《未来今日研究所（FTI）：2025年科技趋势报告（第18版）（英文版）（1000页）.pdf》由会员分享，可在线阅读，更多相关《未来今日研究所（FTI）：2025年科技趋势报告（第18版）（英文版）（1000页）.pdf（1000页珍藏版）》请在三个皮匠报告上搜索。

1、2025 TECH TRENDS REPORT18TH EDITIONTABLE OF CONTENTS003 A Note from FTSG004 2025 Tech Trend Reports005 Time of Impact of Trends on Industry006 Executive Summary040 Artificial Intelligence148 Web3213 Metaverse&New Realities283 Biotechnology360 Energy&Climate432 Mobility,Robotics,&Drones498 Computing5

2、69 Built Environment626 News&Information666 Health Care&Medicine735 Financial Services&Insurance782 Space857 Hospitality&Restaurants900 Supply Chain,Logistics,&Manufacturing944 Entertainment988 Authors&Contributors995 About Future Today Strategy Group997 Methodology998 Disclaimer999 Using the Materi

3、al in the Trend Report2 2025 Future Today Strategy Group.All Rights Reserved.Youre reading the 18th annual edition of what was previously known as the Future Today Institutes Tech Trends Report.Our name has changedwere now Future Today Strategy Group(FTSG)but our goal remains constant:connecting for

4、esight to strategy to drive meaningful organizational transformation.This years analysis spans 1,000 pages divided into 15 comprehensive reports.To access the full report or individual sections,visit .3 2025 Future Today Strategy Group.All Rights Reserved.A NOTE FROM FTSGFuture Today Strategy Groups

5、 2025 Tech Trend Report4 2025 Future Today Strategy Group.All Rights Reserved.Our 2025 edition includes 1000 pages,with hundreds of trends published individually in 15 volumes and as one comprehensive report.Download all sections of Future Today Strategy Groups 2025 Tech Trends report at TECH TREND

6、REPORTSTime of impact of trends will vary by industry.5TIME OF IMPACT 2025 Future Today Strategy Group.All Rights Reserved.AgricultureArchitecture,Built EnvironmentAutomotiveAviation,TravelConstruction,EngineeringConsumer Packaged GoodsFinancial Services,BankingGovernment,PolicyHealth Care Systems&S

7、ervicesHospitalityInsurance(P&C)Insurance(Health&Life)Media Media(News)Pharmaceuticals,Medical ProductsRetailSpace,Aerospace DefenseSupply Chain,LogisticsTelecommunicationsLIVING INTELLIGENCECOMPUTING ARCHITECTUREAIMETA-MATERIALSMOBILITYBIO-ENGINEERINGAR/VR/XRROBOTICSADVANCED SENSORSWEB3 INFRASTRUCT

8、URECLIMATE&GREEN TECHQUANTUMSPACE TECHNOW1-3 YRS3-5 YRS5-7 YRS7-10 YRSLOW RELEVANCEEXECUTIVE SUMMARY2025 TECH TRENDS REPORT 18TH EDITION09 Letter From Amy Webb10 10 Key Takeaways11 FTSG Framework12Key Takeaways in Detail13Living Intelligence15 Large Action Models17 Robotics19Agentic AI21 Metamateria

9、ls23 Unlikely Alliances25Climate Innovation27 Nuclear29 Quantum31 Cislunar33Beyond Trends34Trends vs Trendy35 Trends&Uncertainties36 Trends Opportunities37 Trends Threats38 Whats NextTABLE OF CONTENTS7 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY8 2025 Future Today Strategy

10、 Group.All Rights Reserved.EXECUTIVE SUMMARYINTRODUCTIONBeyond the Rubicon:Navigating Humanitys Point of No Return In the past year,humanity crossed multiple points of no return.This didnt happen gradually,but in sudden,irreversible leaps that have fundamentally altered the trajectory of civilizatio

11、n.Weve moved beyond our mental models,beyond biological constraints,beyond social normsinto territory we can neither fully explain nor comprehend.Just as the first telescopes revealed the vastness of space,todays science and tech advances are revealing how much we dont understand about our own poten

12、tial.Yes,AI has made daily headlines,but its just one piece of a larger transformation.Two other areas of technologyadvanced sensors and biotechnologyare quietly advancing and converging as they evolve.That convergence is creating what we call“living intelligence:”systems that sense,learn,adapt,and

13、evolve.Living intelligence will drive an exponential cycle of innovation,acting as an accelerant for technologies that had previously stalled,from quantum computing to robotics.For some organizations,this will unlock unprecedented opportunities in everything from drug discovery to energy production

14、to financial services.For others,it will trigger overwhelming disorientation as they struggle to adapt to change occurring faster than their ability to process it.The gap between leaders and laggards will widen dramatically,not over decades,but months.Let me be clear:The decisions we make in the nex

15、t five years will determine the long-term fate of human civilization.This isnt hyperboleits the sobering conclusion drawn from our best available data.The convergence of tech isnt just changing how we work or live;its changing what it means to be human.Were building systems that can reprogram biolog

16、y,reshape matter at the atomic level,and process information in ways that defy classical physics.The implications extend far beyond quarterly earnings or market share.This report isnt designed to predict the future.Its purpose is to help you navigate it.While individual trends arent useful in isolat

17、ion,when combined with scenario planning and strategic foresight,they become powerful tools for decision-making.In a world that has moved beyond traditional boundaries,the goal isnt to get the future rightits to get your decisions right in the present.Welcome to the beyond.Amy Webb CEO Future Today

18、Strategy Group9LETTER FROM AMY WEBBEXECUTIVE SUMMARYAmy WebbChief Executive OfficerFTSG 2025 Future Today Strategy Group.All Rights Reserved.10KEY TAKEAWAYSEXECUTIVE SUMMARY 2025 Future Today Strategy Group.All Rights Reserved.10 Key Takeaways from the FTSG 2025 Tech Trends Report.Living intelligenc

19、e merges AI,sensors,and biotech into systems that think,adapt,and evolve beyond our grasp.Action models eclipse language models as AI shifts from talking to doing,reshaping automations frontier.Robots finally break free from factory floors as advanced technology enables real-world adaptability.Agent

20、ic AI systems set their own goals and execute complex decisions,augmenting human expertise.Metamaterials rewrite physical limits,as engineered substances transform how we build our world.Tech giants forge unlikely alliances as AIs demands force former rivals to share computing power and data.The cli

21、mate crisis spurs rapid innovation as extreme weather events accelerate next-gen technology adoption.Nuclear power resurges as AIs energy appetite drives tech giants to invest heavily in small modular reactors.Quantum computing reaches its inflection point as error correction breakthroughs unlock pr

22、actical use cases.Private enterprise colonizes cislunar space,birthing an economy between Earth and the moon that reshapes commerce.1234567891011FTSG FRAMEWORKEXECUTIVE SUMMARY 2025 Future Today Strategy Group.All Rights Reserved.This FTSG Framework helps leaders navigate complexity and make strateg

23、ic decisions in a world of rapid change.Use our trends to shape your futures.Strategic Horizon ScanningIdentify which emerging technologies and trends will directly impact your organizations growth and evolution in the next 12-36 months.Organizational ReadinessEvaluate your current capabilities agai

24、nst future requirements.This isnt just about technology adoptionits about assessing if your culture,talent,and processes can adapt to and thrive in rapidly evolving market conditions.Risk&Disruption MappingPlot potential disruptions from both expected and unlikely sources.Rather than traditional ris

25、k assessment,focus on how technological convergence could create unexpected competitive threats or market opportunities.Action PlanningTransform insights into executable strategies.Move beyond traditional strategic planning to create dynamic response frameworks that allow your organization to pivot

