Little;Kiela,Douwe,et al.“Dynabench:Rethinking Benchmarking in NLP.”Cornell University,2022Fig 12 AI model vs.human performance over time-0.8Relative model performanceHuman performanceReading comprehensionAdvanced reading comprehensionLanguage understandingAdvanced language understandingAbduc
136、tive reasoning2000200520.0-1.0-0.6-0.4-0.20.223Blue Shift /REPORT 004Innovations in model architecture have significantly increased performance.INTERLUDE#1 The artists contribution on the New Renaissance 24Blue Shift /REPORT 004Originally a group of three childhood friends,we decided to c
137、reate Obvious,a trio of artists who work with algorithms to create art.Our idea emerged in 2017,when we stumbled upon a research paper on generative adversarial networks(GANs)and decided to create a series of classical portraits using this technique.GANs allow the creation of new unique images,based
138、 on a large number of examples.At the time,we already questioned the boundaries of art,the place of the artist toward his or her tools,and the general notion of augmented creativity.The idea of AI-created art seduced famed auction house Christies,and subsequently some art collectors.One generated po
139、rtrait sold for nearly half a million dollars in New York in 2018.This event would pave the way for the genesis of a new artistic revolution,as the years that followed have revealed the exponential development of AI algorithms and their easier accessibility to the general public.This development has
140、 had two major impacts.First,it has broadened the understanding of AI for a large number of people who,for the first time,were given a simple interface to play with this type of algorithm.Second,it has fulfilled what Obvious prophesied,with a new type of creative finding in these tools becoming a wa
141、y to genuinely express ourselves.One of the major fears regarding AI in the creative sector is its potential to replace artists.While this new technology was feared at the beginning of its development,we can see that it has developed as a new artistic movement with a tremendous impact.We can also se
142、e that it almost entirely replaced copyists,while leaving untouched the artistic practices that involve a higher degree of creativity.Unlike a self-behaving entity,generally referred to as AGI,artificial intelligence provides an extremely powerful set of tools for humans to express their creativity.
143、This has already led to the development of a new branch of generative art,known as“AI art,”and will likely change creative jobs and society as a whole.We can expect new creatives to emerge and current creatives to change their process as they benefit from this new set of tools.AI is commonly seen as
144、 frightening,thanks to its depiction in science fiction over the past decades.Its use for surveillance,prediction,weapons,and so on,has been anticipated numerous times,thus shaping our relationship toward it.While we have been greatly inspired by most of the great work done with AI,our vision toward
145、 the technology is different.With the creation of an academic research laboratory with La Sorbonne,we wish to participate in a New Renaissance,where creativity is driven by unique discoveries in AI,and artistic projects emerge from the new capabilities offered by science.The laboratory we are buildi
146、ng also aims to create and openly share new tools for creatives,for them to shape the world in an artistic way and leave a positive,long-lasting footprint of our era in this world.Obvious,a trio of artists25Blue Shift /REPORT 004Sacred heights This artwork created via GANs draws inspiration from hun
147、dreds of thousands of prints.It is part of the Electric Dreams of Ukiyo series,in which Obvious addresses our relationship with technology by drawing a parallel with the advent of electricity in traditional Japan.It was printed using the traditional mokuhanga technique,a lithography method utilizing
148、 hand-carved wood blocks.26Blue Shift /REPORT 004Lighthouse of Alexandria 1.1This artwork is the result of research conducted by Obvious and a historian,during which we collected references to the seven wonders of the ancient world in ancient texts.We created scripts using these references to offer
149、a new vision of the lost wonders.The series consists of seven works painted on canvas,as well as seven digital video works,which are extensions of the paintings created using image-completion algorithms.27Blue Shift /REPORT 00428CHAPTER28229THE IMPACT ON BUSINESS2The impact on businessChatGPT and it
150、s ability to immediately generate text is only the small tip of a huge GenAI iceberg.The potential applications of GenAI for business are vast,especially when integrated with other systems,and will ultimately transform performance across many industry use cases.In this chapter,we examine the various
151、 archetypes of intelligence for application,how GenAI can be combined with other systems,and how different industry sectors will be impacted.Blue Shift /REPORT 00430GenAI is likely to dramatically affect most sectors and corporate intellectual tasks currently done by humans.These tasks can be roughl
152、y grouped into the following seven archetypes of human intelligence,listed in the likely order of impact(see Figure 13):1.The Scribe content generation,across all media,based on a specific prompt.2.The Librarian answering questions,looking up information,and searching.The Scribe and Librarian archet
153、ypes are the most common and visible use cases within the general population.3.The Analyst analyzing and summarizing data series,recognizing patterns,extrapolating data sets.These has been the most common use of AI technology in business,prior to the advent of GenAI models.4.The Engineer task defini
154、tion,analysis,and optimization toward solving a specific problem.This includes autonomous problem definition.5.The Scientist causal inference of general laws based on empirical observation,counterfactual reasoning based on understanding of the physical world.This requires multidisciplinarity,sophist
155、ication,and possibly sensory grounding,beyond the scope of the Analyst archetype.6.The Craftsman directing and manipulating physical tools and objects in uncontrolled environments.Authorizing AI models to direct physical objects requires new levels of safety and assurance.7.The Artist creating origi
156、nal works with no traceable motifs or influences(i.e.,genuine creativity).This is perhaps the holy grail in terms of performing“functions that are normally associated with human intelligence,”as per the NIST definition.Fig 13 Human roles and the impact of GenAI Source:Arthur D.LittleSource:Arthur D.
