《理特咨詢（ADL）：2025人工智能驅動的研究、開發與創新突破的新時代研究報告（英文版）（80頁）.pdf》由會員分享，可在線閱讀，更多相關《理特咨詢（ADL）：2025人工智能驅動的研究、開發與創新突破的新時代研究報告（英文版）（80頁）.pdf（80頁珍藏版）》請在三個皮匠報告上搜索。

1、 EUREKA!ON STEROIDSAI-driven research,development,and innovation2025“AI will accelerate scientific research far more than we can imagine.”Jolle Barral,Director of Fundamental Research in AI,Google DeepMindBlue Shift /REPORT 0073EUREKA!ON STEROIDSAI-driven research,development,and innovationAuthorsDr

2、.Albert Meige,Director of Blue Shift,Arthur D.LittleZoe Huczok,Project Leader of Blue Shift,Arthur D.LittleArnaud Siraudin,Associate Director,Arthur D.LittleDonatello Fleck,Business Analyst,Arthur D.LittleGeoffroy Barruel,Consultant,Arthur D.LittleContributorsPierre Blouet,Research,Development,and I

3、nnovation Director,GRTgazCarole Caranta,Deputy Director General of Science and Innovation,INRAEPaul-Jol Derian,Group VP Innovation and Sustainable Development,AvrilPhilippe Mauguin,CEO,INRAEJean-Luc Moullet,Chief of Staff,French Minister of Higher Education and ResearchChristophe Perthuisot,Chief Re

4、search&Innovation Officer,Mot HennessyAntoine Petit,President&CEO,CNRSRick Eagar,Partner Emeritus,Arthur D.LittleExpert-in-residenceAnne Bouverot,French Presidents Special Envoy on AICONTENT-CONTENT-CONTENT-CONTENT-5Executive summary 6Preamble 12Chapter 1.The potential of AI in R&D&I 16Chapter 2.How

5、 to ensure success 24Interlude:Focus on data!36Chapter 3.Tools&providers 38Chapter 4.Navigating the future 44Chapter 5.Strategic actions 646Blue Shift /REPORT 007Executive summary Although AI has been used in specific research,development,and innovation(R&D&I)applications for at least a decade,its b

6、een two years since the recent acceleration began,initiated by the availability of more powerful generative AI(GenAI)and large language models(LLMs).While there is a glut of information on potential applications,widespread integration of AI into R&D&I processes is still relatively immature.Applying

7、AI to many R&D&I use cases poses significant challenges,especially where outcomes need to be error-free,as well as uncertainties in how AI will evolve regarding technology,economics,regulation,and societal acceptability.This in-depth study was conducted by Arthur D.Littles(ADLs)Blue Shift in partner

8、ship with five major leading public and private sector organizations already using AI in their R&D&I efforts:LVMH,Avril,GRTgaz,the French National Centre for Scientific Research(CNRS),and the French National Research Institute for Agriculture,Food,and the Environment(INRAE).The study explored the cu

9、rrent state of AI in R&D&I,the challenges and best practices,the landscape of solution providers,and future scenarios.We gathered evidence through 40+interviews from AI providers,independent AI experts,and current best-in-class users of AI in R&D&I,as well as a survey with over 200 responses from pr

10、ivate companies and public institutions that examined AI maturity,contributions,benefits,and barriers.Choosing among AI systems and other approaches depends on the type and amount of data available and the nature of the question.7Blue Shift /REPORT 007THE POTENTIAL OF AI IN R&D&IChapter 1:AI augment

11、s researchers capabilities across all steps of the R&D&I process through various roles,helping to solve intractable problems and make decisions.No blanket model exists;data availability and problem type determine the best method.Most often,AI models are embedded in a systems of systems.-Benefits abo

12、und for AI.AI augments researchers capabilities,rather than replacing them,as part of a people-centric R&D&I effort.It helps solve intractable problems that researchers couldnt tackle before.It already acts as a knowledge manager,hypothesis generator,and assistant.The planner/thinker archetype,in wh

13、ich AI helps make decisions,is rapidly emerging.-AI-based models support use cases at every step of R&D&I process.These range from technology and market intelligence to innovation strategy,ideation,portfolio and project management,IP management,ecosystem management,knowledge management,and new produ

14、ct/service launch and deployment.-There is no blanket model for R&D&I tasks.Choosing among AI systems and other approaches depends on the type and amount of data available(e.g.,more/less)and the nature of the question(e.g.,open/closed).LLMs often play an orchestrating role in interfacing with or con

15、trolling other systems.AI is not always the answer.Classical science approaches,including traditional regression methods,may perform better on some problems.Most often,AI models are embedded in a systems of systems that also include human intervention.HOW TO ENSURE SUCCESSChapter 2:Ensuring success

16、in AI implementation for R&D&I requires agile methodologies,robust data foundations,strategic prioritization,analytical tradeoffs,scarce data science talent management,IT alignment,rapid benefit demonstration,and continuous monitoring.-Agile methodologies.Agile methodologies that move fast and itera

17、tively are preferable for AI project development,given the speed at which technology evolves.Such approaches ensure that some benefits can be obtained early,even if“perfect”solutions are still some way off.-Robust data foundations.Data collection hygiene,storage,security,and governance are central t

18、o realizing AI benefits.New techniques for processing poorly structured or smaller datasets are becoming more important.Ensuring wide data accessibility,cross-organizational collaboration,and effective data governance is also fundamental.-Strategic prioritization.Organizations must choose strategica

19、lly between making,buying off-the-shelf,or fine-tuning AI models.Most core R&D&I problems lend themselves to fine-tuning existing open source models,whether LLMs,generative adversarial networks(GANs),diffusion models,or reinforcement learning(RL).-Analytical tradeoffs.Tradeoffs must be carefully con

20、sidered in proof-of-concept(POC)development,including acquiring versus synthesizing data,optimizing for precision versus recall,and underfitting versus overfitting data.-Scarce data science talent management.The right organizational solution for accessing expert data science resources depends on nee

21、ds different pros and cons come with using external resources,training internal experts,creating a central AI service center,or embedding data scientists in R&D teams.-Alignment with IT.R&D&I functions need to align with IT departments to address security and compliance requirements while maintainin

22、g the speed needed for experimentation.-Rapid benefit demonstration.Prioritizing AI use cases and demonstrating benefits quickly helps prevent POCs from stalling.-Continuous monitoring and improvement.These are especially important for experimental AI models,as their performance can change over time

23、.TOOLS&PROVIDERSChapter 3:The value chain for AI in R&D&I heavily relies on major open source models,but smaller players also form a key part of the ecosystem.Applications tailored for every part of the R&D&I process exist,as do start-ups targeting vertical-specific problems.Hosting providers also o

24、ffer inference as a service.-Key role of open source.The value chain for AI in R&D&I can be divided into three layers:(1)infrastructure(compute),(2)model development,and(3)R&D&I applications.Open source models are the backbone across the whole chain,developed and trained by major players such as Met

25、a(Llama),Microsoft(Phi),and Nvidia(NVLM,TensorFlow,StyleGAN).Smaller players like Mistral and Cohere(Aya 23)and academic institutions such as CNRS and GENCI(BLOOM)also contribute significantly to the open source ecosystem.Collaboration is fostered through forums that encourage co-creation and sharin

26、g of fine-tuning tools,such as Hugging Face.-Start-ups and players.AI use cases exist for every building block of the R&D&I process,from strategy and intelligence to ideation,portfolio/project management,IP management,knowledge management,ecosystem management,and deployment.The applications market a

27、lso includes vertical-specific start-ups targeting scientific,research,and innovation problems,especially in life sciences.-Inference as a service a promising service.Various hosting providers also offer an inference-as-a-service model,which consists of hosting the compute power needed for inference

28、s run by the model(e.g.,each LLM query)in the cloud to help customers avoid high computational costs.NAVIGATING THE FUTUREChapter 4:How AI in R&D&I will evolve depends on the outcomes of three main factors:performance,trust,and affordability.These lead to six plausible future scenarios on a spectrum

29、 between AI transforming every aspect of R&D&I at one end and being used only in selective,low-risk use cases at the other.We identified 16 trends shaping the future,divided into three main factors:1.Performance whether AI will meet the high bar necessary for many R&D&I problems2.Trust the extent to

30、 which researchers,developers,customers,and the public will trust and accept AI-generated outputs3.Affordability how far AI implementation will be constrained by costs,skills,resources,and environmental impactsThese factors lead to six plausible scenarios across two ends:1.Cheap&Nasty.At one end of

31、the spectrum is the Cheap&Nasty scenario reflecting low performance and trust but high affordability.In this scenario,AI is only used in select low-risk use cases with strict vetting,curtailing productivity gains.2.Blockbuster.At the other end of the spectrum is the Blockbuster scenario,which reflec

32、ts high performance,trust,and affordability.In between are scenarios reflecting other combinations,each with different consequences regarding day-to-day R&D&I work,organizational evolution,and winners and losers.Recognizing these scenarios is important for R&D&I organizations to chart a way forward.

33、8Blue Shift /REPORT 007STRATEGIC ACTIONSChapter 5:We recommend six no-regret moves for organizations regardless of the six future scenarios.These comprise mutualizing compute power,encouraging data sharing,managing AI talent,training the workforce in AI fundamentals,resetting data and AI governance

34、approaches,and improving output controls.Beyond these,organizations should take measured strategic bets aligned with corporate objectives.-Six no-regret moves that R&D&I organizations should take to shape up for the AI future,irrespective of how the scenarios develop:1.Manage and empower talent.Corr

35、ectly access external resources and build in-house capabilities.Given that low-code/no-code(LCNC)AI solutions are becoming increasingly prevalent,data engineering profiles may become more important than data science profiles.2.Control AI-generated content.Scale-up quality and IP control systems for

36、AI-generated content and data.3.Build up data sharing and collaboration.Pursue data-sharing initiatives,looking for mutual benefits and ways to overcome competitive concerns.4.Train for the long run.Continuously deliver AI training to as broad an audience as possible,beyond immediate users.This shou

37、ld touch on AI technical fundamentals,functional capabilities,implementation requirements,and potential risks.5.Rethink organization and governance beyond IT.Set up a suitable governance system for AI.This means allocating clear AI and data governance accountabilities and responsibilities to key ind

38、ividuals within the overall enterprise and digital teams beyond the IT function.For critical AI use cases,centralization of governance may be desirable to overcome coordination processes,reporting directly to executive leadership.6.Mutualize compute resources.Look for opportunities to increase the a

39、ffordability of compute power through mutualization with partners.-Strategic bets to consider.In addition to no-regret moves,R&D&I organizations should consider the scope,costs,and benefits of AI use cases;prioritize efforts;and revisit the overall innovation strategy to suit an AI future.Portfolio

40、logic should be applied to ensure a suitable balance of:-Hedging AI moves to ensure rapid response in case of future disruption-Speculative high-risk/high-reward AI opportunities-Shorting in future AI areas that could be especially susceptible to poor performance,low levels of trust,or excessive cos