26、quickly as technological changes accelerate or decelerate.KEY TAKEAWAYS IN DETAIL12 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYLiving intelligencethe convergence of AI,sensors,and biotechwill create intelligent systems that can perceive,learn,and evolve beyond human progra

27、mming.The Great Tech Convergence is Already HereAIs integration with advanced sensors and biotechnology isnt just another tech trendits the birth of systems that can truly interact with and adapt to the physical world.These technologies are combining to create feedback loops between digital and biol

28、ogical systems,enabling capabilities that would be impossible with any single technology alone.Why Organizations Keep Missing the SignalsMost companies are hyperfocused on AI but are overlooking how sensors and biotechnology will amplify its impact.This myopic view means missing the bigger transform

29、ation:systems that not only process data but actively sense,interpret,and modify their environment in real-time.The next wave of innovation will come from this convergence.13LIVING INTELLIGENCE 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYThe interaction and intersection of

30、these technologies will create compounding effects,pushing the world into a new phase of technological disruption.-Amy Webb,“The Era of Living Intelligence”LIVING INTELLIGENCEArtificial IntelligenceAdvanced SensorsBioengineeringOrganizations that fail to understand and prepare for living intelligenc

31、e systems risk being blindsided by competitors who harness this convergence to create unbeatable advantages.How this shift will reshape business&society through 2030The rise of living intelligence will fundamentally reshape competitive dynamics across industries.Companies that grasp this convergence

32、 early will build systems that can sense market changes,adapt their operations,and evolve their offerings in real-time.This isnt just about automation or efficiencyits about creating organizations that can perceive and respond to opportunities and threats with unprecedented speed and precision.Early

33、 movers will establish data and capability advantages that become nearly impossible for competitors to overcome.How leaders are being influenced by this shift todayOur clients are already experiencing the implications of living intelligence.While most started with narrow AI initiatives,leaders are n

34、ow racing to integrate sensor networks and biological interfaces into their operations.Were seeing health care companies combine AI diagnostics with continuous biometric monitoring,manufacturers deploying adaptive production systems that evolve their processes,and retailers creating environments tha

35、t sense and respond to customer behavior in real-time.14LIVING INTELLIGENCE 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY200 million protein structures the number of proteins in AlphaFold Servers free database.Action models are eclipsing language models as AI shifts from tex

36、t generation to real-world behavior prediction,fundamentally changing how machines learn.From Words to Actions While language models excel at processing text,action models learn from behavioral data captured by ubiquitous sensors.These systems dont just predict what to saythey predict what to do,bre

37、aking complex tasks into executable steps and making real-time decisions based on environmental feedback.The Rise of Personal Action ModelsAs action models evolve,theyll become increasingly personalized,learning from individual behavioral patterns.We believe that PLAMs(Personal Large Action Models)w

38、ill seamlessly manage tasks,negotiate deals,and make decisions based on deep understanding of user preferences,while maintaining privacy through edge computing.15LARGE ACTION MODELS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYMicrosofts work-in-progress LAM started with a t

39、raining dataset comprised of 76,000 task-plan pairs.Ultimately,2,000 successful action sequences were used in the final training set.The shift to action-based AI will create autonomous systems that can execute complex tasks without explicit programming,transforming automation across industries.How t

40、his shift will reshape business&society through 2030Action models represent a fundamental shift in how AI systems operate in the real world.Unlike language models that operate primarily in the realm of text and content generation,action models will enable AI to understand and predict physical behavi

41、ors,movements,and decision-making patterns.This capability will revolutionize everything from robotics to personal assistance to business process automation.As these systems mature,theyll move beyond simple task execution to complex decision-making and strategic planning.How leaders are being influe

42、nced by this shift todayWhile many of our clients were early to invest in LLMs for content generation and customer service,the real transformations in the future will come from LAMs.Leading organizations are already exploring how LAMs could optimize supply chains,predict maintenance needs,and automa

43、te complex operational decisions.The most forward thinking companies in the future will develop hybrid systems that combine language and action models,creating AI that can both communicate and act.16LARGE ACTION MODELS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYBy 2030,mor

44、e than 125 billion connected devices will generate continuous behavioral data,fueling LAMs ability to learn and act autonomously.Robotics will hit an inflection point as AI and advanced sensors enable machines to adapt to unstructured environments and learn complex tasks in real time.The End of Rigi

45、d Robotics Traditional robots were confined to controlled environments,performing repetitive tasks.Now,AI-powered robots can perceive their surroundings,make decisions autonomously,and adapt to changing conditionsmarking the transition from programmed to intelligent automation.Why Scale is Finally P

46、ossibleThe convergence of AI,advanced sensors,declining hardware costs,and edge computing has removed historical barriers to robotic deployment.Combined with improving ROI metrics and labor shortages across industries,these advances are creating perfect conditions for widespread adoption.17ROBOTICS

47、2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYAI-enabled robots that pick and place different parts and materials in our fully automated assembly lines reduce automation costs by 90%.-Stephan Schlauss,Global Head of Manufacturing,Siemens AGAdaptive robotics will transform ind

48、ustries far beyond manufacturing,creating new operational paradigms in health care,agriculture,and construction.How this shift will reshape business&society through 2030Robotics will expand beyond traditional industrial applications into more complex,human-centric environments.In health care,surgica

49、l robots will enhance human capabilities;in agriculture,autonomous systems will enable precision farming;in construction,robots will perform dangerous or repetitive tasks.At least initially,this shift wont replace human workers but augment them,creating new roles focused on robot supervision and str

50、ategic decision-making.How leaders are being influenced by this shift todayManufacturing leaders we advise are rapidly reevaluating their automation strategies as adaptive robots become more viable.Were seeing health care executives explore robotic surgical assistants that could triple procedure eff

51、iciency,while construction firms are piloting autonomous equipment for site preparation and basic assembly.However,most organizations are struggling with integration challenges and workforce concerns.The most successful deployments focus on augmenting human capabilities rather than replacing workers

52、.18ROBOTICS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYThe convergence of advanced sensors and AI will increase robotic autonomy by more than 60%.(Boston Dynamics)Agentic AI marks the transition from passive tools to autonomous systems that can set goals,make decisions,and

53、 execute complex strategies independently.The Rise of AI That Acts Beyond pattern recognition and prediction,agentic AI systems can understand context,formulate strategies,and take independent action.These systems dont just respond to commandsthey identify opportunities,set objectives,and orchestrat

54、e resources to achieve them.Multi-Agent Collaboration Changes EverythingThe real power emerges when multiple AI agents work together,each specializing in different tasks while coordinating toward common goals.This creates networks of AI systems that can handle complex,interconnected challenges that

55、would overwhelm single agents.19AGENTIC AI 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY72%of enterprises using AI agents achieve business process efficiency gains.(Stanford HAI Survey,2024)Organizations must prepare for a world where AI systems make and execute decisions au

56、tonomously,fundamentally altering business operations.How this shift will reshape business&society through 2030Agentic AI will transform how organizations operate,moving from human-directed automation to AI-orchestrated autonomy.These systems will manage supply chains,optimize resource allocation,an

57、d coordinate complex business processes with minimal human oversight.The shift will be gradual but profoundstarting with discrete business functions before expanding to cross-functional operations.Success will depend on building trust,establishing clear governance,and creating new frameworks for hum

58、an-AI collaboration.How leaders are being influenced by this shift todayWhile executives recognize Agentic AIs potential,most struggle with implementation challenges.Leading organizations are starting small,deploying autonomous agents in controlled environments like inventory management or predictiv

59、e maintenance.Were seeing increased concern about security,compliance,and control as these systems become more autonomous.The most successful companies are investing heavily in training,governance frameworks,and change management to prepare their organizations for this transition.20AGENTIC AI 2025 F