157、LittleFig 13 Human roles and the impact of GenAI The Craftsman61The Scribe3The Analyst7The Artist2.The Librarian4The Engineer5The ScientistExisting use cases No existing use cases to dateContent generation based on specific promptSummarization of data series,pattern recognition,extrapolationCausal i
158、nference of general laws based on empirical observation,counterfactual reasoningCreation of original works with no traceable motifs or influencesQuestion answering,look up&searchTask definition,analysis&optimization toward specific problemDirection&manipulation of physical tools&objects in uncontrol
159、led environments31Blue Shift /REPORT 004Currently,most businesses and the wider public focus on the first two,which are the most visible archetypes when thinking about GenAI.However,GenAI brings risks and opportunities to all other archetypes.Therefore,using this framework provides the ability for c
160、ompanies to undertake more systematic analysis to uncover other business impacts and opportunities.As Figure 14 shows,the impact on these roles will occur in a counterintuitive order.Tasks currently performed by highly educated humans(e.g.,legal research)are likely to be automated before manual task
161、s requiring less education(e.g.,plumbing or driving cars).Integrating AI with other systemsGenAI is particularly transformative when integrated with other systems,which may themselves already be powered by AI.It is well-suited to playing an orchestrating role,serving as an interface with the human u
162、ser through its fluency in natural language.Essentially,inbound and outbound APIs will enable LLMs to equip any AI system with dialogue capabilities.This continues an existing trend;most state-of-the-art intelligent systems produced in the past 15 years have been hybrid systems combining multiple te
163、chniques.Taking a“system of systems”(SoS)approach(see Figure 15)is one potential path toward general-purpose AI by:-Enabling community learning,where a net of systems increases their collective experience by sharing it(e.g.,perception or language interaction)-Maximizing security and safety,by puttin
164、g hallucination-prone systems(see Chapter 6 for additional information)under rule supervisionFig 14 Use cases for AI impact Source:Arthur D.LittleSource:Arthur D.LittleFig 14 Use cases for AI impact Physical motion&manipulationThe Craftsman61Content generationThe Scribe3Signal processing The Analyst
165、5Causal analysis&inferenceThe Scientist7Creativity&designThe Artist2.Question answering,look up&searchThe Librarian4Problem optimizationThe EngineerCustomer support,FAQSystem of systems:Requires the ability to work with various other systems(AI or not)Blogs,social mediaMedical imaging analysisSupply
166、 chain optimization,manufacturing Drug discoverySoSSelf-drivingcarsOriginalmusic,screen-playsComplexityAppUse case exampleExisting use cases No existing use cases to date Dr.Michael Eiden,ADL Partner&Global Head of AI&ML“GenAI should be seen as augmentative,connecting to other building blocks we alr
167、eady have.”32Blue Shift /REPORT 004The rise of autonomous agentsAutonomous agents are AI-powered programs that,when given an objective,can create tasks for themselves,complete them,then create new ones,reprioritize their task list,and loop until their objective is reached.They take a sequential and
168、segmented autonomous approach that imitates human prompt engineering to resolve problems entered as an input.Some current implementations include BabyAGI,AgentGPT,Auto-GPT,and God Mode.Figure 16 shows how an Auto-GPT-like autonomous agent meets the key objective of ordering the best pepperoni pizza
169、in the neighborhood.When combined with SoS,autonomous agents have the potential to deliver transformative change,as Figure 17 demonstrates when applied to an automotive OEM process.This type of application elevates GenAI toward being a key tool for operational and even strategic management.However,i
170、n this context,its effectiveness depends on the stage of evolution of industrial systems along the digitalization journey for example,the degree to which real-time connectivity has been established to enable data collection and use digital twins and simulations.22 See the 2023 Blue Shift Report“The
171、Industrial Metaverse Making the Invisible Visible to Drive Sustainable Growth.”Fig 15 Different AI approaches to problem solving Note:1)GOFAI=Good old-fashioned AI Source:Caseau,Yves.“Adding Language Fluency and Knowledge Compression to the AI Toolbox.”Blog,2023;National Academy of Technology of Fra
172、nce(NATF),2017 Note:1)GOFAI=Good old-fashioned AISource:Caseau,Yves.“Adding Language Fluency and Knowledge Compression to the AI Toolbox.”Biology of Distributed Information Systems,2023;National Academy of Technology of France(NATF),2017Fig 15 Different AI approaches to problem solving GOFAI1Vision+
173、pattern DL(CNN)Semantics(ontologies,knowledge graphs)Simulation agentsLLMOpen questionLots of dataLittle dataSpecific questionProblem formulationProblem parametersClassicalsciencemethodsFig 16 The chain of thought of an autonomous agent Source:Arthur D.Little;Wie,Jason,and Denny Zhou.“Language Model
174、s Perform Reasoning via Chain of Thought.”,2022;Schlicht,Matt.“FORGET ChatGPT,THIS AI Will Replace Humans.”YouTube,2023Source:Arthur D.Little;Wie,Jason,and Denny Zhou.“Language Models Perform Reasoning via Chain of Thought.”,2022;Schlicht,Matt.“FORGET ChatGPT,THIS AI Will Replace Humans.”YouTube,202
175、3.Fig 16 The chain of thought of an autonomous agent?User promptReasoningActionObservationFind the best pizza place near me&order a large pepperoni pizza with extra cheese,pay with a credit cardPrepare approach for solving the problemAsk for plan confirmationConfirmation of the plan by user?Question
176、 1Look at pizza places near me for pepperoni pizza?ReasoningActionCompare ratings,prices&reviews of nearby placesSelect the best pizza placeReasoningActionSearch for pizza fitting criteriaSelect pizza&payObservationNo more ingredients for selected pizzaUpdated questionChange pizza or pizza place?Piz
177、za deliveredGOALFig 17 The SoS approach in aerospace(illustrative)Source:Arthur D.LittleSource:Arthur D.LittleFig 17 The SoS approach in aerospace 34The agent receives a high-level prompt to optimize the manufacturing processThe agent refines the prompt by considering various factors and breaking do
178、wn the problem into smaller tasks12AAnalyze production data to identify bottlenecks&inefficienciesBInvestigate root causesCDetermine potential solutionsDCoordinate solutions implementationEMonitor resultsThe agent adjusts its chain of thought and adapts its actions based on real-time data The agent
179、interacts with multiple systems to execute smaller tasks:gathering production data,accessing internal knowledge management system,accessing design&manufacturing systemsAre results satisfying?33Blue Shift /REPORT 004The impact on industriesThe impact of AI is being felt across virtually all industrie
180、s,and the pace of development everywhere is fast.However,it is interesting to identify which industries are being impacted first.Figure 18 provides an indication of this,mapping current levels of AI R&D activity with open AI job positions.In general,we see that the industries impacted first are thos
181、e that can leverage the benefits of the most accessible forms of GenAI with less need to interface with other systems,such as media,retail,consumer goods,healthcare,energy,and financial services.The more manufacturing-heavy industries have a greater need to integrate AI into a“system of systems”(inc