41、ts.9Blue Shift /REPORT 0072AI will power all steps of R&D&I including those involving creativityAIs impact is not just hype its impacting both the productivity and the creativity dimensions of R&D&I.Companies that quietly leverage AI using general-purpose LLMs and smaller specialized models are alre

42、ady seeing 10 x productivity gains in some situations.1AI as the orchestrator,not the solo playerAI should serve as a coordinator between diverse digital tools,such as simulation,good old-fashioned AI(GOFAI),GenAI,graphs,rules and heuristics,and Bayesian networks,while also keeping the human in the

43、loop.Automated agents will empower researchers to run entire workflows autonomously,speeding up discovery.Maturity gap in AI adoption not a barrier but an opportunityMost R&D&I organizations are still new to AI,with many researchers unaware of its current and future impact.3Focus on solving problems

44、,not just deploying AIThe focus shouldnt be on flashy AI tools but on using AI to solve specific,high-impact problems.Defining the right problems will ensure that AI is a tool for innovation rather than just a trend.4Leverage LLMs for productivity use casesFine-tuned LLMs deliver high value even tho

45、ugh we dont fully understand their workings at scale.LLMs are particularly interesting for productivity use cases.While cross-silo data integration could unlock even greater potential,methods such as Low-Rank Adaptation(LORA)make fine-tuning affordable and effective today.510117Smaller models for bi

46、gger creativebreakthroughsSmaller and more specialized AI models or other approaches,such as Bayesian networks,will increasingly excel in solving complex R&D&I problems.These tailored models are more effective in certain areas.6Data is the game changer,not algorithmsData management will be the diffe

47、rentiator in the AI-driven future as algorithmsbecome commoditized.Centralized,structured,and cleaned data will be the foundation for building competitive R&D&I systems.Preparing data for the first POC may take time(up to 18-24 months),but it will speed up with each iteration.Trust is everything bui

48、ld it carefully and maintain it diligentlyBuilding trust in AI systems is critical and fragile.In R&D&I,where the stakes are high and outputs arent immediately tangible,ensuring transparent processes and human oversight is essential to avoid setbacks in AI adoption.8AI talent shortage the race to up

49、skill R&D teamsThe supply of AI talent will lag behind demand until 2030,making upskilling of existing R&D teams crucial.Organizations that invest earlyin training their talent will avoid falling behind in the AI race.9Inference as a service a paradigm shift for product developmentLike cloud computi

50、ng transformed IT infrastructure,inference as a service willrevolutionize how companies develop and scale AI-driven products.This model will be key in democratizing AI and fostering new business models.10PreambleOctober 2018:Christies auction house sells a work created by the French artist collectiv

51、e Obvious in collaboration with AI for nearly half a million dollars.December 2023:A group of researchers reports in Nature that they have solved a previously unsolved mathematics problem by collaborating with GenAI.June 2024:Researchers at the University of Pennsylvania leverage machine learning(ML

52、)to analyze microbial dark matter,uncovering nearly 1 million potential antibiotic compounds.Published in Cell,this breakthrough accelerates the discovery of new antibiotics,with dozens showing activity against antibiotic-resistant bacteria,compressing years of research into mere hours using AI.A se

53、ries of ripples in the pond of creativity.Blue Shift /REPORT 00712As businesses and decision-makers,we believe its only natural to ask whether AI has truly become creative.Or perhaps more importantly,can it help us humans become more creative?This question is crucial,as AI has been primarily focused

54、 on improving performance and productivity over the past 10-15 years.If AI can indeed foster creativity,it could revolutionize problem solving across various sectors,including R&D&I.The question becomes even more relevant following a landmark event in October 2024:two Nobel Prizes were awarded for d

55、iscoveries related to ML.In chemistry,Demis Hassabis and John Jumper from DeepMind received the Nobel Prize for AlphaFold2,a model capable of predicting the structure of 200 million proteins.Since its release in 2020,over 2 million researchers worldwide have already used this deep learning model.Thi

56、s breakthrough has revolutionized drug discovery,demonstrating how AI can directly fuel innovation.In physics,John Hopfield and Geoffrey Hinton were awarded the Nobel Prize for their contributions to AI,specifically ML.They applied statistical physics principles to develop foundational neural networ

57、k models,including the Hopfield network and the Boltzmann machine.Their work laid the groundwork for modern ML,allowing AI systems to learn and recognize patterns from complex data,revolutionizing the field and demonstrating how science fuels AI.In this watershed moment for science and research,one

58、might wonder:how do professions defined by human intelligence adapt to the rise of artificial intelligence?What role does AI play in the“Eureka!”moment and the countless tasks that define R&D&I?What specific challenges do organizations face in areas such as problem definition,data availability,secur

59、ity,system interpretability,and costs and what are the best practices to address them?Furthermore,how will AIs uncertain future affect the landscape of R&D&I?If AI can foster creativity,it could revolutionize problem solving across various sectors,including R&D&I.13Blue Shift /REPORT 007As a teaser,

60、heres an intriguing anagram of the Reports title:“Eureka!On steroids”also transforms into“Edison treasure:OK!”Anagrams move in mysterious ways,but this one feels particularly fitting.Thomas Edison,a great innovator,stands as a symbol of discovery.This anagram hints that AI may uncover valuable treas

61、ures of innovation in the future.Albert Meige,PhD14Blue Shift /REPORT 00715Blue Shift /REPORT 00716CHAPTER16117THE POTENTIAL OF AI IN R&D&I1THE POTENTIAL OF AI IN R&D&IAI augments researchers capabilities across all steps of the R&D&I process through various roles,helping to solve intractable proble

62、ms and make decisions.No blanket model exists;data availability and problem type determine the best method.Most often,AI models are embedded in a systems of systems.Blue Shift /REPORT 00718THE BENEFITS OF AI IN R&D&IEvery building block of R&D&I has benefits and use cases,from technology and market

63、intelligence to innovation strategy,ideation,portfolio and project management,and IP management.When we look to understand these benefits,three key factors emerge:1.AI augments researchers,rather than replacing themIn the 40-plus interviews for this Report,none of the decision makers was looking to

64、replace their R&D&I workforce with AI,now or in the future.Currently,AI doesnt operate fully autonomously in any use case deployed in our extensive survey sample.Instead,it augments researchers,frees up their time,and enables them to be more productive and creative,often by automating previously man

65、ual tasks.Particularly since the advent of GenAI,researchers have been able to automate repetitive tasks,such as drafting emails or documents and synthesizing the contents of multiple papers.For example,in the case of the food company Roquette,researchers and developers saw their jobs evolve from op

66、erating lab machines to operating AI that operates machines and learning how to handle data.AI works best when it is applied in the context of a“people-centric”lab.12.AI helps solve intractable problemsEquipped with AI,researchers can solve problems they couldnt before because of the technologys spe

67、ed and ability to scale and learn.For instance,to optimize nutrition plans,agro-industrial group Avril developed a model to process historical data that was unusable without AI.Google DeepMind created AlphaFold,an AI model that could examine millions of protein combinations,enabling the discovery of

68、 proteins in novel fields.Without AI,neither these nor many other use cases would be possible.3.AI will assume a“planner-thinker”positionAI embodies an increasing range of intelligence archetypes and is moving beyond content generation and search to cover more complex roles.These include becoming a

69、knowledge manager,hypothesis generator,and assistant to R&D&I teams.Today,the“planner-thinker”archetype is emerging.For example,a military organization has seen AI usage help make decisions,basing recommendations on weak signals from various sources.However,progress needs to be made before fully int

70、egrated workflows emerge in companies/institutions to enable this type of decision-making.AI-BASED MODELS SUPPORT USE CASES AT EVERY STEP OF R&D&I PROCESSADLs Innovation Excellence Model is a proven framework for innovation management within organizations(see Figure 1).It contains 10 elements that a

71、re essential to strong innovation performance.AI can currently augment and enable all of these,albeit at different maturity levels.Johan Aubert,Chief Technology&Digital Officer,LOral“AI does not replace researchers but augments them by automating tasks to free up time for innovation and combining th

72、eir expertise with up-to-date knowledge.”Florent Brissaud,R&D Project Manager for AI Applications,GRTgaz“We had old data,but it was hard to get something out of it.With AI and the access to this data it provided,we trained models to identify characteristics of unreliable assets.”Erez Raanan,CEO,Math

73、labs“Every organization should ask themselves how do we make our analysts evolve?And it will certainly not stop at the analyst level.”19Blue Shift /REPORT 007NO BLANKET MODEL FOR R&D&I TASKSWhen deciding whether to use AI to solve a specific R&D&I use case and which AI approach will give the best re

74、sults organizations need to focus on two factors,as shown in Figure 2:1.The type and amount of data available(from a little to a lot)2.The nature of the question being asked(from open to specific)Figure 1.ADL Innovation Excellence ModelSource:Arthur D.LittleFigure 2.AI systems in relation to other m

75、ethods by problem typeSource:Arthur D.Little,Yves Caseau,National Academy of Technologies of France(NATF)Source:Arthur D.LittleFigure 1.ADL Innovation Excellence Model KNOWLEDGE MANAGEMENTKnowledge management&transferSoftware dev&testing for R&D toolsIP MANAGEMENTDocument management&complianceECOSYS

76、TEM MANAGEMENTAssisted grant&funding applicationsSUPPORTING PROCESSESMARKET INTELLIGENCEAutomated market analysis&competitive intelligenceRegulatory affairs&compliance monitoring070809TECHNOLOGY INTELLIGENCEAutomated technology trends&weak signals watchINNOVATION STRATEGY SETTINGScenario planning&st

77、rategy development 01INNOVATION ENGINE10R&D&I PROJECTS Formulation Materials design&engineering Product design&prototyping Experimental design&automation Safety&environmental impact assessmentPORTFOLIO MANAGEMENTResource allocation&project mgmtIDEA GENERATION&MANAGEMENTIdeation&concept developmentDE

78、PLOY-MENT(market launch,manufacturing and/or scale-up)Automated quality control&defect detection0504060302Source:Arthur D.Little,Yves Caseau,National Academy of Technologies of France(NATF)Figure 2.AI systems in relation to other methods by problem typeGood old-fashioned AIPattern deep learning&visi

79、on modelsSemanticsReinforcement learningClassical scientific approachesOpenLittleSpecificProblemDataLotsGenAI20Blue Shift /REPORT 007Approaches fall under the following categories:-Good old-fashioned AI rule-based systems for specific,structured problems with minimal data-Classical scientific approa

80、ches standard scientific methods that rely on experimentation and validation-Pattern deep learning and vision models pattern recognition and vision tasks requiring extensive labeled data(e.g.,convolutional neural networks)-Semantics structured knowledge representation for understanding context and m

81、eaning(e.g.,knowledge graphs,ontologies)-GenAI Large models handling open-ended,data-rich tasks,often with human feedback(LLM+RL from human feedback RLHF)-Reinforcement learning used for exploring open-ended problems,relying on simulation and relatively limited dataFigure 3 offers guidance in select

82、ing the best approach.However,AI is not always the answer classical science techniques,including traditional regression methods,may perform better on some problems.Tasks essential to the core of R&D&I have no blanket model;specific models and systems deliver optimal results for specific problems and