60、uture Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYAI-powered agents could automate80%of coding tasks by 2030.(MIT CSAIL)Metamaterials are revolutionizing construction and manufacturing,creating substances with properties that transcend natural limitations.Natures Rules Are Being Rewrit

61、tenMetamaterials,designed at the microscopic level using advanced tech,can manipulate light,sound,heat,and mechanical stress in ways previously impossible.These engineered substances represent a fundamental shift from simply discovering materials to designing their properties from scratch.From Theor

62、y to Commercial RealityAI has accelerated metamaterial development from theoretical models to practical applications.What required decades of research can now be simulated and optimized in hours,enabling rapid prototyping and commercialization of materials with unprecedented capabilities.21METAMATER

63、IALS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYIn metamaterials,we go beyond natural arrangements to a new level of organization.We design our own collection of tiny structures using multiple materials(such as gold and glass).Like regular materials,the electromagnetic pro

64、perties of metamaterials depends on how we shape and arrange these structures.The difference is,we can create new properties that are not found in nature.And thats what metamaterials are all about.-Dr.Nader Engheta,H.Nedwill Ramsey Professor,School of Engineering and Applied Science at University of

65、 Pennsylvania School of Arts and SciencesMetamaterials will transform the built environment,enabling self-cooling buildings,ultra-resilient infrastructure,and adaptive structures.How this shift will reshape business&society through 2030Metamaterials will revolutionize industries from construction to

66、 energy to telecommunications.Buildings will regulate their own temperature,infrastructure will adapt to environmental stresses,and communication systems will achieve unprecedented efficiency.The technology will be crucial for climate resilience,enabling structures that can withstand extreme weather

67、 while dramatically reducing energy consumption.This shift will create new design paradigms and force industries to rethink traditional approaches.How leaders are being influenced by this shift todayConstruction and engineering executives are scrambling to understand metamaterials implications for t

68、heir industries.While some view the technology as distant,leading firms are already forming partnerships with metamaterial startups and research institutions.Were seeing increased investment in R&D and pilot projects,particularly in energy efficiency and structural resilience.However,most organizati

69、ons still lack the expertise to evaluate and implement these new materials.22METAMATERIALS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYAcoustic metamaterials can reduce sound transmission by up to 94%,enabling quieter buildings,aircraft,and industrial environments.(Boston U

70、niversity)Tech giants are forming unprecedented partnerships as AIs massive computational demands force former competitors to share resources and infrastructure.Competition Gives Way to Coopetition The sheer scale of AI developmentfrom computing power to specialized hardwarehas made going it alone i

71、mpossible.Even the largest tech companies are finding they must collaborate with rivals to remain competitive and innovative.The Cloud Becomes the New Battleground As AI workloads grow exponentially,control of cloud infrastructure becomes crucial.Strategic alliances between cloud providers,chip manu

72、facturers,and AI companies are creating new power dynamics that will reshape the tech landscape.23UNLIKELY ALLIANCES 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYTech giants and sector leaders have a synergistic relationship:industry expertise helps make technology advanceme

73、nts actionable,and industry leaders cant advance without new computational and AI capabilities.The era of tech companies operating in isolation is ending,as AIs demands create complex networks of interdependent partnerships.How this shift will reshape business&society through 2030These strategic all

74、iances will fundamentally alter how technology is developed and deployed.Cross-company collaboration will become the norm,with shared infrastructure,data,and research accelerating innovation.However,this consolidation raises concerns about market concentration and competition.Organizations will need

75、 to navigate complex partnership networks while maintaining their competitive advantage.How leaders are being influenced by this shift todayWeve observed that business leaders are challenged by a transformed vendor landscape where traditional competition lines blur.Many are finding their strategic p

76、lanning complicated by uncertain alliances and shifting partnerships.While some embrace multi-vendor strategies to maintain flexibility,others are forming deeper partnerships with specific tech ecosystems.The most sophisticated organizations are creating partnership strategies that balance access to

77、 innovation with vendor lock-in risks.24UNLIKELY ALLIANCES 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYAmazon committed up to$4 billion to support Anthropics AI research,embedding its Claude models into AWS infrastructure.Extreme weather events are accelerating technologica

78、l innovation as climate adaptation becomes an urgent business imperative across every industry.Crisis Drives Commercial Breakthroughs Climate disasters are forcing rapid advancement in resilience technologies.What began as defensive measures is evolving into new markets for climate adaptation,spanni

79、ng infrastructure,agriculture,and emergency response systems.Smart Systems Reshape Climate ResponseThe convergence of AI,sensors,and biotechnology is enabling unprecedented capabilities in climate prediction,response,and adaptation.These technologies are creating early warning systems and resilient

80、solutions previously thought impossible.25CLIMATE INNOVATION 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY82%of investors believe that publicly held financial services companies that better anticipate environmental risks are more likely to succeed financially.(Harvard Busine

81、ss Review)Organizations must integrate climate adaptation into their core strategy as extreme weather reshapes markets and creates new business imperatives.How this shift will reshape business&society through 2030Climate adaptation technologies will become central to business operations and infrastr

82、ucture development.Advanced materials will protect against extreme conditions,while emerging tech will optimize resource usage and predict environmental risks.Biotechnology breakthroughs will create climate-resistant agriculture and carbon-capture solutions.Organizations that fail to adapt will face

83、 increasing operational disruptions and market disadvantages.How leaders are being influenced by this shift todayCorporate leaders are shifting from viewing climate technology as a compliance issue to seeing it as a strategic necessity.Were seeing increased investment in resilient infrastructure,AI-

84、powered climate modeling,and a host of other solutions.Leading organizations are integrating climate adaptation into their core business strategies,while others struggle to balance short-term pressures with long-term climate resilience needs.26CLIMATE INNOVATION 2025 Future Today Strategy Group.All

85、Rights Reserved.EXECUTIVE SUMMARYBy 2050,climate change could put$26 trillion in global financial assets at risk,forcing central banks to integrate climate risk into monetary policy.(IMF)Nuclear powers revival will reshape energy markets and corporate strategy as techcompanies become major players i

86、n power generation.Big Tech Drives Nuclear Renaissance Tech companies are bypassing traditional utilities to invest directly in nuclear power.The push for reliable,carbon-free energy to power AI systems is making nuclear innovation a Silicon Valley priority.SMRs Change the Nuclear Equation Small mod

87、ular reactors offer a new paradigm:scalable,safer,and faster to deploy than traditional nuclear plants.Their standardized design and reduced complexity are transforming nuclear powers risk-reward profile.27NUCLEAR 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYMicrosofts new n

88、uclear plant at Three Mile Island is expected to open in 2028 and will be renamed the Crane Clean Energy Center.The plant will power Microsofts data centers.Small modular reactors emerge as tech giants answer to AIs massive energy demands,marking nuclear powers transformation from pariah to savior.H

89、ow this shift will reshape business&society through 2030The rise of SMRs could democratize nuclear power,enabling new deployment models beyond traditional utility structures.Tech companies will emerge as major energy producers,potentially disrupting traditional utility markets.This shift will accele

90、rate the transition to carbon-free energy while raising new questions about power generation control and infrastructure security.How leaders are being influenced by this shift todayEnergy-intensive industries are closely watching tech companies nuclear initiatives.Many are reevaluating their power s

91、trategies,considering direct investment in SMRs or partnerships with nuclear developers.Some organizations are now developing comprehensive energy strategies that include nuclear as part of their sustainability and operational resilience plans.However,concerns about public perception and regulatory

92、uncertainty remain.28NUCLEAR 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYSmall modular reactors can be manufactured in factories and deployed within 3-5 years,accelerating nuclear adoption.Quantum computing reaches its inflection point as error correction breakthroughs and

93、hybrid systems bring practical applications within reach for the first time.Error Correction Changes Everything After decades of theoretical promise,quantum error correction breakthroughs are finally enabling stable qubit operations.This fundamental advance removes the key barrier that has held quan