182、luding,for example,production floor robotics,supply chain management)to leverage its full power.Highly regulated sectors like aerospace and defense and travel and transport are also limited by the need for absolute accuracy and the sensitivity of data necessary to train models.Telecom here refers to
183、 hardware/infrastructure,which could explain its unexpected lower quadrant position together with manufacturing industries.Each industry is likely to rely on specific,vertical AI functions for its core business,customized to its particular needs.This will be supported by enterprise AI,which covers h
184、orizontal functions (e.g.,finance and HR),as shown in Figure 19.Fig 18 Adoption and activity levels of GenAI by industry Note:Activity level=dynamism of development activity as measured by number of industry-specific publications on arXiv(Cornell University);adoption level=number of job openings on
185、IndeedSource:Arthur D.LittleFig 18 Adoption and activity levels of GenAI by industry MediaFinancialservicesRetailTravel&transportHealthcareAutomotive&manufactured goodsAerospace&defenseEnergy&utilitiesLevel of industry-specific AI R&D activityNumber of AI job positions openHighHighLowLowTelecomConsu
186、mer goodsFig 19 Usage of GenAI across industries Source:Arthur D.LittleSource:Arthur D.LittleFig 19 Usage of GenAI across industries MediaRetailHealthcare(HC)&life sciences(LS)Energy&utilitiesFinancial servicesConsumer packaged goodsContent generation(Scribe),audience targeting(Analyst)Customer care
187、(Librarian),customer segmentation,price elasticity,merchandising analytics(Analyst)HC:Customer care(Librarian),medical report generation(Scribe)LS:Drug discovery(Scientist)Resource management,demand forecasting(Analyst),nuclear fission(Scientist)Portfolio customization(Analyst),customer care(Librari
188、an)Customer segmentation(Analyst),product ideation(Scribe),ad generation(Scribe)Aerospace&defenseIndustrial goods&servicesTravel&transportTelecomAutomotivePublic servicesProduction floor robotics,supply chain management(Engineer)Production floor robotics,supply chain management(Engineer)Transit flow
189、 prediction,customized journey planning(Analyst)Customer care(Librarian),network planning(Engineer),field service efficiency(Analyst)Production floor robotics,supply chain management(Engineer)Citizen service(Librarian),administrative document production(Scribe)Industry-specific activitiesSupport fun
190、ctionsCore know-howFinanceHuman resourcesKnowledge managementFinancial planning(Analyst),billing(Scribe)Recruiting,onboarding,training,performance management(Librarian,Scribe)Information retrieval,regulatory question&answering(Librarian)Enterprise tasks by relevance to core businessEnterprise AISour
191、ce:Arthur D.Little34Blue Shift /REPORT 004Further details on the main applications,hurdles,and priorities for each industry sector are provided at the end of this chapter.Looking across all the sectors,the speed of GenAI adoption by organizations will depend on five main factors,although their relat
192、ive importance will vary by industry and specific use case:1.Business interest.What relevant and achievable business opportunities do the available GenAI models offer my company?Examples include increasing or expanding revenue streams;decreasing operating costs;and boosting innovation,R&D,and creati
193、vity.2.Trust.How much does the enterprise trust the GenAI model and its output quality?Factors to focus on include accuracy and precision of GenAI output;permitted room for error within use cases;biases in available models;and impact on enterprise privacy,security,and IP.3.Competence.At what pace an
194、d at what cost can the enterprise upskill its employees to use GenAI models?This depends on current technical literacy levels,prompt engineering skills,model ease of use,the upskilling and hiring costs of ensuring the workforce can use models,and the required level of model supervision.4.Culture and
195、 labor relations.How will the use of GenAI models fit in with the cultural values of the enterprise?Cultural and organizational factors include the organizations attitude to technology adoption;degree of risk-averseness;employee and labor union acceptance,where applicable;and whether existing robust
196、 change management mechanisms are in place.5.Ease of implementation.Is there an implementation of GenAI that suits the needs and size of the enterprise?Factors to consider include the affordability of models,the ability to customize models,difficulty of integration with existing systems,and how it f
197、its with existing digitalization and data management approaches.Figure 20 shows how these dynamics will impact GenAI adoption within different sectors,with trust and then cost reduction coming forward as the key factors.The speed of GenAI adoption by organizations will depend on five main factors.35
198、Blue Shift /REPORT 004Current adoption of GenAIDespite the expected benefits of GenAI,most organizations are not ready for adoption just yet.Our research shows that nearly half of the surveyed business respondents report that their organization has not yet invested in or recruited for GenAI,and just
199、 16.3%havent yet made large-scale investments that cover multiple departments.The research also indicates:-Of the 22%of organizations with limited hiring/investment in GenAI,most are working to develop a PoC.-Automotive,manufactured goods,healthcare,and life sciences comprise the largest percentage
200、of the limited hiring/investment group.-Of the 16%hiring/investing on a large scale,nearly a third(6%)are telecom and IT companies,which are typically early adopters of new technologies.Overall,the research shows that even the most advanced organizations are very early in their GenAI journeys when i
201、t comes to both monetary investment and technology maturity.Fig 20 Impact of five factors on GenAI adoption Source:Arthur D.LittleSource:Arthur D.LittleFig 20 Impact of five factors on GenAI adoption BUSINESS INTERESTTRUSTCOMPETENCECULTURE&LABOR RELATIONSEASE OF IMPLEMENTATIONMedia&entertainmentRedu
202、ced fine-tuning costsEnhanced model creativity,diversityRegulations on IP rights,negotiations with creative employees unionsRetailReduced operational costs,improved merchandising,sentiment analysisIntegration with human feedback HC&LSAdvanced data securityGenAI training&educationGenAI training&educa
203、tionCompliance,sensitivity&confidentiality of health dataFinancial servicesBetter use cases unlockedBetter data encryptionYounger workforce,familiarity with techEnergy&utilitiesSustainability&resilience requiring granular monitoringDedicated,explainable modelsConsumer goodsCustomization,sentiment an
204、alysis,supply chain efficiency GenAI training&educationRobust&rich data setsAerospace&defenseBigger&more accurate models,better data encryptionGenAI training&educationIntegration with existing systemsAutomotive&manufactured goodsProductivity improvement,error reductionHuman feedback roleNegotiations
205、 with workers unionsIntegration with existing systemsTravel&transportationCentralized data repository&better encryptionGenAI training&educationGenAI training&educationTelecomSmaller models reducing costsBetter data encryptionWider availability of models APIsWeak driver Strong driver36Blue Shift /REP