83、 data.Vincent Champain,Senior Executive VP,Digital Performance&IT,Framatome“We have been running AI models on our super computers for decades,and there is clearly an acceleration since 2022.We also have a deep expertise in exact science models(neutronics,physics,fluids,etc.):after considering a prob

84、lem from every angle,if there is an exact science model available,we use it most of the time because it usually offers better results and predictability.But for the other cases,such as documents or language analysis,or for having a first guess of a solution(that will need to be confirmed by exact sc

85、ience models),GenAI can also bring very interesting results if you fully master its limitations.Even before the AI Act,our approach was to control AI results and always have human operators or engineers doing the final decision.”Figure 3.AI/ML architectures for R&D&I problemsNote:Map focuses on mach

86、ine learning and causal inference methods;it does not focus on symbolic approaches(“good old-fashioned AI”)Source:Arthur D.Little,Yves CaseauNote:Map focuses on machine learning and causal inference methods;it does not focus on symbolic approaches(“good old-fashioned AI”)Source:Arthur D.Little,Yves

87、CaseauFigure 3.AI/ML architectures for R&D&I problemsSequential data(e.g.,language,time series data)Non-sequential data with spatial locality(e.g.,images,grids)Non-structured data(not sequences or grids)(e.g.,molecules,social networks)Discrete states&actions(e.g.,game moves)Continuous states&actions

88、(e.g.,robot moves)Causally linked data(e.g.,treatment effects,incident costs)Classifi-cation/detectionRecurrent Neural Network(RNN),Transformer(e.g.,BERT)Convolutional Neural Network(CNN)(e.g.,AlexNet),K-Nearest Neighbors(KNN),Random ForestGraph Neural Network(GNN)(e.g.,GraphSAGE)NANANAPredictionRNN

89、,incl.long-/short-term memory algorithms(e.g.,Neural Hydrology)CNN,Transformer,Variational Autoencoder(VAE)GNN(e.g.,Attentive FP)NANASimulation(incl.digital twins)+any relevant ML prediction method,Bayesian networks,Causal ForestsControlsCommand&control algorithm+RNNCommand&control algorithm+CNNComm

90、and&control algorithm+GNNModel-free RL(MFRL)(e.g.,AlphaGo)MFRL(e.g.,Q-learning),Model-based RL(e.g.,DemoStart)NAGenerationRNN(e.g.,Google Smart Compose),Transformers(e.g.,GPT-4)Diffusion(e.g.,Stable Diffusion),GAN(e.g.,StyleGAN),VAE+GAN(GANverse3D)VAE+GNN,GANSimulation+Monte Carlo Tree Search(MCTS)S

91、imulation+MCTSNAINPUT DATA TYPETASK TYPEMore specificMore openNON-EXHAUSTIVE21Blue Shift /REPORT 007System of systemsAt the same time,a single AI approach may not deliver optimal results most state-of-the-art intelligent systems produced in the past 15 years have been systems of systems.These are in

92、dependent AI systems,models,or algorithms designed for specific tasks,which,when combined,offer greater functionality and performance.For example,most LLM chatbots,such as ChatGPT,use a transformer architecture coupled with RL from human feedback(RLHF).Image generation model DALL-E 2 brings together

93、 an auto-regressive transformer and a diffusion algorithm.LLMs can often play an orchestrating role in interfacing with or controlling other systems because of their language fluency.Robotics use cases often require a system-of-systems approach,as set out in Figure 4 for manufacturing robotics.A com

94、puter vision algorithm can process the visual data gathered by the robots camera;a symbolic,rule-based engine can inform the decisions of the cognitive controller,which,in turn,can be influenced by an RL algorithm that learns from past experiences or other external knowledge.Figure 4.System of syste

95、ms example,including AI in robotics for manufacturingSource:Arthur D.Little;Oliff,Harley,et al.“A Framework of Integrating Knowledge of Human Factors to Facilitate HMI and Collaboration in Intelligent Manufacturing.”Procedia CIRP,Vol.72,2018.Source:Arthur D.Little;Oliff,Harley,et al.“A Framework of

96、Integrating Knowledge of Human Factors to Facilitate HMI and Collaboration in Intelligent Manufacturing.”Procedia CIRP,Vol.72,2018.Figure 4.System of systems example,including AI in robotics for manufacturingExamples of possible AI modules augmenting system capabilitiesDATA COLLECTIONROBOTICSCOGNITI

97、VE LAYERACTIONPerceptionDecision-makingInformation retrievalData preprocessingExternal knowledgeLearning mechanismCognitive controllerI/O controllerActuatorsMotorsSample/eventCognitive controlRL modelwith a policyComputer vision modelEnvironmental&process dataData collectionHistorical dataRule-based

98、 engineRobot controllerFEEDBACK22Blue Shift /REPORT 00723Blue Shift /REPORT 007“The measure of intelligence is the ability to change.”Albert Einstein24CHAPTER24225HOW TO ENSURE SUCCESS2HOW TO ENSURE SUCCESSEnsuring success in the implementation of AI for R&D&I requires agile methodologies,robust dat

99、a foundations,strategic prioritization,analytical tradeoffs,scarce data science talent management,IT alignment,rapid benefit demonstration,and continuous monitoring.Blue Shift /REPORT 00726To gain stronger knowledge of good practices in R&D&I,we conducted 40-plus interviews with researchers;AI scien

100、tists;founders;and heads of R&D in digital,manufacturing,marketing,and R&D teams.Respondents are from 57 companies and three research institutions with interviews taking place June 2024October 2024.Companies are headquartered on three continents and span the defense,chemicals,automotive,software dev

101、elopment,and consumer goods sectors with revenues of US$10,000 to over US$1 billion.AI projects range from a few weeks to multiple years in length.THEMES UNDERPINNING GOOD AI PRACTICEWe have distilled these interviews into eight good practices(agile methodologies,robust data foundations,strategic pr

102、ioritization,analytical tradeoffs,scarce data science talent management,IT alignment,rapid benefit demonstration,and continuous monitoring)across four categories(data and project management,strategic implementation,organizational structures,and sustained adoption and impact)that help maximize AI suc

103、cess in R&D&I.Establish strong data&project management foundationsAdopt agile methodologiesAI is a new,evolving discipline,meaning projects will likely change as they progress.Thus,R&D&I teams need to apply agile methodologies to work quickly and iteratively something that is all the more necessary

104、given the rapid pace of technological progress in AI.Successful projects are likely to follow the agile development best practices set out in Figure 5 from project launch onward.Figure 5.Agile development best practices for AI projectsSource:Arthur D.LittleSource:Arthur D.LittleFigure 5.Agile develo

105、pment best practices for AI projectsBUILD TEAMForm cross-functional core team with IT,data science&domain expertise.Ideally,colocate team in single workspaceADJUSTOnce in production,collect model outputs&expected values for post-deployment adjustmentsDEFINEDefine objectives of incrementDESIGNDesign

106、model:developing/modeling/fine-tuning based on datasetVALIDATEValidate model using separate validation datasetREVIEWReview&adapt:decide on model modifications,dataset adjustments&out-of-sample verificationIMPROVEConduct retrospective on increment&decide actions to be implemented on next incrementSEL

107、ECTSelect right tool&create initial dataset in zero increment phaseTRAINTrain business users outside core teamDEVELOP INCREMENTALLY TO MAINTAIN FLEXIBILITY IN MODEL REFINEMENTDenis Gardin,Innovation Director,MBDA“Move fast and dont wait for the perfect technology.AI will evolve;it is better to have

108、something that was released on the market two years ago than nothing at all.”27Blue Shift /REPORT 007Build robust foundationsStrong,structured data management capabilities are central to realizing the benefits from AI.These need to span data quality,collaborative data management,and successfully lev

109、eraging proprietary data:-Focus on data quality.High-quality,structured data is crucial,but emerging techniques offer new possibilities for using smaller datasets to achieve significant results.For example,Alysophil,Gourmey,and Integrated Biosciences rely on high-quality,structured data from experim

110、ental runs for accurate predictions of the properties of new compounds.After each run,models are retrained with the newly produced outputs and experimental parameters adjusted to produce new data.Unsupervised learning models(e.g.,LLMs)can process unlabeled data,potentially unlocking value from histo

111、rical datasets.-Encourage cross-organizational collaboration.Successful AI implementation requires cross-organizational collaboration,data accessibility,and effective governance.Demonstrating this,Veolia,J&J,and Avril emphasize breaking down data silos between and inside teams.For example,at Avril,t

112、he initial cleansing of historical data highlighted the need for standardized data collection.Based on this,the R&D department has implemented a uniform and refined method for data acquisition,laying a solid foundation for future projects.-Leverage proprietary data.Harnessing proprietary data provid

113、es companies with a significant competitive edge as models fine-tuned with proprietary data provide more relevant outputs specific to teams needs.For example,by using their own data,LOral and Air Liquide created solutions that were unavailable to the competition.LOral replaced animal testing with AI

114、-driven methods based on its historical data for cosmetics safety assessments.Air Liquide created a custom model to find new polymers based on 30 years of internal polymer-related data.Adopt a strategic approach for AI implementationMake a strategic choice between building,buying&fine-tuning modelsR

115、&D&I organizations have three possibilities for AI models they can build their own from scratch,buy from a specialist provider,or fine-tune an existing AI model,often available via open source.The strategic choice should be based on the specific use case and internal capabilities,following the decis

116、ion flow in Figure 6.Carlos Escapa,Data AI/ML Global Practice Lead,Accenture AWS Business Group“Companies that excel in AI adoption are those with a structured approach to data management.They collect sufficient amounts of data,maintain strong compliance practices,efficiently utilize data lakes,and

117、move data quickly and seamlessly throughout their organization.”28Blue Shift /REPORT 007Most core R&D&I problems lend themselves to fine-tuning existing open source models,whether LLMs(e.g.,Llama,Mistral,Cohere),state-of-the-art GANs and diffusion models(e.g.,Stable Diffusion),or RL(e.g.,TensorFlow)

118、.Research ecosystems are particularly amenable to fine-tuning,as the open source community is largely fed by academia,as shown by the development of the LoRA matrices.In some very specific use cases,an in-house model developed from scratch may outperform a fine-tuned public model at an acceptable co

119、st.Examples include hybrid models seeking to embed scientific knowledge in the models mechanics,largely experimental architectures that are less demanding in terms of compute(e.g.,recurrent neural networks RNNs),or very small models developed to run on specific edge devices.However,pure“productivity

120、”AI use cases are best bought off-the-shelf,including from specialized R&D&I application providers.For LLMs,prompting can provide satisfactory customization at a very low cost,with no coding skills required(e.g.,OpenAIs GPTs).The retrieval augmented generation(RAG)technique also enables customizatio

121、n,tailoring an LLM to an organizations knowledge base without fine-tuning.Figure 6.Choosing between buy,fine-tune,and make for AI modelsNote:(1)Even then,models are preferably only developed in-house when there is a significant analytical or efficiency upside to doing so Source:Arthur D.LittleNote:(