94、tum computing back from practical applications.Hybrid Systems Bridge the Gap The integration of quantum and classical computing systems is creating immediate value,even before full quantum advantage.Organizations can begin capturing benefits while the technology continues to mature.29QUANTUM 2025 Fu

95、ture Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYFrom AWS Ocelot to Microsofts Majorana 1,the focus is quickly shifting from AI chips to quantum computing chips,indicating another step closer to commercial viability.Organizations must prepare for quantums impact on encryption,optimizat

96、ion,and simulation as the technology moves from research labs to real-world deployment.How this shift will reshape business&society through 2030Quantum computing will revolutionize fields requiring complex simulations and optimization,from drug discovery to financial modeling.Early applications will

97、 focus on specific use cases where quantum offers clear advantages,gradually expanding as the technology matures.Organizations must balance preparation for quantums transformative potential with realistic expectations about implementation timelines.How leaders are being influenced by this shift toda

98、yWhile most executives acknowledge quantums long-term opportunity,they struggle with strategic importance and timing their investments.Some organizations are building quantum literacy,identifying potential use cases,and developing quantum-safe security protocols.The most sophisticated companies are

99、already experimenting with hybrid quantum-classical systems,gaining practical experience while preparing for quantums broader impact.30QUANTUM 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYQuantum algorithms could cut energy grid inefficiencies by 20%,saving billions annually

100、.(Siemens Quantum Energy)Private enterprise is colonizing the space between Earth and the moon,creating a new economic frontier that will reshape commerce and resource extraction.Space Infrastructure Goes Commercial The privatization of space is moving beyond launches to include orbital manufacturin

101、g,refueling stations,and maintenance services.This emerging infrastructure network will enable sustainable operations throughout cislunar space.New Resources Drive New Markets The discovery of lunar water ice and rare minerals,combined with zero-gravity manufacturing capabilities,is creating unprece

102、dented economic opportunities.Space resources will transform industries from pharmaceuticals to semiconductors.31CISLUNAR 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY$1.8 billionEstimated size of the space economy by 2035(World Economic Forum)The cislunar economy will creat

103、e new industry leaders as space capabilities become critical for competitive advantage across sectors.How this shift will reshape business&society through 2030The commercialization of cislunar space will extend Earths economic sphere to lunar orbit.In-space manufacturing will enable the production o

104、f materials impossible to create under gravity,while lunar resources will reduce dependence on terrestrial mining.This expansion will create new logistics networks,insurance markets,and financial instruments.Organizations that establish early positions in this economy will gain significant advantage

105、s.How leaders are being influenced by this shift todayWhile space remains a frontier market,forward-thinking executives are already developing cislunar strategies.Manufacturing companies are exploring zero-gravity production possibilities,while logistics firms plan for orbital supply chains.However,

106、most organizations struggle to evaluate space opportunities against terrestrial investments.Leading companies are forming partnerships with space startups to gain early access to these capabilities.32CISLUNAR 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYMore than$100 billion

107、 is being invested by private companies and national space agencies in cislunar infrastructure.33 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYBEYOND TRENDSMany organizations confuse whats trendy with true strategic trends.Trends are measurable changes occurring over time,no

108、t buzzy headlines or viral tech fads.Yet most organizations chase shiny objects while missing real strategic signals.This problem compounds when companies only track trends in their own industry,which blinds them to the powerful convergences happening at the intersections.True disruption rarely emer

109、ges from a single trend.It comes from the collision of multiple forces across different domains.By tracking the right trends consistently and understanding their broader implications,any organization can develop stronger foresight capabilities and make better strategic decisions.34TRENDS VS TRENDY 2

110、025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYUnderstanding the difference between trends and uncertainties shapes better strategic decisions.Trends are what we can know Measurable changes occurring over time,backed by data and research.Observable patterns that show consistent

111、 movement in a specific direction.Developments that can be tracked,quantified,and validated through evidence.Uncertainties are what we cannot know Future conditions that defy precise prediction or measurement.Variables that could develop in multiple different directions.Events whose outcomes remain

112、unknown despite careful analysis.35TRENDS&UNCERTAINTIES 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYOur research uncovers what others are missing,such as:Topological qubits could finally solve quantum computings stability problem.Microsofts breakthrough shows a path to scal

113、able quantum systems without the massive error correction overhead.Spatial computing will transform workplace collaboration,enabling seamless integration of virtual and physical spaces.Early movers will reshape how teams interact and solve problems.Generative AI is revolutionizing how robots learn,c

114、ombining sensor data,human demonstrations,and internet-scale training.This breakthrough will finally make robots adaptable enough for real-world deployment.Nanotech breakthroughs in materials science will enable self-healing infrastructure and smart surfaces.Researchers can develop products with unp

115、recedented properties and capabilities.Edge computing combined with 5G will enable real-time processing at unprecedented scale.Organizations can deploy new solutions anywhere,creating truly distributed operations.Synthetic biology will create new possibilities for sustainable manufacturing and carbo

116、n capture.Teams can redesign supply chains and create innovative climate solutions.GenAI-powered search is evolving from link lists to conversational answers.Marketers can transform how they reach customers by optimizing for this new paradigm of discovery and engagement.Augmented reality will transf

117、orm customer experiences and worker training.HR leaders can create immersive interfaces that blend digital and physical worlds.Smart materials will enable adaptive infrastructure and self-optimizing systems.Designers can create buildings and products that respond to environmental changes.36TRENDS OP

118、PORTUNITIES 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARYTrends reveal transformative opportunities,but only if you ask,“What if?”These trends also create new vulnerabilities.Heres where your organization might be most exposedSmall language models make AI deployment easier a

119、nd cheaper,creating security blind spots as departments bypass IT to implement their own solutions without proper oversight.Synthetic biology democratizes biotech capabilities,but also creates new risks.Safety protocols designed for traditional threats cant handle these emerging biological hazards.P

120、rivacy regulations cant keep pace with AIs data analysis capabilities.Legal teams using traditional compliance frameworks will leave organizations exposed to new forms of liability.Spatial computing creates new attack surfaces in physical spaces.Facility managers and security teams lack protocols fo

121、r securing augmented and mixed-reality environments.Decentralized systems challenge traditional governance structures.Risk managers using centralized control models will struggle as operations become more distributed.Computer vision advances enable unprecedented surveillance capabilities.Ethics boar

122、ds lack frameworks to address the responsible use of these powerful monitoring tools.Deep learning models can make critical decisions without clear audit trails.Compliance teams cant explain AI decisions using traditional accountability frameworks.Cross-platform data flows create new vulnerabilities

123、.IT teams focused on securing individual systems miss risks in the interconnections between emerging technologies.Fast-learning robots could make entire skill sets obsolete overnight,forcing rapid workforce transitions.HR teams arent prepared for this pace of role displacement and retraining.37TREND

124、S THREATS 2025 Future Today Strategy Group.All Rights Reserved.EXECUTIVE SUMMARY Organizational leadership should embed foresight into strategy by regularly assessing tech disruptions and aligning long-term vision with emerging trends.Require tech literacy at the board level to ensure informed decis

125、ion-making on disruptive innovations.Allocate capital for innovation,balancing short-term returns with long-term investments in emerging technologies.Integrate scenario planning and strategic foresight into annual planning to anticipate volatility and new opportunities.Monitor weak signals and track

126、 emerging tech,geopolitical shifts,and societal trends to anticipate disruptions early.We recommend that every organization do the following now.To thrive in a world of rapid change,organizations must be agile,resilient,and future-ready.38WHATS NEXT 2025 Future Today Strategy Group.All Rights Reserv