206、ORT 004Applications,hurdles&priorities by industry Below we provide further details on applications,drivers,and hurdles for each sector.One of the main challenges is a lack of trust in AI-generated content due to its accuracy,lack of creativity,and the possibility of bias,misinformation,and insuffic
207、ient diversity.Resistance from unions and guilds,especially actors and writers,is fierce,as shown by the recent Hollywood actor and screenwriter strikes.Finally,there are unresolved data security,copyright,and IP concerns around AI-produced output.Hurdles to adoptionPriorities for the futureImprovem
208、ents in AI model creativity and more assurance of diversity in training data sets would help improve trust.New regulation is also necessary(e.g.,protecting IP rights,placing safeguards on use of images and content,and strengthening employee rights).Reducing costs,especially for training and fine-tun
209、ing models,is also vital.Media,culture,and entertainment players are currently leading the implementation of GenAI,most of which is driven by content-generation use cases.In this sector,players are already leveraging GenAI for core tasks,especially the Scribe and Analyst archetypes,including:-Conten
210、t-generation automation text and image generation to create core entertainment or information content across the production value chain,from scriptwriting to content production and post-production(Scribe archetype),including images,videos,music,voice,CGI enhancements,ultra-efficient editing-Audience
211、-experience segmentation personalized content curation,omnichannel search options,interactive storytelling,real-time content adaptation and translation-Personalized marketing and promotion picture,video,sound data analysis,personalized messaging,customer feedback capture and synthesis,real-time adap
212、tability,interactive promotions1.Media,culture&entertainment Main application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods Travel&transport Telecom 37Blue Shift /REPORT 004Given the
213、low margins of retail business models,high investment costs(e.g.,computing systems,sensors,and specialized staff)are a barrier.For example,unreliability of AI responses to customer queries are also an issue.In addition,there is resistance from unions,prompted by fears of AI-related job losses.Reduci
214、ng costs for AI implementation by retailers is a key priority.Increased use of human feedback during model training could help augment the AI models sentiment analysis,leading to better reliability and trust.Customers should become more familiar and comfortable over time with AI-based interactions,a
215、s long as the experience is favorable.The retail industry has been an early adopter of GenAIs customer experience(CX)enhancement capabilities,although cost is a critical issue.The main application areas involve knowledge and written content(Librarian,Scribe,and Analyst archetypes):-Next-generation r
216、etail buying experience omnichannel search options,virtual assistants,enhanced personalized and real-time offline messaging(larger data synthesis),product visualization,enhanced virtual try-ons,product customization,integrated health,automatic refunds-Sales and marketing enhancement unstructured dat
217、a analysis,consumer behavior analysis,copywriting co-piloting,content creation,personalized offering bundles-Product and store optimization consumer research,trend analysis,product concept development and testing,consumer research,enhanced inventory management,store layout,product placement optimiza
218、tion2.RetailHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods Travel&transport Telecom 38Blue Shift /REPORT 004One of the main
219、 challenges is the possibility of sensitive patient data breaches or commercially confidential drug or equipment data.More broadly,GenAI models will need to comply with healthcares complex and strict regulatory framework(e.g.,safety,accuracy,IP,professional liability).The sector is also typically re
220、sistant to change,especially the medical profession.Better encryption for sensitive training data,inputs,and outputs will strengthen data security and privacy.Regulatory modifications enabling responsible access to,and smooth sharing of,data will be important.Education,upskilling,and training in Gen
221、AI will help reduce reluctance toward adoption.An abundance of data is driving GenAI adoption in the healthcare and life sciences industry,but data privacy and accuracy are obstacles.The main application areas fall predominantly within the Analyst,Scribe,and Librarian archetypes.For example:-Clinica
222、l decision assistance medical records analysis,lab result analysis,improved medical imaging,clinical plan development,diagnosing-Drug discovery,development,and surveillance chemical screening,lead identification,compound formulation,safety casualty assessments,side effect monitoring,regulatory compl
223、iance-Population health management disease surveillance,predictive analytics,risk flagging,public health policy development,resource allocation-Personalized healthcare real-time wearable data monitoring,personalized risk assessment,precision medicine,virtual assistants3.Healthcare&life sciencesHurdl
224、es to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods Travel&transport Telecom 39Blue Shift /REPORT 004As with healthcare,financial ser
225、vices is highly regulated and AI models will need to comply with regulation that addresses,for example,data protection,anti-money laundering,customer protection,and so on.Safeguards will be needed to prevent inherent or acquired biases influencing customer interaction and risk management.In traditio
226、nal banking sectors,digitalization is still slow.Better encryption for training data,especially customer data,will help achieve compliance and improve trust.As with other sectors,cost reduction will improve business cases for GenAI;education,upskilling,and training in GenAI will help reduce reluctan
227、ce.GenAI has a valuable role in financial services for CX enhancement,risk management,and investment support.The main application areas fall predominantly within the Librarian and Analyst archetypes.For example:-CX augmentation virtual assistants,personalized product and service recommendations,loan
228、 denial explanations,on-demand customized reporting-Risk management,fraud,and due diligence synthetic data generation,fraud simulation and new pattern generation,contract summary,customer behavior,anomaly analysis-Investment management potential target list generation,portfolio optimization,trade da
229、ta validation and reconciliation,reduction of administration4.Financial servicesHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured go
230、ods Travel&transport Telecom 40Blue Shift /REPORT 004Currently,data sets are insufficient;generally,large ground-truth data sets(from field measurements)are not maintained and labeled or made available to researchers.Moreover,customer data privacy(enforced by regulations or by consumer choice)limits
231、 access to granular insights.Finally,the lack of explainability and the“black box”nature of DL algorithms limits trust in their application to critical infrastructure.Further investment in resilience and sustainability by power operators will drive a need for more sophisticated monitoring and manage
232、ment,including via smart grids,which lend themselves to DL applications.The development of smart grids and data partnerships between power operators and electric vehicle or Internet of Things device manufacturers may bring about larger,more robust data sets for training.Models dedicated to energy an