122、1)Even then,models are preferably only developed in-house when there is a significant analytical or efficiency upside to doing soSource:Arthur D.LittleFigure 6.Choosing between buy,fine-tune,and make for AI modelsBUYMAKEFINE-TUNEIs there a suitable solution for this problem off the shelf?Is data con

123、fidentiality/sovereignty a concern?Is pretraining the model affordable?Do you have the skills/resources?YESNONOYESNOYESYES1NOReady-to-use,quickly implementableUser-friendly,Optimized for productivityNo customizationSaaS model:recurring costs can become significantAffordable&quickAccess to open sourc

124、e communitys creativityNo expert skills requiredLimited customizationNot applicable to most novel or specific problemsBuilt from scratch for a specific problemUpdatable,flexibleMore interpretableCostly in compute&resourcesPerformance may be worse than publicly available models(esp.for LLMs)PROSCONS2

125、9Blue Shift /REPORT 007Consider analytical tradeoffs to ensure progress during POCOrganizations face a range of constraints during AI projects and should therefore consider analytical tradeoffs in three areas to move projects forward:1.Acquiring versus synthesizing data.Organizations must choose bet

126、ween generating/acquiring more data,which could take longer,or adopting other approaches.They could use GenAI methods to create synthesized data or data augmentation to slightly tweak their initial datasets,although this could increase training set biases/errors.Best practices include transfer learn

127、ing,which helps overcome the lack of data by using the backbone network output of previously trained models as features in further stages of new models,or embracing techniques such as Box-Behnken,which helps optimize data collection.2.Precision versus recall.Does your model output favor false negati

128、ves or encourage false positives?Recommended best practices include assigning a specific cost to each type of error to understand which metric to prioritize.Alternatively,in some task types(open versus closed),depending on the stakes involved(e.g.,production line monitoring versus brainstorming),som

129、e imbalance can be valuable,as it could generate completely new ideas.3.Underfitting versus overfitting.The choice here is between decreasing training data loss,which could increase test data loss and mean the model is not able to generalize or be creative,and decreasing test data loss,which might i

130、ncrease training data loss,leading to the model not being accurate.Regularization techniques(e.g.,L1,L2,Dropout,and Early Stopping)mitigate overfitting by penalizing excessive model complexity,ensuring they generalize better to new data by reducing the magnitude of high coefficients without necessar

131、ily reducing the number of variables.If none of these tradeoffs proves satisfactory,organizations should reformulate the problem.Align organizational structures to overcome barriers to AI adoptionBe proactive in leveraging available data science talentAI talent is a precious,often scarce resource.Di

132、stributing it effectively and,if necessary,plugging resource gaps are essential if R&D&I organizations are to ensure successful deployment at scale.Depending on the size of any talent gaps,they can choose between the five organization models set out in Figure 7.The first three work best for those wi

133、th insufficient resources;the last two best suit those with the right levels of skills in place.Carlos Escapa,Data AI/ML Global Practice Lead,Accenture AWS Business Group“We see cases where the POC appears to work fine on a limited set of data,but low accuracy and hallucinations appear as we try to

134、scale.In the case of RAG,the main reason is data quality.Prompt engineering cannot fix errors and inconsistencies in the source data.”30Blue Shift /REPORT 007Align with IT to balance security&compliance with experimentation speedIT departments face four common concerns with introducing new AI tools:

135、1.System maintenance and integration challenges.IT departments can face difficulties integrating AI solutions with existing systems,which can slow deployment.In particular,scalability concerns that demand thorough planning for future-proof architectures often lead to less agile implementations.2.Com

136、pliance with internal policies.Strict internal policies regarding compliance and cybersecurity can create resistance to AI adoption,particularly when IT must ensure adherence to these internal regulations.This may lead to overly cautious approaches when deploying new technologies.3.Legal and data pr

137、otection regulations.IT departments are responsible for ensuring compliance with legal standards,such as the EUs General Data Protection Regulation(GDPR)and intellectual property(IP)laws.For example,Roquettes IT department receives constant pitches from AI providers.While the promised benefits may b

138、e appealing,the terms of service are often too constraining,requiring a careful review process before adopting an AI tool.4.Supervision of deployment processes.Conflicts may arise between the need for rapid experimentation in R&D,ITs need to oversee deployment,and resource constraints.All these requ

139、ire close alignment with IT,building an understanding of differing needs and working together to move forward.Figure 7.Organization models for leveraging data science talentSource:Arthur D.LittleSource:Arthur D.LittleFigure 7.Organization models for leveraging data science talentLess neededMore need

140、edMODELPROSCONSDESCRIPTIONExternalizeTrainPairService centerEmbed+Testing new ideas is quick+Talent is available immediately+Convenient for punctual uses+Tools are highly relevant+Fear of replacement is limited+Minimum training can help an expert set up a simple model+Enables mentoring approach that

141、 benefits each+Among most efficient usages of data scientists&data+Duplicates are avoided+You can mutualize real data science expertise+Data scientists gain knowledge on subject+Project execution capabilities are improved-Recurring costs-Could be a limit to customization-No real knowledge capture fr

142、om internal resources-Training takes time-Subject to employee churn-Data science experts still needed-Difficult to roll outat large scale-One data science resource per project still required-Feedback might be limited-Might not be enough“as is”to answer very specific needs-Creativity is limited-Commu

143、nication&collaboration can be challenging between researchers&data scientistsSeek support from data science providers,use their pre-build and/or pre-trained models&access their training data setsADL supported a big pharma for an AI model to hire patients more efficientlyTrain subject matter experts

144、in data science In 2022,Air Liquide deployed its internal AI readiness program with the objective of training 300 employees on operations in data science/AI by 2025 in addition to the ones in R&D,digital&ITPair a data scientist with an expert MaiaSpace pairs young researchers with experienced ones t

145、o couple energy&ideas with experienceCentralize all demands with unique service centerNestl has a team in Switzerland developing models for the whole company,which made it possible to implement AI in every processImplement group of data scientists in each R&D teamSolvay,LOral created hybrid teams to

146、 leverage both domain expertise&data science skills in highly specific R&D domain RESOURCESCarlos Martin,Managing Director,MACAMI Group“You must deal with internal politics.Something observed in almost any company is that researchers are on board,but IT is not willing to cooperate.”Research Director

147、,consumer goods player“There were internal politics games:researchers were onboard,but IT was not willing to cooperate.”31Blue Shift /REPORT 007Ensure sustained adoption&impact of AI solutionsDemonstrate benefits quickly&get user buy-inAll AI projects face challenges and bottlenecks,including employ

148、ee fears,which can cause POC projects to stall.A transparent,people-first approach that aligns solutions with needs and builds trust over time can help overcome such challenges.Maintain&monitor system performance continuouslyExperimental AI models can stray from their expected behavior over time,lea

149、ding to inaccurate results if not supervised continuously,focusing on performance monitoring and model improvement:-Performance monitoring.Comprehensive monitoring of AI model performance is crucial for maintaining accuracy over time.Establish baseline metrics and set performance thresholds based on

150、 initial model validation,including accuracy,precision,recall,and F1 score.Then,monitor input characteristics.Model performance can be impacted by the change in quality or distribution of input data over time compared to the training set.Inconsistencies or errors in input data can arise from changes

151、 in data pipelines,source data schema modifications,or data corruption.The overall distribution of input data can change compared to the training data,causing the model to be less relevant,while extreme or unexpected data points may appear in input data,potentially skewing predictions.-Model improve

152、ment.Continuous improvement of AI models is essential to maintain their relevance and performance over time.Organizations should focus on three areas.They need to fine-tune the model by keeping it updated with new,relevant data to adapt to changes in the underlying patterns or distributions of the d

153、ataset.Then,they must adjust the model architecture,modifying its structure or hyperparameters to better suit the current data landscape.Finally,they should build new versions of the model,using A/B testing to compare new versions against current production models.Jrme Christin,VP,R&D group,Air Liqu

154、ide“A small project that is successful can build more trust in AI than many explanations.”Bruno Guilbot,Data Science,AI,and New Technologies Director,Louis Vuitton“MLOps is crucial for maintaining the effectiveness of predictive models.Once they are in production,business and operational teams rely

155、on these models for critical tasks and decision-making.Its essential to monitor the models ongoing accuracy,compare predictions with real outcomes,and ensure the model doesnt diverge or drift.This requires a robust process for continuous evaluation and improvement to maintaining reliable and sustain

156、able performance.”32Blue Shift /REPORT 007GOOD PRACTICE CASE STUDIES OF AI ADOPTION IN R&D&IBased on our interviews and further research,we have collected six best-in-class experiments with AI in R&D&I across corporate and public research.Public research institution:Resource management Public resear

157、ch institution:Resource management 20ObjectiveForm the best multidisciplinary teams for cross-functional research projects with the help of AIProblemResearch institutions have difficulty assigning appropriate reviewers to interdisciplinary research projects,making it time-consuming to find the right

158、 combination.They must consider expertise,location,availability,affinity,and use rate.The same issue occurs when creating research teams of experts in different fields for interdisciplinary research projects.Technical designInitial engine:regular convolutional neural network(CNN),then updated to dee

159、p learning model;knowledge graph for expert research area relationshipsData usedExpert profiles,historical proposal-reviewer matching,research proposal content weightsAI rolesAnalyst,engineer,scientistImplementation sequenceFirst,a massive reviewer profile database was developed,and then a knowledge

160、 graph linking experts to research areas was extracted while training the AI model using historical human-made matches.The model was initially deployed with a CNN-based engine,which has recently been updated to a deep learning engine.Best practicesUse a multipronged approach combining expert profile

161、s and proposal analysis;leverage historical data on successful human-made matches;start with a small,dedicated team(two people,25%time over two years);continuously improve the model(e.g.,by upgrading from CNN to deep learning).BenefitsImproved allocation of reviewers,increasing efficiency and releva

162、nce,while enabling the faster creation of cross-disciplinary teams.Cosmetics industry player:Idea generation&managementCosmetics industry player:Idea generation&management21ObjectiveCapture expert knowledge by training models with Bayesian networksProblemThe cosmetics company lacked sufficient data

163、to train reliable AI model and faced loss of knowledge once experts left company.Technical designBayesian networks,with LLMs as a first layerData usedExpert interviews,internal data to complete the modelAI rolesLibrarianImplementation sequenceWhen statistical approaches are not enough,Bayesian netwo

164、rks are an option.In this case,an off-the-shelf model was adopted and fine-tuned by the AI provider based on a series of expert interviews.Best practicesCombine experts from different fields for the best results;dont underestimate the human side,as experts can see this exercise negatively and fear b

165、eing replaced instead,present it as a collective undertaking.BenefitsRetaining expertise when employees leave enables real innovation,as AI is not limited to existing data.33Blue Shift /REPORT 007Food industry player:Innovation project managementFood industry player:Innovation project management22Ob

166、jectiveIntegrate AI in every step of product development,reducing time to market and industrial failure ratesProblemIn highly competitive markets,this food company needed to transform product development,reducing time to market and increasing end-to-end efficiency all the way through to manufacturin