127、ed.EXECUTIVE SUMMARY Strengthen infrastructure by upgrading networks,cloud systems,and cybersecurity to handle rapid shifts in technology.Develop cross-industry partnerships and collaborate beyond traditional sectors to drive innovation and expand market reach.Expand global intelligence capabilities

128、 to track geopolitical,economic,and tech shifts to anticipate disruptions and opportunities.Adopt agile governance and implement flexible policies that can evolve with emerging technologies and global uncertainties.Develop experimental sandboxes to test emerging tech,fostering a culture of rapid pro

129、totyping and iteration.ARTIFICIAL INTELLIGENCE2025 TECH TRENDS REPORT 18TH EDITION44 Letter From the Authors46 Top 5 Things You Need to Know47 State of Play49 Key Events Past50 Key Events Future51 Why Artificial Intelligence TrendsMatter to Your Organization52 When Will Artificial Intelligence Trend

130、sDisrupt Your Organization?54 Pioneers and Power Players56 Opportunities and Threats57 Investments and Actions to Consider58 Important Terms61 Artificial Intelligence Trends62Models,Techniques,and Research63Generative AI Modalities Expand63 Fine Tuning64 Automated Reinforcement Learning64 Evolutiona

131、ry Composition64 Mixture of Experts65 Autonomy-of-Experts66 LLMs as Operating Systems66 LLMs:Bigger and More Expensive66 Chain-of-Thought Models 68 Small Language Models68 Grounding and Context Augmentation68 Overcoming the Data Shortage69 Open-Source AI69 Modular AI 70 Large Action Models70 Persona

132、l Large Action Models72Safety,Ethics,and Society73Explainable AI(XAI)73 AI Optimization73 Decentralized AI Alignment74 Mission Drift in AI Alignment74 Indexing Trust74 Realtime Deepfake Detectors75 Watermarking75 Child Safe AI76 Politically Biased AI76 AI as a Tool to Address Political Bias76 Gender

133、 and Race Biased AI77 Nefarious AI Misuse78 Data Poisoning:A Double-Edged Sword78 Citizen Surveillance78 Worker Surveillance79 School Surveillance79 Posthumous AI80 Privacy Risks in Behavior Biometrics81AI and Energy82Resource-Hungry AI82 AI Nuclear Renaissance 82 Efficient AI Architectures83 Effici

134、ent AI Algorithms84 Energy Optimization85AI Geopolitics,Defense and Warfighting87AI Nationalism87 The AI-Driven Chip War88 AI Diplomacy88 Tech Pivots on Defense89 Autonomous Weapons Policies89 Automated Target Recognition and AI-Guided Strikes90 AI-Assisted Humanitarianism in War90 AI-Assisted Situa

135、tional Awareness90 AI as a Shield91 Simulating Warfare91 AI in Cyber Defense92 Policy and Regulation94 United States:Accelerating AI Fast 95 European Union:Driving Hard on AI Governance41TABLE OF CONTENTS 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE96 China:State-Dire

136、cted Strategy and Tight Oversight97 Brazil:On the Path to AI Legislation98 United Arab Emirates:Balancing Innovation with Guidelines99 Emerging Capabilities100 AI in Mathematics100 Computer-Using Agents101 AI Reasoning101 AI-to-AI Communication102 Detecting Emotion102 Embodied Agents103 Neuro-symbol

137、ic AI104 Human-AI Interactions105 AIs Persuade Humans105 Humans Persuade AI106 Prediction and Prescience into our Human Lives106 On-Device AI107 Wearable AI107 Generative User Interfaces108 The Business of AI 109 Vertical Integration From Hardware to LLMs110 Pricing Bifurcation110 Optimizing AI to R

138、un On and For the Edge110 The AI Training Data Market111 AI Breathes life into Legacy Systems112 Talent and Education113 AI Brain Drain from Academia113 AI Education Surge113 AIs Two Speed Economy114 Agents:From Assistants to Actors114 Complementary Work115 AI-Assisted Education115 AI Native Educati

139、on116 Creativity and Design117 GAN-Assisted Creativity117 Neural Rendering117 Generating Virtual Environments 118 AI as a Content Medium118 AI Democratizes Music Production118 Automatic Ambient Noise Dubbing119 AI-Assisted Invention120 Industries121 Pharmaceuticals121 Protein Folding121 AI-First Dru

140、g Development122 Generative Antibody Design122 NLP Algorithms Detect Virus Mutations123 Health Care123 AI-Assisted Diagnosis and Clinical Decision-Making123 Anomaly Detection in Medical Imaging123 AI-Empowered People124 Health Care-Specific LLMs124 Medical Deepfakes126 Science126 Multistep Scientifi

141、c Reasoning126 AI-Driven Hypotheses126 AI-Driven Experimentation127 AI-Powered Analysis and Interpretation127 AI to Speed Up New Materials Development128 Animal Decoding129 Finance129 AI Assisted Asset Pricing and Management129 Mitigating Fraud129 Predicting Financial Risk130 Customized Portfolios13

142、0 Consumer-Facing Robo-Advisers131 Insurance131 Predicting Workplace Injuries131 Improving Damage Assessment131 AI Powered Fire Prevention42TABLE OF CONTENTS 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE132 The Connected Worker132 Liability Insurance for AI133 HR133 Au

143、tonomous Talent Acquisition133 AI Onboarding and Integration133 Employee Engagement and Retention134 Benefits Selection and Management135 Marketing135 AI Shifts Search135 Dynamic Engagement Through Deep Personalization135 AI-Assisted Campaigns136 Anecdotal Observations,Now Usable Marketing Data137 A

144、uthors&Contributors140 Selected Sources43TABLE OF CONTENTS 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEAIs bleeding edge now changes by the hour,not year.What next?AI is moving at breakneck speed,reshaping industries,workflows,and everyday life faster than we can docu

145、ment.On the day we wrote this,people were still breathlessly marveling at Chinas DeepSeek,which achieved OpenAIs top-tier performance with a fraction of the usual price tag and computing powerchallenging everything we thought we knew about what it takes to build advanced AI.Hours later,researchers a

146、t Stanford and the University of Washington debuted yet another new model,s1,which outperformed both DeepSeeks R1 and OpenAIs o1 reasoning models using even fewer resources.Thats the nature of AI right now:Whats bleeding-edge today might be old news later today.Heres what we know for certain.Last ye

147、ar,OpenAI CEO Sam Altman met with sovereign wealth fund managers and investors,hoping to raise up to$7 trillion for an AI chip company.In January 2025,Stargate,a newly formed joint venture between OpenAI,Oracle,and SoftBank,said it would raise$500 billion for chips,AI data centers,and their massive

148、power requirements.Not to be outdone,Microsoft,Meta,and Google have each announced plans to invest hundreds of billions of dollars in AI infrastructure.But if now anyone can replicate a multimillion dollar model with only modest resources,wont AI models quickly become commoditized?If so,this would p

149、ressure Big Tech to move very fast,building and scaling ever-advancing AI systems in order to stay competitive in the market.44Amy WebbChief Executive OfficerSam JordanTechnology&Computing LeadLETTER FROM THE AUTHORS 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEBreakth

150、rough advancements have also accelerated the AI race between the US and China,intensifying it into a full-blown geopolitical contest,with both nations leveraging technology as a tool for global influence.To wit:during Donald Trumps globally televised inauguration,execs from Americas biggest technolo

151、gy companies sat directly behind himwhile his cabinet appointees and family members sat in rows farther back.This US-China rivalry is forcing allies to take sides,escalating tensions and fueling concerns over national security,supply chains,and technological sovereignty.What emerges is a fragmented

152、AI landscape,dominated by a Digital Cold War that threatens to reshape global alliances and economic power structures.Howexactlywill AI reshape our world in the coming months?The honest answer is:nobody can know.At this stage,avoiding costly mistakes,and smart planning for the future,matters more th

153、an predicting exact outcomes.Leaders need a strategic compass,not a crystal ball.Thats the purpose of this trend report:to highlight emerging AI trends and use cases so you can plan for multiple possibilities.Because in an AI landscape moving at warp speed,strategic clarity is your competitive edge.