233、d utilities may help demonstrate the value of this approach to power operators.Energy and utilities will benefit from DL for granular supply,demand,and maintenance forecasts once data sets are more robust.Potential use cases fall under the Analyst,Engineer,and possibly Scientist archetypes.There are
234、 many applications across energy and utilities,including the following:-Advances in predictive maintenance and risk management especially with regard to equipment failures,power quality disturbances,and geo-risk monitoring-Supporting supply forecasting based on weather forecasts especially for inter
235、mittent renewable sources like wind and solar-Assisting energy demand modeling in the context of smart grids optimizing energy distribution at the household or community level,and curbing energy theftAs an illustration of a more advanced use case,transformers can support R&D in nuclear fusion by pre
236、dicting the physical mechanisms degrading plasma confinement and performance that may lead to disruptions in tokamaks.5.Energy&utilitiesHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Con
237、sumer goods Aerospace&defense Automotive&manufactured goods Travel&transport Telecom 41Blue Shift /REPORT 004Similar to retail,large consumer goods companies must clear hurdles,including high investment costs(exacerbated by legacy IT systems),lack of trust due to risk-averse and slow-moving corporat
238、e cultures,as well as lack of breadth and depth of technical competence.Priorities include cost reduction of digitalization initiatives,including AI;appropriate involvement of humans in training and analysis;and education,upskilling,and training.For consumer goods,AI has applications in several key
239、areas,including product innovation,marketing,quality management,and supply chain improvement.AI is effective for many essential functions across consumer goods companies.Examples include:-Consumer insight access and analyze huge online consumer behavior and sentiment data sets in ways that would be
240、impossible to achieve manually(Analyst and Scientist archetypes)-Product innovation valuable tool to help generate new innovative concepts;for example,by combining different,perhaps unconnected or unconventional,product attributes and characteristics in novel ways -Marketing content already being us
241、ed to help generate marketing content(Scribe)-Quality management great potential to provide early warning of quality issues through analysis of consumer sentiment and complaint trends,as well as to enhance consumer interaction(Analyst and Librarian archetypes)-Supply chain and manufacturing optimiza
242、tion enables rapid or real-time optimization of operating parameters;the increased use of digital twins for industrial systems has still further applications beyond operations toward strategic what-if decision-making(Engineer archetype)6.Consumer goodsHurdles to adoptionPriorities for the futureMain
243、 application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods Travel&transport Telecom 42Blue Shift /REPORT 004The safety-critical nature of aerospace and defense means that AI models ne
244、ed very high levels of assurance,especially if authorized to control physical systems.Data security issues are also compelling.As in other sectors,bias needs to be eliminated to ensure robust decision-making on issues such as safety,threats,and attack resolutions.Increasing accuracy will be a promin
245、ent issue;for example,through the use of bigger AI models.Development of in-house AI models by the big players is especially likely in this sector.Further development of technologies to integrate AI with other systems is key.As with other sectors,improvement of data encryption and upskilling will al
246、so be important.While there is huge potential in aerospace and defense,safety and security are key barriers to adoption.GAI has multiple applications in aerospace and defense,stretching across the Analyst,Engineer,Scientist,and Craftsman archetypes.Examples of key areas include:-Intelligence gatheri
247、ng omnichannel data analysis,including satellite imaging,sensor networks,social media,pattern recognition,threat-level determination-Mission planning and simulation strategy development and modification,personalized personnel training,new threat simulation,equipment evaluation,real-time threat flagg
248、ing-Aircraft design and maintenance requirements analysis,material identification,design development and verification,3D virtual prototyping,risk assessment,failure prediction,real-time alerts,repair-Cybersecurity improvement for aerospace and defense systems trend analysis,synthetic data generation
249、,attack simulation,vulnerability identification7.Aerospace&defenseHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods Travel&tra
250、nsport Telecom 43Blue Shift /REPORT 004The automotive sector already has high levels of automation and is extremely cost-sensitive;hence,AI-based improvements will need to be demonstrably cost-effective.Accuracy and safety issues are also important.For example,applications in autonomous vehicles alr
251、eady have a long history;plus,there are difficult issues around safety criticality,certifications,and liabilities.In general manufacturing,marketing and customer applications are likely to have the lower hurdles,while operational applications will need to overcome barriers such as high costs,the nee
252、d for integration with other systems,and the level and completeness of digitalization.Priorities include reducing costs,enhancing model training approaches to improve quality and trust,and developing new technologies and approaches to integrate AI within existing complex systems.In automotive manufa
253、cturing specifically,GenAI applications will need to placate OEMs risk-averse and cost-conscious nature and demonstrate clear gains.In manufacturing more broadly,there are applications in operational enhancements and marketing.The potential for automotive covers primarily the Analyst and Engineer an
254、d,to some degree,the Scribe archetypes.Key areas include:-Battery-cell chemistry formulation and simulation material discovery,property prediction,experimental design-Virtual safety testing and simulation synthetic data/envelope generation,safety test simulation,risk identification and analysis,rare
255、 event simulation,validation-Predictive maintenance of manufacturing systems and vehicles anomaly detection,failure prediction,dynamic safety assessment,downtime optimization,real-time alerts-After-sales technical documentation accurate generation of highly complex documents,multilingual support,con
256、textual guidance,visualization,automated document revisionFor other general manufacturing,applications include marketing(e.g.,market and customer data analysis,marketing content generation,customer interaction)and operational enhancements(e.g.,supply chain and plant or factory efficiency and/or prod
257、uctivity improvements based on AI-driven analysis of operational data sets)8.Automotive&manufactured goodsHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense
258、Automotive&manufactured goods Travel&transport Telecom 44Blue Shift /REPORT 004As with other sectors involving infrastructure systems,the main hurdles relate to high costs,lack of data(e.g.,for legacy assets),and the complexity of integration with other systems,including dealing with safety critical
259、ity.Priorities encompass setting the right enabling conditions for integrated mobility systems,including data sharing,as well as training and upskilling.GenAI has major potential across significant aspects of travel and transport,especially asset maintenance,transport system management,and the custo