167、g.Technical designModels developed in-house,centrally at headquarters,covering end-to-end process:AI augments everyday tasks and knowledge management.Trend identification:AI scans mature markets and finds key characteristics of successful products.Formulation prediction:desired flavors are entered,a

168、nd AI provides rapid solutions,or if new regulations require ingredients to be replaced,AI finds best candidates.Experimental design:AI crafts tests that comply with regulations.Manufacturing troubleshooting:through a digital twin of the manufacturing line,AI identifies product or machine problems t

169、hat cause potential manufacturing issues.Data usedInternal research,customer habits,product and machine characteristicsAI rolesAnalyst,engineer,scientistImplementation sequenceModels are first developed at headquarters and fed with data collected across every process.Change management team drives AI

170、 acceptance,explaining the benefits,kickstarting adoption,and running quarterly bottom-up feedback sessions.Best practicesBuild skilled teams that understand that physical and AI functioning should be combined in AI training and fine-tuning;carry out regular checks on the model to monitor output qua

171、lity.Benefits30%time-to-market reduction and 40%drop in industrial failure rates Retail industry player:Market,customer,operations insights&analyticsRetail industry player:Market,customer,operations insights&analytics23ObjectiveBetter understand customers,position brand fast,and gain unexpected cust

172、omer insightsProblemThis retail industry player needed to be able to develop products quickly,requiring faster and more in-depth market research capabilities to understand evolving customer needs better.Technical designML models,interface to access platform,or API for technical teams.While off-the-s

173、helf solutions were available,the model was developed in-house.Data usedCustomer data,basket contents,and preferences(surveys,online reviews,etc.)AI rolesAnalystImplementation sequenceFirst,the company collected and cleaned in-house data,including information collected from all departments,and purch

174、ased proprietary data to supplement this.The model was then trained on this proprietary database and regularly retrained to keep up with the latest trends.AI is now used to recognize patterns in customer baskets,identifying what sells well,trends,and more complex links(e.g.,the combination of produc

175、ts bought together and seasonality).Best practicesProduce well-structured data as early as possible or quickly clean historical data to ensure the model is tailored to your customers and brand,differentiating from competitors.Retain employees who previously carried out this role manually;they will t

176、rain the model better than data scientists alone and be able to interpret outputs.BenefitsEnhanced probability of successful product launches34Blue Shift /REPORT 007Chemical industry start-up:Innovation product managementLeading university:Ecosystem managementChemical industry start-up:Innovation pr

177、oduct management24ObjectiveConduct experiments and retrain the model every time a result is addedProblemReaching desired chemical properties requires exploratory experiments and mapping formulation properties dependent on the proportion of each element.Researchers can estimate the most promising pro

178、portions but still need to do thousands of experiments to confirm this.This chemical industry start-up wanted to reduce the number of experiments conducted to increase speed and efficiency.Technical designIn-house Bayesian optimization and Gaussian process model,based on tweaked open source algorith

179、ms,available via existing interface and APIsData usedSolvent combinations and proportions,viscosity measurementsAI rolesEngineerImplementation sequenceAI can help design experiments based on selected proportions to test,adapt them to the first results,and predict the properties of all possible exper

180、imental combinations.The process began by setting up an experiment,including a robotic arm to automate the process.Several hundred experiments were then carried out following the classical statistical approach.This data was used to train the model and output the first experimental designs.The compan

181、y then conducted further experiments and retrained the model with every new result produced.Even when finding the final“perfect”combination,it let the model try to diverge to potentially find unexpected results.Best practicesWhen starting with incomplete data,have experts steer the model to compensa

182、te;fully include experts in the process,rather than replacing them;do not generalize the model,as it is problem-specific.BenefitsConfirm desired properties by carrying out 10 experiments,rather than up to 100.Leading university:Ecosystem management25ObjectiveFacilitate technology transfer from resea

183、rch institutions to companies ProblemAcademic incentives focus on publication,which makes achieving commercial or industrial impact an afterthought,particularly as researchers often lack commercialization awareness or training.Given that technology transfer offices are understaffed and limited in sc

184、ope,many innovations struggle to find a go-to-market route.This university wanted to increase success rates and efficiency by augmenting every step of the technology transfer process using AI while matching projects with commercial needs.Technical designClosed AI systems tailored for technology tran

185、sfer-specific tasks(e.g.,agreement generation and review),HIPAA-compliant systems for handling confidential IP information,and LLMs to match innovations with commercial needsData usedResearch reports and publications,company interests,and focus areasAI rolesAnalyst,engineer,scientistImplementation s

186、equenceThe university first developed an AI-powered information-extraction system for research reports while creating a database of company interests and focus areas,leveraging its existing ecosystem.It then established an open model sandbox for tech transfer and implemented closed AI systems to han

187、dle technology transferspecific tasks.Finally,it explored further partnerships with scientists and companies for AI-driven innovation.Best practicesUse private AI instances of off-the-shelf solutions to ensure data confidentiality and security;make significant up-front time investment to train staff

188、;encourage evangelists who can champion AI adoption and guide others to foster a culture of innovation;emphasize user accountability for AI-generated outputs to maintain quality and prevent the treatment of AI as a scapegoat for errors or poor decisions.Benefits95%greater efficiencies for some techn

189、ology transfer workflow tasks35Blue Shift /REPORT 007INTERLUDE Focus on data!by Anne BouverotAnne Bouverot is French President Emmanuel Macrons Special Envoy for the Artificial Intelligence Action Summit,taking place in Paris on 10-11 February 2025.She spent most of her career in the technology sect

190、or and now advises several public and private technology companies and scale-ups.Ms.Bouverot chairs the Board of the Ecole Normale Suprieure,Frances leading grande cole in science and humanities and cochairs the AI&Society Institute.In 2017,she cofounded the Fondation AbeonaChampioning Responsible A

191、I,addressing the societal impacts of AI.Programs include a visiting chair on social justice and AI and an introductory massive open online course(MOOC)on AI,followed by more than 350,000 people.Ms.Bouverot spent the first 20 years of her career at Orange in various positions,then became Director Gen

192、eral of the GSMA(Global Mobile Telecommunications Association)and later CEO of Morpho(digital security and identity solutions).She is a mathematics graduate of Ecole Normale Suprieure and holds both an engineering degree in telecommunications and a PhD in AI.36Blue Shift /REPORT 007AI&SCIENCE:A SYNE

193、RGETIC REVOLUTIONAs my coauthors and I noted in“Our AI:Our Ambition for France,”2 a report presented to French President Emmanuel Macron in March 2024,AI is an inescapable technological revolution impacting all fields of activity.R&D&I is no exception.In fact,AI and R&D&I are intrinsically linked,no

194、t just coexisting but fueling each other.The most recent Nobel Prizes illustrate this perfectly:the Nobel Prize in Chemistry awarded to Demis Hassabis and John Jumper showcases how AI now drives protein discovery,enabling significant advances in biology.Similarly,the Nobel Prize in Physics was award

195、ed to John Hopfield and Geoffrey Hinton for their work on statistical physics that laid the foundation for neural networks,which today power complex AI algorithms.Science and AI intersect and strengthen each other,and its precisely in this symbiosis that the future of R&D&I lies.DATA ACCESS&COMPUTAT

196、IONAL POWER:ESSENTIAL DRIVERSAs Jolle Barral,head of AI research at Google DeepMind,recently noted,“AI will accelerate research far beyond what we currently imagine.”However,for AI to fulfill its promises,two elements are crucial:access to data and computational power.The achievements of Hassabis an

197、d Jumper were made possible by the availability of extensive protein datasets already collected.However,in other fields,data remains the limiting factor.The critical challenge will be to collect and share data openly and at a sufficient scale to allow researchers to fully leverage AIs capabilities.A

198、ccess to powerful computing platforms that can process these vast datasets,such as supercomputers,must be secured for both the public sector and private enterprises.It is essential that these infrastructures are made available on an equitable basis.Beyond technical challenges,international cooperati

199、on on data governance is critical.Building a shared ecosystem with standardized data models and interoperable frameworks is the key to maximizing synergies between AI and R&D.EUROPE FACES GLOBAL CHALLENGEEurope,particularly France,is home to a rich AI talent ecosystem,with renowned figures like Barr

200、al,Yann LeCun(Meta),and Arthur Mensch(Mistral AI).While some of these experts have moved abroad to further their careers,many have returned or continue to shape strategic sectors in Europe,whether by founding companies or reinforcing local labs of global tech giants.To retain them,however,we must of

201、fer attractive working conditions in both the public and private sectors,particularly by enhancing investment capacities in critical scale-up phases.The ambitious 5 billion annual investment plan proposed in“AI:Our Ambition for France,”is a concrete example of what can be done to ensure Europe remai

202、ns competitive.This investment should focus not only on infrastructure but also on continuous training and upskilling of R&D teams to address the talent gap.Additionally,fiscal incentives,such as tax credits for supercomputer usage,could play a key role in democratizing access to these advanced tech

203、nologies.UNDERSTANDING&SHAPING THE FUTUREIn conclusion,I would like to quote physicist Marie Curie:“Nothing in life is to be feared;it is only to be understood.”This quote reminds us that while AI brings new risks,it also offers immense opportunities to build a more innovative,prosperous,and resilie

204、nt future.But to achieve this,we must invest now in data access,computational infrastructure,and talent development to position Europe as a leader in this new revolution.37Blue Shift /REPORT 00738CHAPTER38339TOOLS&PROVIDERS3TOOLS&PROVIDERSThe value chain for AI in R&D&I is heavily reliant on major o

205、pen source models,but smaller players also form a key part of the ecosystem.Applications tailored for every part of the R&D&I process exist,as do start-ups targeting vertical-specific problems.Hosting providers also offer inference as a service.Blue Shift /REPORT 00740THE R&D&I AI VALUE CHAINAs with

206、 most AI use cases,the R&D&I value chain comprises three layers(see Figure 8):1.Infrastructure.Compute is delivered by super calculators,GPU providers,and cloud computing companies.2.Model developers.Predominantly via open source models,developed and trained by major players such as Meta(Llama),Micr

207、osoft(Phi),and Nvidia(NVLM,TensorFlow,StyleGAN),along with smaller players and academic institutions.3.Applications.At the application end of the chain,general and specialist R&D&I apps have been created to meet most use cases.Demonstrating the spread of these AI applications,over 500 are now availa

208、ble for different R&D&I use cases,covering the entire R&D&I process(see Figure 9).Figure 8.R&D&I AI value chainSource:Arthur D.LittleFigure 9.R&D&I AI applications mapped to ADL Innovation Excellence ModelNote:Research conducted to find providers with solutions based on AI:a list of 500 was sifted d

209、own to 90,which were mapped to the framework enriched with the interviews,this constitutes a long list of 130 relevant providers;some solutions could manage more than one innovation function Source:Arthur D.Little,Yves Caseau,National Academy of Technologies of France(NATF)Source:Arthur D.LittleFigu

210、re 8.R&D&I AI value chainUsersApplicationsSoftware solutions developers leveraging GenAI models to deliver intelligent&automated features(off-the-shelf or custom)Foundation modelsDevelopers of AI models:designing architecture,training,fine-tuning,optimizing performanceDevelopment toolsCloud computin