154、In the race to win AI,critical evaluation has become a casualty of speed.We meet regularly with the research teams building SOTA models,heads of frontier labs working to advance AI,and executives at the big tech giants.While we are certainly excited about the incredible technological progress being

155、made,there is the practical reality of organizational readiness.The makers of AI systemsand the professional service firms promising overnight transformationare operating in a reality far removed from everyday organizations.What weve observed in the past year advising CEOs and their management teams

156、 on AI strategy and implementation is that regardless of AIs tantalizing developments,most organizations face substantial technical debt in data standardization and maintenance,creating operational friction in deployment.They are also struggling with the basics of change management,which is often de

157、prioritized(or forgotten entirely)ahead of implementation.As a result,we are seeing new strategic risks for organizations that overindex on technological readiness without addressing fundamental operational and cultural barriers to deployment.45LETTER FROM THE AUTHORS 2025 Future Today Strategy Grou

158、p.All Rights Reserved.ARTIFICIAL INTELLIGENCEExpect a continued frenzy of activity as AI companies compete for market share,though investments and policies will concentrate influence among several key players.Harder,better,faster,stronger DeepSeeks R1,and s1 from Stanford and the University of Washi

159、ngton achieved strong reasoning capabilities while remaining cost-efficient,challenging the convention that progress requires ever-larger models and raising questions about future AI scalability.Your AI now has eyes and earsRecent advancements in multimodal AI,like Google Gemini Live and OpenAIs Sor

160、a,are quickly transforming how machines process and generate text,audio,and video,unlocking new possibilities for richer and more interactive AI experiences.Learning how to thinkAI performance is improving,elevating models like OpenAIs o1 and Googles Gemini 2.0 Flash Thinking Mode from mere informat

161、ion engines to thought partners.In late 2024,OpenAIs o3 scored 85%on the ARC-AGI benchmark,matching the average human score.From assistance to autonomyAgentic AI is evolving from supporting tasks to autonomously reasoning and taking action across workflows.This year,AI agents will not only assist bu

162、t also execute complex processes,transforming industries with greater efficiency and automation.US and China race aheadThe US and China are locked in a high-stakes competition for AI dominance,shaping the future of technology and global power.As both nations invest in AI research,infrastructure,and

163、regulation,they are redefining innovation,security,and economic influence.46TOP 5 THINGS YOU NEED TO KNOW 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE13452Its no secret that AIs landscape has transformed dramatically since we wrote the State of Play section last year.

164、GPT-4 set early benchmarks with its multimodal capabilities and professional-level performance,but since then,Google DeepMinds Gemini Ultra raised the bar further,exceeding GPT-4 on most benchmarks.The field has split between proprietary and open-source approaches.Metas release of Llama 2 sparked an

165、 open-source revolution,with developers rapidly fine-tuning variants that rival larger commercial models.This success prompted even OpenAIs CEO Sam Altman to admit that the company might have been“on the wrong side of history”regarding closed systems.Cloud partnerships have become crucial.Microsoft

166、bet big on OpenAI($10B),while Amazon and Google split their support for Anthropic($4B and$2B respectively).These alliances provide AI companies with massive compute power while securing cloud providers positions in the AI race.However,this concentration of resources raises concerns about market cons

167、olidation.China has emerged as a formidable AI power.Baidus Ernie 4.0 claims GPT-4-level performance,while Alibaba released more than 100 open-source models under Qwen 2.5.ByteDances Doubao chatbot gained significant market share.Also making huge strides are Zhipu AI,MiniMax,Baichuan Intelligence,Mo

168、onshot,StepFun,and 01.AIcollectively known as the countrys“Six Little Tigers.”Despite US chip restrictions,Chinese firms are adapting with domestic alternatives like Huaweis Ascend and Baidus Kunlun chips.Investment has exploded,with more than$22 billion flowing to generative AI startups last year a

169、lone,representing nearly half of all AI funding.Traditional VCs are competing with tech giants throwing billions at AI,driving valuations skyward.Artificial intelligence will fundamentally rewire dynamics and competitive frontiers in 2025.47STATE OF PLAY 2025 Future Today Strategy Group.All Rights R

170、eserved.ARTIFICIAL INTELLIGENCEWe see six macro themes emerging:Big Tech will continue to dominate funding,often through strategic partnershipsValuations are reaching dot-com era levels,raising legitimate bubble concernsTraditional tech sectors are seeing funding dry up as AI sucks the oxygen out of

171、 the roomInvestment is flowing directly to cloud providers for compute powerCompetition is intensifying between proprietary and open-source modelsChina is rapidly closing the gap with Western AI capabilitiesLooking ahead,funded AI companies must prove real business value.The winners will likely be t

172、hose that can balance innovation with sustainable monetization while navigating increasing regulatory scrutiny.48STATE OF PLAY 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE123456JANUARY 2025 Nvidia enters the AI PC market Nvidia unveils its Project DIGITS,a personal AI

173、 supercomputer.JANUARY 2025DeepSeek DisruptsChinese AI company DeepSeek releases R1,its reasoning model and competitor to OpenAIs o1.JANUARY 2025 Tech Giants Introduce 500B AI planPolitical,tech,and financial leaders announce Stargate,a joint venture that aims to invest$500 billion in US AI infrastr

174、ucture.DECEMBER 2024O3 Closes In on AGIOpenAI says its o3 model has passed the ARC-AGI challenge,considered a leading benchmark for artificial general intelligence.AI giants raced to AGI as Chinese rivals proved formidable.49PASTMAY 2024OpenAI Launches GPT-4oThe new AI model is capable of real-time

175、reasoning across audio,visual,and text inputs.KEY EVENTS PAST 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEMARCH 2025Huang to Preview Next-Gen AI ChipsNvidias GPU Technology Conference will feature CEO Jensen Huangs keynote on what to expect from this critical chip man

176、ufacturer.MAY 2025 Nvidias Growth Is TestedNvidia could show record-breaking revenueunless DeepSeek portends an alternative future requiring fewer chips.JUNE 2025Apple Bets Big on a Smarter SiriIn a“make or break”moment for proving AIs necessity in consumer products,Apple is poised to debut a new Si

177、ri with generative AI.JULY 2025China Shows Its AI HandThe World AI Conference will showcase Chinas latest AI innovations and initiatives.MAY 2025AI IntegratesGoogle I/Os new generative AI updates will signal how AI will further integrate into consumer services and enterprise tools.Tech titans will f

178、ace pivotal AI tests in 2025.50FUTUREKEY EVENTS FUTURE 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEBeyond the AI hype,these six structural changes are already determining which organizations will thrive.51Speed Is the New ScaleFor better or worse,AI is compressing dec

179、ision cycles from weeks to minutes.The advantage is shifting to organizations that can harness AI for rapid experimentation and learningas long as theyre making good decisions about data governance,vendor selection,and change management.Your Competitors Wont WaitSomeone in your industry is already u

180、sing AI to cut costs and boost productivity by 30%40%.AI is automating certain labor intensive knowledge tasks,but it will soon lead to new workflows and business models.The question isnt whether your organization will adapt to AI,but whether youll do it before or after your margins get squeezed fro

181、m all directions.The Middle Office Is MeltingAI is automating coordination and decision-making tasks that traditionally required human middleware,and organizations clinging to manual coordination and approval processes will find themselves structurally uncompetitive.The future org chart is flatter a