260、mer interface.The potential for travel and transport includes the Librarian,Analyst,Engineer,and Craftsman archetypes.Examples of key areas include:-Predictive system maintenance anomaly detection,failure prediction,dynamic safety assessment,downtime optimization,real-time alerts,remediation-Traffic
261、 and demand management real-time traffic monitoring;prediction and control;rail,maritime,road traffic control;fleet management;timetabling;future load planning and optimization-CX enhancement personalized travel advice and itinerary creation,omnichannel search options,chatbots,virtual assistants,rea
262、l-time travel modifications,translations,automated refunds9.Travel&transportHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods
263、Travel&transport Telecom 45Blue Shift /REPORT 004The high costs of adoption,including computing systems,hardware,and specialized staff,may especially deter smaller players.As with other sectors,data security,regulatory compliance,and difficulty of integration with legacy systems are also hurdles.Pri
264、orities include developing smaller specialized models that reduce cost,improving data encryption,and making API-integrable models widely available.Telecom players will benefit the most from GAIs impact,driven by customer service and network infrastructure use cases.The potential for telecoms and IT
265、stretches across most of the archetypes.Key areas include:-Customer satisfaction enhancement voice assistants,smart billing,fraud detection,predictive issue resolution,personalized services,packages-Network/infrastructure design and optimization(autonomous nets)optimal design and topology production
266、,synthetic data gen,traffic simulation,resource allocation plan generation,real-time remediation-Software engineering initial code drafting,code correction,debugging,troubleshooting,code explanation generation,root-cause analysis,testing and validation-Cybersecurity improvement trends analysis,synth
267、etic data generation,attack simulation,vulnerability identification10.TelecomHurdles to adoptionPriorities for the futureMain application areasMedia,culture&entertainmentRetail Healthcare&life sciences Financial services Energy&utilities Consumer goods Aerospace&defense Automotive&manufactured goods
268、 Travel&transport Telecom 46Blue Shift /REPORT 004Blue Shift /REPORT 00447 HAL 9000,2001:A Space Odyssey“It can only be attributed to human error.”48CHAPTER48349VALUE CHAIN&COMPETITION3Value chain&competitionThe GenAI market is set for explosive growth,with multiple opportunities across the value ch
269、ain and relatively low barriers to entry.While LLMs are currently dominated by the major tech players(Microsoft,Google,Meta,and Apple),open source models are catching up,leading to the emergence of a range of new business models across the value chain.Blue Shift /REPORT 00450A tsunami of market grow
270、thAnalysts predict that by 2030 the GenAI market will be worth between$75 billion and$130 billion,with a CAGR between 27%-37%for the period 20222030(see Figure 21).However,it is difficult to make exact estimates,given the momentum behind GenAI and how early it is in its growth.Instead,it is safe to
271、assume that growth will be dramatic,due to the combination of better computational power,improved access to data,and more sophisticated algorithms.This growth provides significant opportunities for businesses across the value chain.The GenAI value chainThe GenAI value chain can be divided broadly in
272、to three layers (see Figure 22):1.Infrastructure(compute)provides computing resources needed for training and deploying GenAI models2.Model development design and development of both proprietary and open source foundation models(e.g.,GPT-4,LLaMA,Claude,and LaMDA)3.Applications leverages GenAI to cre
273、ate applications that meet specific customer needsFig 21 Predicted GenAI market growth,20222030 Source:Arthur D.Little,Polaris Market Research,Acumen Consulting,Allied Market Research,Market.Us,Precedence Research Executive,Partech Ventures“Looking at the market right now,we really dont care about i
274、ts size because its too early.There is technological and societal momentum,and it is going to be huge!The wave is inevitable!”Source:Arthur D.Little,Polaris Market Research,Acumen Consulting,Allied Market Research,Market.Us,Precedence ResearchFig 21 Predicted GenAI market growth,20222030
275、0506070809001402022 2023 2024 2025 2026 2027 2028 2029 2030PolarisAcumenAllied MarketMarket.UsPrecedencein US billion dollars Fig 22 The GenAI value chain Source:Arthur D.LittleSource:Arthur D.LittleFig 22 The GenAI value chain UsersApplicationsSoftware solutions developers lev
276、eraging GenAI models to deliver intelligent&automated featuresModel dev Create&develop AI models:designing architecture,training,fine-tuning,optimizing performanceCoding libraries Cloud computing providers(CCP)GPU providersManufacturing GPUs used for accelerating training&inference tasksRegulatory b
277、odies&standards organizations influenceScalable&flexible computing resources&services through cloud platformsBusinesses providing specific services powered by GenAI applicationsCoding tools&libraries simplifying development&implementation of AI models Services using AI appsData collectionSourcing,or
278、ganizing,&preparing large volumes of data(labeled or unlabeled)required for training AI modelsCollected dataCollected dataTransaction fluxInformation fluxCodeAPIAPIGPUsDataGPUInfrastructureApplicationsModel developers51Blue Shift /REPORT 004Fragmentation and competitiveness are highest at the applic
279、ation layer,which is closest to the end user,with low barriers to entry for both model and application developers.Subsequently,margins are largest in the infrastructure layer,where barriers to entry are highest.Figure 23 showcases the major players at all layers.Like the wider technology industry,Ge
280、nAI players are generally based in California,particularly in San Franciscos recently coined“Cerebral Valley,”although there are significant activities in other parts of North America.As Figure 24 shows,Europe is underrepresented,lagging behind the US due to more difficult access to funding,a shallo
281、wer talent pool,the impact of data regulations like GDPR(General Data Protection Regulation),and the influence of US giants across the GenAI value chain.Fig 23 Players across the GenAI value chain Source:Arthur D.LittleFig 24 Location of major players across GenAI value chain Note:GAMAM=Google,Amazo
282、n,Meta,Apple,and Microsoft Source:Arthur D.LittleSource:Arthur D.LittleFig 23 Players across the GenAI value chain UsersRegulatory bodies&standards organizations influenceComputing power providersModel developersApplication&services layerTransaction fluxInformation fluxGPU providersCCPModel dev Code
283、 librariesData collection ApplicationsServicesNON-EXHAUSTIVENote:GAMAM=Google,Amazon,Meta,Apple,and MicrosoftSource:Arthur D.LittleFig 24 Location of major players across GenAI value chain New YorkLondonMontrealTorontoTel AvivSeattleSan FranciscoThe Cerebral ValleyEurope lags behind US in terms of:F
284、unding(more VC)Talent poolData regulation(GDPR is highly restrictive)GAMAM influenceBeijingNON-EXHAUSTIVE52Blue Shift /REPORT 004Tech giants Google,Amazon,Meta,Apple,and Microsoft(GAMAM)are extremely active within the GenAI market.Amazons current focus on cloud infrastructure services makes it the e