211、g providersSuper-calculatorsData processing&managementData flowTraining dataINFRASTRUCTUREAPPLICATIONSMODEL DEVELOPERSGPU providersCollaborative platformsRetrieval dataServices using AI appsOut of scope of this studyCode flowNote:Research conducted to find providers with solutions based on AI:a list

212、 of 500 was sifted down to 90,which were mapped to the framework enriched with the interviews,this constitutes a long list of 130 relevant providers;some solutions could manage more than one innovation functionSource:Arthur D.LittleFigure 9.R&D&I AI applications mapped to ADL Innovation Excellence M

213、odelKNOWLEDGE MANAGEMENTIP MANAGEMENTECOSYSTEM MANAGEMENTSUPPORTING PROCESSESMARKET INTELLIGENCE070809TECHNOLOGY INTELLIGENCEINNOVATION STRATEGY SETTING01INNOVATION ENGINE10R&D&I PROJECTSPORTFOLIO MANAGEMENTIDEA GENERATION&MANAGEMENTDEPLOY-MENT 050406030241Blue Shift /REPORT 007PROMISING R&D&I AI AP

214、PLICATION PLAYERSBased on desk research and interviews,we have identified 12 promising offerings gaining traction within different parts of the R&D&I space.While not exhaustive,it provides a snapshot of currently available tools,their capabilities,and promised benefits:1.GetFocus technology forecast

215、ing and monitoring based on analysis of patent data2.Quid(Netbase)platform that uses GenAI to provide comprehensive view of customer sentiment and behavior3.Wazoku platform to help capture,evaluate,and implement ideas4.Sakana.ai solution promising fully automatic scientific discovery using foundatio

216、n models and LLMs5.Osium AI platform to predict the properties of chemicals and materials and screen and test candidates6.Alysophil solution that combines AI and flow chemistry technologies to enable agile,intelligent manufacturing7.LandingAI computer vision cloud platform for production lines in sh

217、ort time-to-market industries8.Patsnap automated IP intelligence and management platform9.Bayesia tools for building,analyzing,and reasoning with Bayesian network models coupled with AI10.Sinequa search-powered AI assistant platform11.Causality Link AI platform that analyzes economic indicators/mark

218、et events to predict future trends12.Simporter AI-powered forecasting to predict the required characteristics of new productsApplications tailored for every part of the R&D&I process exist.42Blue Shift /REPORT 00743Blue Shift /REPORT 007 Demis Hassabis,cofounder,Google DeepMind“We are at the beginni

219、ng of a revolution that is fundamentally changing our understanding of intelligence.”44CHAPTER44445NAVIGATING THE FUTURE4NAVIGATING THE FUTUREHow AI in R&D&I will evolve depends on the outcomes of three main factors:performance,trust,and affordability.These lead to six plausible future scenarios on

220、a spectrum between AI transforming every aspect of R&D&I at one end and being used only in selective,low-risk use cases at the other.Blue Shift /REPORT 00746OUTCOMES SHAPING FUTURE OF R&D&IThree factors will drive the adoption and benefits of AI within R&D&I moving forward:1.Performance.How good wil

221、l AI be at solving R&D&I problems?The bar for AI performance is high in R&D&I.While it will likely improve for several R&D problems,whether those gains will suffice to improve R&D&I efficiency by 2030 is uncertain.2.Trust.Will teams trust AI models and outputs?Greater trust in AI is a key adoption f

222、actor.It is likely to grow among researchers and developers with exposure,interpretability,and performance gains.But trust could be hampered by unreasonable expectations,public attitudes,or fear of job replacement.3.Affordability.What will be the financial,environmental,and operational affordability

223、 of AI systems?The affordability of AI matters to R&D&I,where budgets are smaller and use cases less scalable.While the implementation of AI models is likely to become more affordable(in time,money,skills,and resources),it could be constrained by insufficient data and price hikes in a consolidated A

224、I market.In turn,these three factors will be shaped by 16 underlying trends that we can see emerging today:Performance1.Maturity of multimodal models.A multimodal model is capable of processing information from different modalities(including images,videos,audio files,text,and 3D representations)to e

225、ither“convert”from one medium to another or learn from various media inputs to reach a prediction with a single output(often text).The recent release of multimodal foundation models(GPT-4,Gemini Ultra,Claude 3.5,Llama 3.2)showcases their versatility in managing both images and text and,in some cases

226、,audio(OpenAI Whisper).Multiple R&D&I applications have already been deployed,including in life sciences(patient diagnosis prediction from multiple document types)and consumer goods(multimodal sentiment analysis to optimize product development).2.Rise of Graph Neural Networks(GNNs)for unstructured d

227、ata.GNNs can operate and learn from“unstructured”or“graph-structured”data(as opposed to sequential data such as language or grid-structured data such as images).GNNs can capture the complex relationships between different nodes in a graph,making them particularly suitable for analyzing social networ

228、ks and molecular structures.GNNs advanced applications are growing across many different research fields.For example,they are already used in environmental research such as weather forecasting (e.g.,Googles GraphCast),chemistry to research the graph structure of molecules or compounds(e.g.,Google De

229、epMinds AlphaFold),and materials science to explore new materials and predict their stability(e.g.,Googles GNoME project).The affordability of AI matters to R&D&I,where budgets are smaller and use cases less scalable.47Blue Shift /REPORT 0073.Emergence of hybrid models.Hybrid models mix a probabilis

230、tic architecture with a symbolic architecture(e.g.,first-order logic or the laws of physics).Depending on the system and implementation,the benefits include more robustness(limited hallucinations),greater generalizing power,incorporation of existing knowledge(e.g.,physical laws),improved computation

231、al efficiency,better handling of uncertainty,and better interpretability all of which are important qualities in R&D&I use cases.I-JEPA,an image recognition model implemented by Meta based on Yanna Le Cuns hybrid JEPA framework,achieved state-of-the-art performance in June 2023 after being trained w

232、ith just 16 H100 GPUs in 72 hours.Hybrid models are expected to demonstrate significant benefits in fields that require“sensory grounding,”such as experimental physics.4.Further scientific exploration via RL.Reinforcement learning is a type of machine learning in which an agent learns through trial

233、and error by interacting with an environment(real or modeled).Based on its actions,the environment delivers a positive or negative reward to the agent,which learns to maximize total cumulative rewards and updates its policy for future actions accordingly.RL models have already been successfully appl

234、ied in physics to nuclear fusion,in medicine to drug discovery,in math to theorem proving,and in the surveying of large spaces of objects to discover new patterns.RL is suitable for the open exploration of new ideas,a capability particularly attuned to more fundamental R&D problems for example,desig

235、ning new chips or writing assembly code from scratch.RL has also been broadly used in robotics,including autonomous cars,and will likely enable robotic manipulation in laboratories.However,RL is very computationally intensive,and the ecosystem of hosting services for RL is not yet as commoditized as

236、 it is for LLMs.5.Advances in agentic workflows.Agentic systems are AI-powered frameworks designed to perform tasks with a degree of autonomy and intelligence reminiscent of human agents.These systems are characterized by their ability to perceive their environment,make decisions,take actions,and le

237、arn and adapt.More advanced workflows include different types of AI agents(e.g.,reflective,tool using,planning,or collaborative)working together,sharing goals,and making collective decisions to tackle tasks more effectively.The first widely used open source framework for multi-agent orchestration wa

238、s Microsoft AutoGen(September 2023),followed by MetaGPT,CrewAI,and LangChains LangGraph.A popular open source example of a multi-agent system is ChatDev AI,in which a group of AI agents work together to build software programs.Sakana.ais AI Scientist is an example of an agentic workflow tailored to

239、research(see Figure 10).RL is suitable for the open exploration of new ideas.48Blue Shift /REPORT 007Trust1.Progress in mechanistic interpretability.AI models,especially those based on neural networks,do not provide explanations or rationales for their predictions and are not easily auditable,underm

240、ining users trust.Techniques vary from external behavioral to in-depth mechanistic analysis(see Figure 11).Mechanistic interpretability aims to articulate the rationale for a model prediction by“reverse engineering,”or making the models inner workings explicit by looking at its bottom-up components

241、offering more compelling explanations for model predictions than competing approaches.Since 2019,the number of papers on transparency and explainability submitted to major academic conferences has more than tripled,but successes at scale remain limited.2.Acceptance of AI by R&D teams.The acceptance

242、of AI use by researchers and developers is driven by various motivations and factors,from perceptions of AI performance to ethical concerns about data use and bias,as well as the human fear of being replaced.According to our survey respondents,“user resistance to change”ranks as the third most impor

243、tant obstacle to implementing AI,cited by 50%.However,on average,over 80%of researchers expect benefits from AI across all aspects of R&D&I(including innovation,cost,quality,and speed).Figure 10.Sakanas AI Scientist,an AI-driven system for automated scientific discoveryNote:(1)GPT-4o-based agent to

244、conduct paper reviews based on neural information processing systems conference review guidelines;(2)generated reviews can be used either to improve project or as feedback to future generations for open-ended ideation Source:Arthur D.Little SakanaNote:(1)GPT-4o-based agent to conduct paper reviews b

245、ased on neural information processing systems conference review guidelines;(2)generated reviews can be used either to improve project or as feedback to future generations for open-ended ideationSource:Arthur D.Little SakanaFigure 10.Sakanas AI Scientist,an AI-driven system for automated scientific d

246、iscoveryBrainstorm with a starting code“template”of an existing topic to be exploredGiven idea&template,execute proposed experiments Paper production based on templateReview1 feeding a continuous feedback loop2Main resultsInitial runs generated scientific papers on various topics(e.g.,“diffusion mod

247、eling,”“language modeling,”or“grokking”10-page documents comprising abstract,introduction,related work,background,method,experimental setup,results,conclusion&references“Unexpected”outcomeAI scientist sometimes attempts to boost likelihood of success by altering&executing own scriptsFor instance,if

248、experiments take too long to finish,it would modify its code to extend time-out periodIDEA GENERATIONEXPERIMENT ITERATIONPAPER WRITE-UPLLM idea/plan innovationNovelty check sem.scholarIdea scoring/archivingExperiment templateCode via LLM&aiderExperiment exec scriptManuscript templateText via LLM&aid

249、erLLM paper reviewingManuscriptUpdate planExperimentsNumerical data/plotsStage 1Stage 2Stage 3Stage 4Feedback loop2132449Blue Shift /REPORT 0073.Acceptance of AI by the public.Public acceptance of AI is varied,with benefits around speed and innovation balanced by concerns over ethics,data use/privac

250、y,job losses,and sustainability,given the technologys high energy consumption.Some concerns regarding the negative use of AI,such as deepfakes,bias,or hallucinations,are seeing increased public attention.A 2023 Ipsos survey3 indicates that the public has a“cautious”attitude toward AI,with 54%believi

251、ng AIs benefits surpass its drawbacks.Trust in AI varies widely by region and is generally much higher in emerging markets and among people under 40 than in high-income countries and among Gen Xers and Boomers.Public acceptance of AI matters greatly for publicly funded research organizations(e.g.,pu