182、nd faster.The Talent War Has New RulesThe best talent now expects to work with the best tools.Theyre not just looking for good paythey want AI-enabled workplaces that multiply their impact.Your ability to attract and retain top performers increasingly depends on your AI readiness.A Hidden AI TaxThe

183、cost of AI isnt in buying the technologyits in powering it.As demand for AI computing skyrockets,organizations will face a new economic reality:Buy in early to start their AI transformation,but unwittingly pay an increasingly steep premium later.Your Interface Is Costing You CustomersNatural languag

184、e is eating your user interface,turning it into abandonware.Every app,database,and system will soon be accessible through simple conversation.Organizations clinging to traditional interfaces will find themselves with the corporate equivalent of a flip phone in an iPhone world.Scale Out Over Scale Up

185、History shows that core technologies inevitably become cheaper and more widely distributed.The same pattern is emerging in AI,and the companies that only chase ever-larger models and data centers risk being disrupted by nimble upstarts that can spin up smaller,more efficient systems,at lower cost.WH

186、Y ARTIFICIAL INTELLIGENCE TRENDS MATTER TO YOUR ORGANIZATION 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEGenerative technologies advance over the next several years,while computing methods like organoid and neuromorphic computing drive developments in the long term.WH

187、EN WILL ARTIFICIAL INTELLIGENCE DISRUPT YOUR ORGANIZATION?FORECASTED TIME OF IMPACT012618315911371941610214820+YEARS51711Agentic Integration AI-Augmented Genetic RewritingAI Becomes the Foundation of Many IndustriesChain-of-ThoughtDecentralized AI AlignmentExplainable AI(XAI)Generalized AI AgentsHum

188、an-Level AI EmbodimentLarge Action ModelsLLMs as Operating SystemsModel Context ProtocolNatural Language Processing 2.0Neural Interface IntegrationNeuromorphic AI Systems Organoid IntelligencePLAMs Quantum AI Optimization Scaling AI-Generated New Materials Self-Assembling AI Networks Self-Improving

189、AI Systems Small,Efficient AI Models(SLMs)Turing-Complete Neural Interfaces 52 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEBelow,we highlight high level near-term developments to keep an eye on across industries.SCALINGEnormous amounts of training data are still requi

190、red for most AI models to learn.For example,recommender systems coupled with generative AI could lead to deep personalization for the hospitality and health care sectorsas long as data is made available.Historically,data is locked inside proprietary systems built by third parties,and regulation ofte

191、n hinders access to certain forms of data.INVESTMENTAI has seen cycles of enthusiasm and disillusionment,leading to either too much or not enough capital.Investors prioritize commercialization over basic R&Dthough the latter yields bigger impact and often stronger returns.In-vestors patience will in

192、fluence progress and commercialization.R&D DEVELOPMENTSThe pace of new research breakthroughs cant be scheduled to coincide with a board meeting or earnings report.Fac-tors like funding,quality,size of staff,and access to resources can improve the likelihood and speed of new discoveries.We closely m

193、onitor R&D developments but treat them as wild cards.MEDIA MENTIONSIncreased awareness and enthusiasm can influence the momentum of a technology,even when theres been no real break-through.Future media bursts will drive AI momentum,especially if those stories are easily understood by the public.PUBL

194、IC PERCEPTIONHow the public understands and responds to AI advancements will create or quell demand.This is especially true of gener-ative AI and education/creativity/intellec-tual property/misinformation,as well as the role assistive technologies will play in shaping the future workforce.CONSTRAINT

195、S ON ADOPTIONEven if a technology is maturing,con-straints on its adoption can hinder its im-pact.For example,a business may refuse to adopt an automated system because it challenges existing orthodoxy or an exist-ing successful strategy.This is especially true in health care,insurance,and finan-cia

196、l services.REGULATIONSAdvances in technology typically outpace regulatory changes.This has benefited AI,which until very recently was not targeted for regulation.Additionally,factors like whether local regulations are conflicting or complementary can influence adoption in the marketplace.53WHEN WILL

197、 ARTIFICIAL INTELLIGENCE DISRUPT YOUR ORGANIZATION?2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE Dr.Adji Bousso Dieng,assistant professor at Princeton University,for her work in deep probabilistic graphical modeling.Alexandr Wang,founder and CEO of Scale AI,for reveali

198、ng DeepSeeks open-source AI model and for creating a leading data annotation platform that accelerates AI model development across various industries.Dr.Abeba Birhane,senior fellow at Mozilla Foundation,for her research on the ethical implications of AI and critiques of algorithmic biases.Dr.Anima A

199、nandkumar,Bren Professor of Computing and Mathematical Sciences at Caltech,for developing AI algorithms that accelerate scientific discovery,including frameworks like neural operators for efficient simulations.Dr.Chinasa T.Okolo,computer scientist and fellow at the Brookings Institution,for her work

200、 in advocating for responsible AI adoption in the Global South and contributing to international AI safety reports.Clment Delangue,CEO at Hugging Face,for democratizing access to state-of-the-art NLP models and fostering an open-source AI community.Dr.Cynthia Rudin,the Gilbert,Louis,and Edward Lehrm

201、an Distinguished Professor at Duke University,for her work on interpretable machine learning models and ethical AI.Dr.Dan Hendrycks,director at the Center for AI Safety,for pioneering research in AI safety and developing Humanitys Last Exam,a test designed to evaluate AI risks,in collaboration with

202、Scale AI.Dean Ball,research fellow on AI&Progress at Georgetown Universitys Mercatus Center and author of the AI-focused Substack“Hyperdimensional,”for his analysis on AI governance.Dr.Devi Parikh,professor at Georgia Tech and co-founder of Yutori,for pioneering work in visual question answering and

203、 vision-language models,which helped establish foundational benchmarks for how AI systems understand and reason about visual information in natural language contexts.Dr.Ilya Sutskever,co-founder at Safe Superintelligence,for pioneering work in deep learning and leading efforts to develop AI that sur

204、passes human intelligence while remaining aligned with human interests.We expect to hear often from the worlds largest technology and AI companies.For that reason,our 2025 list highlights individuals flying deeper under the public radar.54PIONEERS&POWER PLAYERS 2025 Future Today Strategy Group.All R

205、ights Reserved.ARTIFICIAL INTELLIGENCE Dr.Joelle Pineau,vice president of AI research at Meta,for leading advancements in AI research and promoting open science,contributing to developments like the open-source language model LLamA.Dr.Kazumi Fukuda,research scientist at Sony AI,for her research in e

206、mbodied intelligence,particularly in developing computational models that enable robots to perceive,plan,and act in dynamic environments.Dr.Li Deng,chief AI officer at Vatic Investments,for his contributions to speech recognition and deep learning.Dr.Liang Wenfeng,CEO at DeepSeek,for developing the

207、R1 AI model,which rivals top competitors in capability but operates at a fraction of the cost.Dr.Lila Ibrahim,chief operating officer at Google DeepMind,for guiding the integration of AI research into practical applications and leading initiatives to apply AI in consumer products.May Habib,CEO and c

208、o-founder at Writer,for leading the development of enterprise AI tools that help businesses generate and manage high-quality content while ensuring brand consistency and compliance.Dr.Nathan Lambert,research scientist at the Allen Institute for AI(Ai2)and author of the Interconnects blog,for his con

209、tributions to the open science of language model fine-tuning.Dr.Pieter Abbeel,director of the Berkeley Robot Learning Lab and co-director of the Berkeley Artificial Intelligence Research Lab,for his work in robotics and reinforcement learning.Dr.Sasha Luccioni,climate and AI lead at Hugging Face,for

210、 developing tools to measure the carbon footprint of AI models and advocating for environmentally responsible AI practices.Dr.Sheng Shen,research scientist at Google,for coauthoring“Mixture-of-Experts Meets Instruction Tuning:A Winning Combination for Large Language Models,”exploring the integration