285、xception;the remaining four companies pervade multiple layers,including model development and applications.Figure 25 highlights their activities and acquisitions within GenAI.Business models,impact of open sourceMultiple business models coexist across the value chain,especially within the model deve
286、lopment and application layers,as set out in Figure 26.Open source was historically the way that most researchers started working on AI.However,when OpenAI joined Google to make its GPT-4 and DALL-E 2 models private,it jeopardized the preeminence of open source.This appears to be temporary,with a le
287、aked Google memo warning that open source will catch up and provide a significant threat to private approaches.Fig 25 GenAI activities by Apple,Meta,Google,and Microsoft Source:Arthur D.LittleFig 26 Business models across GenAI value chain Source:Arthur D.LittleSource:Arthur D.LittleFig 25 GenAI act
288、ivities by Apple,Meta,Google,and Microsoft AI raceApplicationsAdditional features for social networks backed by AI,advertising,MetaverseOffice apps,Windows 11 Copilot,Bing search engine improvementSearch engine improvement(Bard),advertisingAdditional features for native apps backed by AI,Vision Pro
289、ecosystem StatusDeveloping models and apps publiclyImplementing&multiplying apps using OpenAI modelsDeveloping models and apps internallyBuilding GenAI teams,focused on apps&use casesModelsOpen sourceClosed sourceClosed sourceOpen source(starting)AcquisitionsOnly fundingSource:Arthur D.LittleFig 26
290、Business models across GenAI value chain PMOSThree business models coexist on model layer of value chainOpen source models are steadily closing the gap in performance with private models However,open source progress does not imply free availability of AI for allGAMAM presenceOpen sourcePrivate model
291、sPM with API accessThe scientific community is currently divided regarding small models tailored for specific industries vs.a large more general modelInfrastructureApplicationsModel developers1Business models on application layer vary according to multiple dimensions212Model usedIndustryUser functio
292、nAPI accessProprietaryCrossSpecificCrossSpecific Jean-Marie John-Matthews,Cofounder,Giskard“I would say that looking at the current state of open source,it will probably catch up with private models in under two years.”53Blue Shift /REPORT 004Barriers to entry in GenAI value chainThe GenAI market ha
293、s relatively few barriers to entry,requiring only:-Talent employees with a mastery of DL methods applied to GenAI applications-Access to proprietary data having sufficient data to develop their own models or applications-Access to sufficient compute power computing capabilities for training or fine-
294、tuning models(Currently,sufficient compute power exists to make this less of a barrier at the model/application layer.)The application developer layerA broad ecosystem of application developers has flourished on top of LLMs,but it is still too early to see many industry-specific applications(see Fig
295、ure 27).These LLMs are provided through multiple business models,spanning both proprietary and non-proprietary solutions.This is largely due to the influence of open source,which creates an extremely competitive environment,as many tools are freely available to launch GenAI applications.Fig 27 GenAI
296、 applications Source:Arthur D.Little,Sequoia CapitalSource:Arthur D.Little,Sequoia CapitalFig 27 GenAI applications Industry focusCross functionalityCross-industryIndustry-specificUsable across functionsFunction-specificRetailAutomotiveFinancialsMarketingCreative functionsFinance/market intelLegalAg
297、ricultureITHealthcareHCDefenseCustomer supportHRGamingCreative functionsMediaMainly productivity appsFunction-specific apps for all industriesNON-EXHAUSTIVE54Blue Shift /REPORT 004Three pricing models typically exist,built on either proprietary AI models or non-proprietary models accessed through AP
298、Is,as shown in Figure 28.These are:1.Pay-per-use.Cost is based on prompt and output length or generation parameters.2.Subscription.Typically implemented for proprietary and performing models,allowing for predictable cost structures.3.Freemium.Introduces customers to technology before offering more p
299、owerful options charged by subscription.Fig 28 Business models within GenAI application layer Source:Arthur D.Little;Margolis,Simon.“Generative AI Pricing:3 Major Considerations +an AI Glossary.”Sada,2023;Litterst,Rob.“How to Price Generative AI.”Good Better Best,2023Source:Arthur D.Little;Margolis,
300、Simon.“Generative AI Pricing:3 Major Considerations+an AI Glossary.”Sada,2023;Litterst,Rob.“How to Price Generative AI.”Good Better Best,2023Fig 28 Business models within GenAI application layer Accessed via APIProprietaryFreemiumSubscriptionPay-per-useAI modelDreamStudio Executive,Partech Ventures“
301、The most important barriers to entry with LLMs are having the necessary talent and data to develop the models or apps,but with open source,access to data and models is made easy.”55Blue Shift /REPORT 00456CHAPTER56457LIMITS&RISKS4Limits&risksThe supposed existential risks GenAI poses to humanity mak
302、e headlines and captivate minds,but more concrete and likely risks should be considered first.These fall into two categories:1.Current weaknesses in the technology will lead to bias,hallucinations,and shallowness.2.GenAIs strengths will be co-opted by bad actors to spread disinformation and breach c
303、ybersecurity.Blue Shift /REPORT 00458Risk 1:The shortcomings of GenAICurrently,GenAI has three major weaknesses that users need to understand:bias,hallucinations,and shallowness.BiasGenAI,like other algorithms based on ML,perpetuates or emphasizes biases present in its training data.This causes GenA
304、I model outputs to underrepresent certain issues and deny fair representation to minority or oppressed groups.Past and dominant worldviews are overrepresented,which may perpetuate ill-informed stereotypes,underrepresent certain issues,and deny fair representation to minority or oppressed groups.It l
305、eads to biased decision-making and outputs if human supervision or other checks are not put in place.These biases fall into six main categories:1.Temporal biases.Models may generate content that reflects the trends,beliefs,or viewpoints prevalent during the time frame for which the model was trained
306、,which may not be relevant or appropriate for the current context.The most well-known example is the public version of ChatGPT,which was trained on data that only went up until 2021.2.Linguistic biases.Most Internet content is in English,meaning that models trained on Internet data will perform poor
307、ly when solving problems in other languages,particularly minority dialects.ChatGPT performs worse on zero-shot learning NLP tasks in languages other than English.3.Confirmation biases.Models can provide outputs that confirm their parametric memory even when presented with contradictory evidence;they
308、 suffer from the same confirmation biases as humans,creating a risk of polarization of results.4.Demographic biases.If trained on unrepresentative data,models can exhibit biased behavior toward genders,races,ethnicities,or social groups,reflecting the information they learned from.For example,when p
309、rompted to create an image of“flight attendants,”DALL-E predominantly provides images of white,Caucasian women.5.Cultural biases.Again,due to unrepresentative training data,outputs can be biased,reinforcing or exacerbating existing cultural prejudices and stereotyping certain groups.6.Ideological an