252、blic concerns about GMOs had a chilling effect on research in the domain but less so in corporate R&D&I).Affordability1.Generalization of small language models(SLMs).SLMs are ML algorithms trained on much smaller,more specific,and often higher-quality datasets than LLMs.They have far fewer parameter

253、s(usually under 10 billion,compared to over 100 billion for common LLMs)and a simpler architecture.The ongoing rise of SLMs has been fueled by the deployment of open source models from leading technology companies or universities,with around 10“foundation”small models released since 2023(see Figure

254、12).Smaller transformer-based architectures,such as Mistral 7B,Llama 7B,and the Phi family,perform on par with very large models on general language benchmarks(Measuring Massive Multitask Language Understanding MMLU)while being significantly faster,cheaper,easier to fine-tune,and more sustainable in

255、 terms of power consumption.Figure 11.Interpretability landscapeSource:Arthur D.Little,Bereska&Gavves,2024Source:Arthur D.Little,Bereska&Gavves,2024Figure 11.Interpretability landscapeAnalyze input-output Quantify individual input feature influences Identify high-level representations governing beha

256、viorUncover precise causal mechanisms from inputs to outputBEHAVIORALATTRIBUTIONALCONCEPT-BASEDMECHANISTICExample:SHapley Additive exPlanations(SHAP)SHAP values are a method used in ML for explaining the output of a model by attributing each features contribution to the final prediction For instance

257、,GTRgaz uses Shapley value for result interpretation related to AI-driven maintenance analysis activityExample:Sparse Autoencoders(SAEs)SAEs are an unsupervised technique for decomposing activations of a neural network into sum of interpretable componentsMatureExperimentalLevel of maturity50Blue Shi

258、ft /REPORT 0072.Deployment of compute at the edge.Edge computing involves implementing AI algorithms and models on local devices,including smartphones,sensors,or Internet of Things(IoT)devices.This allows immediate data processing and analysis without continuous dependence on cloud infrastructure.In

259、dustry research estimates that the market will expand at+20%p.a.between 2023 and 2030,mainly driven by the IoT.Demonstrating this trend,OpenAI and Apple partnered in June 2024 to integrate ChatGPT into Apple consumer products,while Apples new M4 chip is focused on AI within its iPad Pro tablets.-lab

260、Sweden(in partnership with the ESA)is developing solutions around edge computing and learning in space.Analysis of large datasets aboard a sailing drone is providing marine ecologists with real-time insights into the Baltic Sea ecosystem.3.Popularization of open source hosting and fine-tuning servic

261、es.A dynamic ecosystem of providers is developing around open source models.Coding tools such as PyTorch and TensorFlow and collaborative platforms such as Hugging Face offer decentralized open source libraries,including various model fine-tuning modalities(e.g.,LoRA,explained in Figure 13).Hosting

262、providers offer model inference in the cloud(inference as a service),enabling organizations of all sizes to experiment with AI deployment without investing in costly infrastructure.4.Increasing regulatory constraints.Recent advances in AI in both business and consumer applications have compelled gov

263、ernments and institutions across the world to plan and pass laws and regulations to address the risks associated with the ethical deployment,development,and distribution of AI technologies(see Figure 14).Figure 12.Ongoing rise of small language modelsSource:Arthur D.LittleSource:Arthur D.LittleFigur

264、e 12.Ongoing rise of small language modelsFebruary 2023July 2024March 2023September 2023February 2024April 2024March 2023December 2023May 2023April 2024LlamaVicunaMistralGemmaOpenELMAlpacaOrcaPhi-2Phi-3Llama 30.27B-3B2B,7B7B13B7B,13B8B3.8B,7B2.7B13B7BxModel sizes(#parameters)51Blue Shift /REPORT 007

265、Two major laws have recently been passed in Europe and the US.Some believe the EU AI Act,which came into force on 1 August 2024,imposes unreasonable liability and restrictions on downstream developers,creating a potentially chilling effect on the open source community.US California Senate Bill 1047,

266、signed on 28 August 2024 and vetoed by Governor Gavin Newsom on 30 September 2024,sought to impose strict safeguards on foundation model developers.Its critics cited longer development timelines,lower-performing models,and curtailing of use cases as potential risks.AI regulations,along with other da

267、ta-related laws,such as the GDPR in Europe,are expected to continue to evolve.Figure 13.Low-resource fine-tuning via LoRASource:Arthur D.Little,The Hugging FaceSource:Arthur D.Little,The Hugging FaceFigure 13.Low-resource fine-tuning via LoRA The original equation is output=W0 x+b0,where x is the in

268、put,W0 and b0 are the weight matrix and bias terms of the original dense layer(frozen)The LoRA equation is output=W0 x+b0+BAx,where A and B are the rank-decomposition matrices=0+=W0+BA Pretrained weightsW Merged weightsWmerged B=0A=(0,2)DURING TRAININGAFTER TRAININGFigure 14.AI-related bills passed

269、into law by country/state/regionMETI=Ministry of Economy,Trade and Industry;NIST=National Institute of Standards and Technology;NITI=Indian governments think-tank Source:Arthur D.LittleMETI=Ministry of Economy,Trade and Industry;NIST=National Institute of Standards and Technology;NITI=Indian governm

270、ents think-tankSource:Arthur D.LittleFigure 14.AI-related bills passed into law by country/state/regionNON-EXHAUSTIVECanadaArtificial Intelligence and Data Act(2nd reading)UKAI regulation white paper as a flexible framework(Feb.24)EUArtificial Intelligence Act(Aug.24)JapanGovernance Guidelines for I

271、mplementation of AI Principles(METI)CaliforniaCalifornia Senate Bill 1047(Aug.24)USAI Risk ManagementFramework(NIST)(Jul.24)BrazilAI Regulation(pending review)IndiaResponsible AI(guidelines from NITI Aayog)(Aug.21)ChinaInterim AI Measures is 1st administrative regulation on GenAI(Aug.23)016-251-5No

272、available data6-1011-15Legally-binding regulations 32 countries have enacted at least one AI-related bill as of 2023Number of AI-related bills passed into law by country(20162023)52Blue Shift /REPORT 0075.Cementing of data oligopolies.However,in several domain areas,critical data is owned or control

273、led by players in dominant market positions,such as telecom operators(e.g.,Verizon,AT&T),wearable device suppliers(chiefly Apple),social media companies(Meta,Google),and e-commerce players(e.g.,Amazon).Additionally,some public data may not be usable directly for training models because of concerns a

274、bout fair use,as demonstrated by The New York Times v.OpenAI lawsuit.6.Dissemination of a“data culture”in organizations.Organizations with a long data generation history can leverage it to train R&D&I AI models from new product development to predictive maintenance.The development of a data culture,

275、which encourages data to be systematically collected,labeled,stored,and governed,is critical to enabling AI use cases in R&D&I.Data collection and labeling services reached revenues of around$3 billion in 2023,with an expected CAGR of approximately 25%between 2023 and 2030.This shows organizations i

276、ncreasing appetite for valorizing their own data and their need for outsourced support in ambitious data projects.A paradigm shift could occur for public research organizations,as they would assume new roles as data collectors,wardens,and auditors for scientific purposes.7.Greater consolidation/inte

277、gration of AI providers.Large AI players have been making strategic moves up and down the AI value chain.For example,chipmaker Nvidia has developed foundation models,including Megatron(Megatron-LM),EG3D(Efficient Geometry-Aware 3D),and StyleGAN(image synthesis).OpenAI launched its app store in Janua

278、ry 2024,which enables third parties to create and sell apps through a future revenue-share model.Persistent rumors claim OpenAI is exploring avenues to secure its own compute capabilities and lessen its dependence on Nvidia and Microsoft Azure.Moreover,the current oligopoly on LLMs concentrated on O

279、penAI,Anthropic,Google,and Meta may persist or fail based on the relative success of these companies business models.The future configuration of the market will have important implications for users,such as fewer model options,lock-in effects,and price hikes.It is also uncertain how sustainable the

280、development and training of open source models(a not-for-profit activity)will be for Meta,Nvidia,and Google,as well as,to a greater degree,smaller players such as Mistral.8.Persistent scarcity of talent.The competition for AI-skilled individuals is intensifying,with businesses in various sectors see

281、king to capitalize on AIs potentially transformative capabilities.Demand for trained data scientists is broadly expected to outpace supply up to 2030.Cited by 64%of our survey respondents,“lack of in-house skills”is the most important obstacle to implementing AI,which shows the scale of the issue.Th

282、ese concerns will be somewhat,but not fully,eased by the development of LCNC solutions for fine-tuning and multiple outsourcing offerings for ML operations.Large AI players have been making strategic moves up and down the AI value chain.53Blue Shift /REPORT 007By combining alternative outcomes on al

283、l three dichotomies,we obtained eight possible combinations,of which we have selected six scenarios.4 Each of these sketches a different future for AI in R&D&I,from a fully mature market where everyone benefits from AI(Blockbuster)to a world where AI is affordable yet remains solely used for low-val

284、ue use cases because of other constraints(Cheap&Nasty).Blue Shift /REPORT 00754SIX SCENARIOS FOR THE FUTURE OF AI IN R&D&IBLOCKBUSTER:AI becomes top of mind throughout the R&D cycle,reshaping organizations along the way.Data becomes the new frontier.CROWD-PLEASER:AI is convenient,affordable,and adop

285、ted for daily productivity tasks but falls short of delivering scientific/creative value.CROWN JEWEL:AI delivers productivity and scientific breakthroughs,but only to those organizations that can afford it leading to a two-speed world in R&D&I.55PROBLEM CHILD:Despite some hallmark use cases and affo

286、rdable solutions,AI fails to demonstrate its value R&D&I organizations remain concerned about data security,deontology,and lack of interpretability.BEST-KEPT SECRET:AI performance improves,but high costs make organizations more risk-averse.Low trust and red tape limit adoption.Few new bold experimen

287、ts are launched.CHEAP&NASTY:AI is broadly used in low-stakes use cases,but only as a prototyping or brainstorming tool.Untrustworthy systems are strictly vetted and outputs are verified,curtailing productivity gains.SIX SCENARIOS FOR THE FUTURE OF AI IN R&D&I56“Just because something thinks differen

288、tly from you,does it mean its not thinking?.”Ava,Ex MachinaBlue Shift /REPORT 00757How did we get here?Foundation models improve through investment and competition,driving AI innovation across architectures,while specialized models thrive with open source support and accessible hosting solutions.On

289、the hardware side,efficient GPUs and edge AI enable widespread local and on-device model deployment.Increased transparency and interpretability boost trust in AI for R&D&I tasks.What does day-to-day work look like?AI is everywhere within R&D&I.It automates productivity tasks,enhancing knowledge mana

290、gement and resource planning,while agentic AI and robotics enable fully automated laboratories.AI assists creativity by executing concept designs for developers and innovators.All this results in major scientific discoveries in multiple fields.How do R&D&I organizations evolve?The sector sees large-

291、scale initiatives that focus on meeting the growing need for training data.Public institutions maintain and disseminate databases,adapting scientific communication methods to benefit from AI.On the skills side,new data engineering talent eases pressure on ML operations while organizations build new

292、career paths for R&D&I staff as analyst roles shift toward planning.Winners&losersWinners can access plentiful,well-structured data through various means,although strict data protection laws may hinder progress in certain fields.Those R&D&I departments tackling AI-friendly problems benefit the most,

293、and organizations skilled at scaling from POCs to deployment gain a unique advantage.Scenario 1:BlockbusterAI becomes top of mind throughout the R&D cycle,reshaping organizations along the way.Data becomes the new frontier.Increased transparency and interpretability boost trust in AI for R&D&I tasks

294、.Blue Shift /REPORT 00758How did we get here?AI performance hits a plateau as scaling laws max out for LLMs.However,the industrialization of lower-value functionalities into AI systems with reduced inference costs leads to large-scale adoption.This means R&D&I teams use AI productivity tools,gaining

295、 modest productivity improvements while simple explainability techniques foster understanding of AIs limitations and workarounds.What does day-to-day work look like?Researchers use AI for non-critical tasks and as a low-quality“sanity check.”At the same time,AI augments resource management and custo

296、mer service systems in daily operations,although unwarranted high trust in AI occasionally leads to costly mistakes.Overall,mature AI use cases become widespread,but major scientific breakthroughs remain elusive.How do R&D&I organizations evolve?Organizations rely on commoditized AI solutions for pr

297、ocurement and administrative tasks,with AI deployments driven by IT,digital,or operations departments rather than R&D&I.Pre-trained models and RAG are sufficient to handle popular use cases,limiting new data investments.R&D&I awaits the next breakthrough before it prioritizes AI within scientific st

298、rategy.Winners&losersOrganizations with limited data science capabilities arent at a significant competitive disadvantage as tools are simple and off-the-shelf.However,due to their size,large,complex organizations benefit most from AI-driven reporting and resource management automation,while smaller

299、 organizations experience subtler changes from AI.Strong quality control helps organizations detect AI-related issues before they impact operations.Scenario 2:Crowd PleaserAI is convenient,affordable,and adopted for daily productivity tasks but falls short of delivering scientific/creative value.AI

300、augments resource management and customer service systems in daily operations.Blue Shift /REPORT 00759How did we get here?AI breakthroughs across architectures improve robustness and efficiency,particularly in RLHF and robotics.Advances in interpretability boost researcher trust,but market consolida

301、tion leads to higher prices,and the open source market consolidates with two main players remaining,limiting smaller model offerings.High inference costs and a widening skills gap persist because of GPU limitations.What does day-to-day work look like?Because of its extreme cost,AI usage focuses on a

302、reas with the highest productivity gains.This means that while major AI-assisted discoveries occur,they are mainly in well-resourced organizations.Researchers design to cost,applying AI selectively to complex problems.Increased model interpretability motivates strong trust in AI outputs.How do R&D&I

303、 organizations evolve?Well-resourced organizations invest in on-premise compute capabilities while public-private partnerships form to provide access to affordable AI compute resources.Enriching enterprise software with AI capabilities enables limited productivity gains,and on the skills side,organi

304、zations encourage researchers to train in data science by adapting their incentive structures.Winners&losersA two-speed market emerges,widening the gap between AI-capable and resource-limited organizations.Organizations leveraging AI for operational gains invest more in compute and talent,while acce

305、ss to supercomputers becomes a significant advantage for some R&D&I teams.Mid-and lower-tier teams in AI-friendly fields experience decreased relative influence.Scenario 3:Crown JewelAI delivers productivity and scientific breakthroughs,but only to those organizations that can afford it leading to a

306、 two-speed world in R&D&I.AI usage focuses on areas with the highest productivity gains.Blue Shift /REPORT 00760How did we get here?Foundation models improve through continued investment and competition in various architectures,while efficient GPUs and edge AI enable widespread local and on-device m

307、odel deployment.However,high-profile AI mishaps erode trust,hindering adoption in R&D&I fields.A lack of progress in interpretability research also erodes trust,perpetuating criticisms of AI systems as opaque black boxes.What does day-to-day work look like?Researchers use AI for low-stakes tasks and

308、 content drafting,enabled by cheap enterprise-wide AI solutions.Regarding R&D&I-specific applications,AI assists in select exploration cases with proven track records,such as protein folding.Given limited trust,strict oversight and testing guidelines are enforced for AI system use,with vetting proce

309、dures creating productivity tradeoffs in researchers tool selection.How do R&D&I organizations evolve?Large-scale initiatives focus on meeting growing needs for training data,with public institutions maintaining and disseminating databases and adapting to new data demands.However,the continued need

310、for verification and oversight means junior analyst roles are maintained and bans on AI-generated content lead to the widespread use of detection tools.Winners&losersOrganizations able to balance AI benefits with managing potential risks gain a competitive advantage,and those dealing with AI-friendl

311、y problems and skilled at scaling POCs achieve dominance.Countries with clear regulatory frameworks promote safer AI use while industries align on AI safety standards to reduce perceived risks and uncertainties.Scenario 4:Problem ChildDespite some hallmark use cases and affordable solutions,AI fails

312、 to show its value R&D&I organizations remain concerned about data security,deontology,and interpretability.High-profile AI mishaps erode trust.Blue Shift /REPORT 00761How did we get here?Foundation models improve through investment and competition,driving AI innovation across architectures.Speciali

313、zed models thrive because of open source support and accessible hosting solutions.Efficient GPUs and edge AI enable widespread local and on-device model deployment.Increased transparency and interpretability boost trust in AI for R&D&I tasks.What does day-to-day work look like?AI automates productiv

314、ity tasks,enhancing knowledge management and resource planning,while agentic AI and robotics enable fully automated laboratories.AI leads to major scientific discoveries across various fields and assists creativity by executing concept designs for developers and innovators.How do R&D&I organizations

315、 evolve?Large-scale initiatives focus on meeting growing training data needs,with public institutions maintaining and disseminating databases and adapting scientific communication methods.An influx of new data engineering talent eases pressure on ML operations while organizations build new career pa

316、ths as analyst roles shift toward planning.Winners&losersWinners can access plentiful,well-structured data through various sources,although strict data protection laws may hinder progress in certain fields.R&D&I departments tackling AI-friendly problems benefit the most,as do organizations skilled a

317、t scaling from POCs to deployment,which gain a unique advantage.Scenario 5:Best-Kept SecretAI performance improves,but high costs make organizations more risk-averse.Low trust and red tape limit adoption.Few new bold experiments are launched.AI leads to major scientific discoveries across various fi

318、elds.Blue Shift /REPORT 00762How did we get here?Enterprise AI adoption spreads through aggressive sales tactics and competitive pricing,with efficient GPUs and open source developments enabling widespread model deployment.However,AI performance plateaus,with no significant breakthroughs in new arch

319、itectures and stalled progress in interpretability,all contributing to an“AI winter”and decreased trust levels.What does day-to-day work look like?Researchers use AI for non-critical tasks and as a low-quality“sanity check,”with prompt-engineering skills in demand to compensate for the multiple inte

320、ractions required to achieve results.AI-augmented systems are used for resource management and customer service.Strict oversight and verification procedures limit AIs productivity gains.How do R&D&I organizations evolve?Because AI use still requires extensive verification,no changes are made to anal

321、yst roles,and organizations only benefit from limited productivity gains in terms of support roles,although RAG is becoming popular for knowledge management.AI deployments are driven by the IT,digital,or operations department rather than being led by R&D&I.Efforts are limited around model fine-tunin

322、g for scientific use cases.Winners&losersLarge organizations benefit most from AI-driven reporting and resource management automation,whereas smaller organizations experience little to no change from adopting AI.Organizations that turned their back on AI maintain their focus on other capabilities,an

323、d these AI-skeptical entities build more durable competitive advantages in non-AI areas.Depicting possible futures with scenarios enables organizations to prepare,identify a path forward with no-regret moves,provide a framework for strategic planning,and make strategic bets based on their needs and

324、capabilities,as we will outline in our concluding chapter.Scenario 6:Cheap&NastyAI is broadly used in low-stakes use cases,but only as a prototyping or brainstorming tool.Untrustworthy systems are strictly vetted and outputs are verified,curtailing productivity gains.Organizations that turned their

325、back on AI maintain their focus on other capabilities.Blue Shift /REPORT 0076364CHAPTER64565STRATEGIC ACTIONS5STRATEGIC ACTIONSWe recommend six no-regret moves for organizations regardless of the six future scenarios.These comprise mutualizing compute power,encouraging data sharing,managing AI talen

326、t,training the workforce in AI fundamentals,resetting data and AI governance approaches,and improving output controls.Beyond these,organizations should take measured strategic bets aligned with corporate objectives.Blue Shift /REPORT 00766EFFECTIVE DECISION-MAKING AROUND R&D&I AIIn some situations,A

327、I is already enabling double-digit improvements in time,costs,and efficiency in formulation,product development,intelligence,and other R&D&I tasks,examples of which are shown in Figure 15.Additional gains are likely,but their exact extent will vary greatly depending on which of the six scenarios mat

328、erialize by 2030.1.Blockbuster widespread,profound benefits for all2.Crowd Pleaser widespread but with shallow benefits3.Crown Jewel deep benefits but limited to best-resourced organizations4.Problem Child benefits limited by red tape and lack of trust5.Best-Kept Secret benefits limited to very comp

329、etent organizations6.Cheap&Nasty very limited benefits across the boardFigure 15.Examples of R&D&I benefits achieved through AISource:Arthur D.LittleSource:Arthur D.LittleFigure 15.Examples of R&D&I benefits achieved through AIFORMULATIONCHEMICAL DISCOVERY30%80%15%10MTIME SAVINGS IN DRUG DISCOVERYRE

330、DUCTION IN REQUIRED EXPERIMENTSSAVED IN MANDATORY TESTINGREDUCTION IN FORMULATION FAILURE RATE With AI,we can predict the characteristics of molecules before synthesis,allowing us to focus our efforts on the most promising candidates.We reduce the number of tests needed to achieve the targeted chara

331、cteristics,and thus time.“”PresidentAI START-UP FOCUSED ON CHEMICAL R&DAI guides us in our experiments and tells what experiments to conduct to fill our data gaps in the most efficient ways.With it,we went from 100k experiments to only a few dozen.“”Head of AI internal AI START-UP,LEADING CHEMICALS

332、MANUFACTURERBy integrating AI,we can input specific target characteristics,enabling the system to predict the most effective tests and adjustments needed to refine the formula.We can save 0.5-1M on avoided regulation-related tests.“”Senior R&D executive CONSUMER PACKAGED GOODS GROUPAI has been instr

333、umental in navigating formulation spaces,reducing failure rate of tests from 25%to less than 10%.We react to regulation changes,competitive launches,and new findings very quickly with it.“”Senior Head of R&D PERSONAL CARE PRODUCTS GROUPPRODUCT DEVTECH INTELLIGENCEINDUSTRIALIZATION Source:Arthur D.LittleFigure 15.Examples of R&D&I benefits achieved through AI30%30%10%IMPROVEMENT IN PRODUCTION EFFIC