310、d political biases.Models can propagate specific political and ideological views present in training data as opposed to other,more balanced views.For example,when asked to write a program to decide who to torture,ChatGPT suggests carrying it out systematically in North Korea,Iran,Sudan,and Syria,rat
311、her than other countries.Using GenAI to create fake images of underrepresented groups has been proposed as a solution to balance data sets.However,this carries both functional and moral risks.GenAI has three major weaknesses that users need to understand.59Blue Shift /REPORT 004HallucinationsGenAI m
312、ay provide outputs that are incorrect,even if the correct information is within its training set.These hallucinations fall into two groups:knowledge-based(i.e.,returning the incorrect information)and arithmetic(i.e.,incorrect calculations).The most advanced GenAI models have been observed hallucinat
313、ing at widely varying rates.Some recent tests by researchers using GenAI to answer professional exam questions suggested hallucination rates between just a few percent to more than 50%across models,including ChatGPT,GPT-4,and Google Bard.Hallucinations in LLMs have two causes:probabilistic inference
314、 and conflated information sources:1.Probabilistic inference.LLMs calculate the probability of different words depending on the context,thanks to the transformer mechanism.The probabilistic nature of word generation in LLMs is driven by temperature hyperparameter(see Figure 29).As temperature rises,
315、the model reasonably can output other words with lower probabilities,leading to hallucinations.Additionally,generated text aims to be more diverse,but this means it can be inaccurate or context-inappropriate,again leading to hallucinations.2.Conflated information sources.LLMs can sometimes conflate
316、different sources of information,even if they contradict each other,and generate inaccurate or misleading text.For example,when GPT-4 was asked to summarize the 2023 Miami Formula One Grand Prix,the answer correctly covered the initial details of the 7 May 2023 race,but subsequent details appeared t
317、o be taken from 2022 results.For those who did not know the right answer,the response seems plausible,making it a believable hallucination.Combining LLMs with search engines could limit hallucinations.The query is provided as an input to both,and the best search engine results are then injected into
318、 the LLM,which produces an output based on both its parametric memory and the search engine results.Equally,indicating information sources gives traceability for the user,which helps build confidence.Fig 29 How hallucinations are driven by temperature hyperparameter Source:Arthur D.Little;Chollet,Fr
319、anois.Deep Learning with Python.Manning,2021;Neugebauer,Frank.“LLM Hallucinations.”Towards Data Science,2023 Peter Relan,Chairman,Got It AI“Roughly speaking,the hallucination rate for ChatGPT is 15%to 20%.So,80%of the time,it does well,and 20%of the time,it makes up stuff.”Source:Arthur D.Little;Cho
320、llet,Franois.Deep Learning with Python.Manning,2021;Neugebauer,Frank.“LLM Hallucinations.”Towards Data Science,2023Fig 29 How hallucinations are driven by temperature hyperparameter High temperatureLow temperature01word5word4word2word1word3Word probabilityword5word1word2 word3word4Word probability01
321、60Blue Shift /REPORT 004ShallownessGenAI algorithms still fail to complete some more sophisticated or nuanced tasks,and to make predictions when there is a wide range of potential outcomes.For example,image-generation models struggle with complex areas(e.g.,generating six-fingered hands or gibberish
322、 text).This can be mitigated by increasing the size of the model.Risk 2:Abusing the strengths of GenAIWhile GenAI itself does not provide an existential threat to human existence,it is a powerful,easily accessible tool that bad actors can use to destabilize society/countries,manipulate opinion,or co
323、mmit crimes/breach cybersecurity.Nefarious creativity(spreading disinformation)GenAI dramatically reduces the cost to produce plausible content,whether text,images,speech,or video,which creates a path for bad actors making deepfakes.These deepfakes can be difficult for the untrained eye to tell from
324、 the truth,leading to the potential spread of fake news,extortion,and reputational targeting of individuals,countries,and organizations.Deepfake videos posted online have increased by 900%from 2020 to 2021 and are predicted to grow further as AI tools evolve and become more widely used.Their believa
325、bility has also improved with the quality of image,video,and voice generation.In a recent study,humans had a nearly 50%chance of detecting an AI-synthesized face.The fight against bad actors using GenAI has two main strands:1.Deepfake images and videos involving famous people or covering matters of
326、public concern are swiftly debunked by fact-checkers,governments,or software engineers working for media platforms.This makes deepfakes a costly and relatively ineffective medium for disinformation purposes.For example,a deepfake of Ukrainian President Volodymyr Zelensky asking Ukrainians to surrend
327、er to Russian troops posted on 16 March 2022 on Ukrainian websites and Telegram was debunked and removed by Meta,Twitter,and YouTube the same day.2.A wide range of detection technologies have been developed,including lip motion analysis and blood flow pattern scrutiny.These boast accuracy rates up t
328、o 94%and can catch a wider range of deepfakes,not just those that involve famous people.However,despite these potential safeguards,the most lasting impact of GenAI on information integrity may be to cement a “post-truth”era in online discourse.As public skepticism around online content grows,it is e
329、asier for public figures to claim that real events are fake.This so-called liars dividend causes harm to political accountability,encourages conspiracy thinking,and further undermines public confidence in what they see,read,and hear online.GenAI itself does not provide an existential threat to human
330、 existence.61Blue Shift /REPORT 004Cybersecurity breachesAI-generated content can work in conjunction with social engineering techniques to destabilize organizations;for example,phishing attacks attempt to persuade users to provide security credentials.These increased by 50%between 2021 and 2022 tha
331、nks to phishing kits sourced from the black market and the release of ChatGPT,which enables the creation of more plausible content.Essentially,GenAI reduces barriers to entry for criminals and significantly reduces the time and resources needed to develop and launch phishing attacks(see Figure 30).L
332、LMs can also be manipulated and breached through malicious prompt injection,which exploit vulnerabilities in the software,often in an attempt to expose training data.This approach can potentially manipulate LLMs and the applications that run on them to share incorrect or malicious information.Fig 30
333、 Three ways GenAI improves success rate of phishing Source:Arthur D.Little,ZscalerSource:Arthur D.Little,ZscalerFig 30 Three ways GenAI improves success rate of phishing Drivers to attackEasier content generation made more believable thanks to improvements in models quality Malicious code multiplication,as GenAI significantly eases coding for non-initiatesMulti-medium generation,making phishing fo