211、 of MoE architectures with instruction tuning to enhance language model performance.Dr.Tim Brooks,leader of the world modeling AI team at Google DeepMind,for his efforts in developing“world models”capable of simulating physical environments,advancing embodied AI in gaming and robotics.Dr.Yang Zhilin

212、,CEO at Moonshot AI,for leading the development of AI models with long context understanding and expanding AI applications globally.55PIONEERS&POWER PLAYERS 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEOPPORTUNITIESTHREATSBuild Internal AI Model Evaluation Frameworks C

213、ompanies that act now will gain first-mover advantage.These critical frameworks will enable rapid assessment and deployment of AI solutions while competitors struggle with ad-hoc evaluation methods.Create AI-Powered Knowledge Management SystemsThese systems transform static documentation into knowle

214、dge bases that continuously learn and adapt,letting businesses unlock significant value from institutional knowledge and retiring employees.Assess Hidden Technical Debt High-value AI clusters are prime targets for sophisticated cyberthreats,making cybersecurity investment essential.The stakes are es

215、pecially high as thieves could use compromised AI clusters to steal proprietary models,impacting industries globally.Delaying AI Adoption Means Rising Talent Costs AI professionals,making it prohibitively expensive to transform later.CEOs often delegate AI strategy to technical teams,ignoring the la

216、rger organizational and cultural changes needed.Embed AI Capabilities Directly Into Core OfferingsMost CEOs see AI as a cost-cutting tool,missing its potential to create new revenue streams through enhanced products and services that transform customer experiences and business models.Invest in Domai

217、n-Specific AI Models These focused models will deliver superior performance in targeted applications while requiring less data and compute resources than general-purpose alternatives,and early investors will benefit.Compounding Data Advantage Could Concentrate PowerAIs first-movers are accumulating

218、massive proprietary datasets and training pipelines that create nearly insurmountable barriers to entry.This advantage compounds over time and threatens to lock out new players.Risks of Digital ColonizationNations without sovereign AI capabilities become vulnerable to digital colonization as foreign

219、 AI systems shape their residents information access and decision-making.AI adoption is creating unprecedented opportunities for value creation.but organizations face growing risks from technical complexity and talent scarcity.56OPPORTUNITIES AND THREATS 2025 Future Today Strategy Group.All Rights R

220、eserved.ARTIFICIAL INTELLIGENCECompanies face steep hidden costs and complex organizational hurdles as they rush to implement AI,from outdated infrastructure to employee resistance and regulatory demands.57Technology DeploymentTalent Development Regulatory InfluenceCapital ExpenditureAs organization

221、s hurry to deploy AI,theyre facing a costly reality:Most of their data infrastructure is decades behind whats required.The investment needed to modernize data architecture often exceeds initial AI project budgets by 510 x,creating a hidden barrier to transformation.While companies eagerly invest in

222、AI,they often overlook crucial investments in change management and employee support.This oversight could lead middle managers to view AI as a threat rather than a tool,fostering resistance that may dramatically slow implementation and adoption.Build AI models with dedicated testing environments to

223、rigorously evaluate financial risk against historical market conditions and stress scenarios.These systems require specialized hardware and compliance monitoring infrastructure to handle complex simulations while maintaining audit trails.Establish rigorous protocols for auditing AI models that enabl

224、e tracking of decisions,data lineage,and performance drift across systems.Integrating automation with human oversight is crucial,along with detailed audit trails that comply with both technical governance and regulations.Know your constraints.Either find trusted external partners with expertise in b

225、oth AI and your domain,or do it in-house.These teams must meld their AI skills with industry-specific knowledge to create solutions that address both regulatory and technical challenges while delivering measurable value.Every piece of technical debt becomes a critical bottleneck as AI systems demand

226、 more flexible,interconnected infrastructure.Smart companies will treat technical debt elimination as a core part of AI strategy,not a separate IT initiative.INVESTMENTS AND ACTIONS TO CONSIDER 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCEAGENTIC AIThis refers to AI sy

227、stems that exhibit autonomous decision-making,goal-setting,and adaptive problem-solving capabilities.Unlike traditional AI models that passively generate responses based on user prompts,agentic AI proactively takes actions,interacts with its environment,and refines its strategies over time.AGENTSAI-

228、powered entities that perceive their environment,make decisions,and take actions autonomously to achieve specific goals.Agents can range from simple automation tools to complex,multimodal AI systems that interact dynamically with users and other systems.AGI(ARTIFICIAL GENERAL INTELLIGENCE)A designat

229、ion for AI systems that match and then exceed the full range of human cognitive abilities across all economically valuable tasks.AGI remains theoretical,but its potential implications for labor markets,governance,and global security are actively debated.AUTONOMOUS AI AI systems capable of independen

230、t decision-making and execution of tasks without human intervention.Autonomous AI is critical in robotics,finance,cybersecurity,and military applications,requiring rigorous safeguards to ensure responsible use.CHAIN OF THOUGHT(COT)REASONINGAn AI reasoning method where models solve problems step-by-s

231、tep,mimicking human-like logical deduction.This improves performance in complex decision-making tasks,including math,legal analysis,and medical diagnostics.COMPUTER VISIONAI-driven technology that enables machines to process,analyze,and derive meaning from digital images and video.Used in security s

232、urveillance,industrial automation,medical imaging,and self-driving vehicles.ALIGNMENTThe process of ensuring that an AI systems goals,behaviors,and decision-making align with human intentions,ethical principles,and regulatory standards.Misalignment can result in unintended consequences,including bia

233、sed or harmful outcomes.ARTIFICIAL SUPERINTELLIGENCE(ASI)A hypothetical future AI system that surpasses human intelligence across all domains,including creativity,general wisdom,strategic planning,and scientific discovery.ASI raises complex questions about control,governance,and existential risk.AUT

234、OMATIC SPEECH RECOGNITION(ASR)AI-driven systems that convert spoken language into written text.ASR powers virtual assistants,transcription services,and multilingual voice interfaces in enterprise and consumer applications.AI ETHICSA multidisciplinary field that studies the societal,economic,and ethi

235、cal risks of AI,including bias,privacy,misinformation,and existential threats.AI ethics frameworks guide policy and regulation to ensure AI development aligns with human values.AI GOVERNANCEThe systems,policies,and international agreements that regulate the develop-ment,deployment,and oversight of A

236、I technologies.AI governance is critical to mitigating risks,ensuring fair competi-tion,and addressing geopolitical tensions around AI capabilities.ALGORITHMA structured set of rules or processes for solving specific problems or performing tasks.In AI,algorithms determine how data is processed,insig

237、hts are generated,and decisions are made.Important terms to know before reading.58 2025 Future Today Strategy Group.All Rights Reserved.IMPORTANT TERMSARTIFICIAL INTELLIGENCEEDGE AI AI models that run directly on edge devices(e.g.,smartphones,IoT sensors,auton-omous drones)rather than centralized cl

238、oud servers.Edge AI enables real-time processing,reduces latency,and enhances data privacy.FOUNDATION MODELA large-scale AI model pretrained on vast amounts of data and adaptable to multiple tasks without requiring retraining from scratch.Foundation models underpin modern AI applications,including g

239、enerative AI,autonomous systems,and enterprise automation.GENERATIVE AI(GENAI)AI technologies capable of generating novel content,including text,images,music,video,and code.GenAI is transforming industries such as media,design,marketing,and customer service while raising concerns about intellectual

240、property and misinformation.QUANTUM AI The intersection of quantum computing and AI,where quantum algorithms enhance machine learning efficiency.Quantum AI has the potential to revolutionize cryptography,materials science,and optimization problems.RECOMMENDER SYSTEMSAI-driven algorithms that analyze

241、 user behavior and preferences to suggest relevant products,content,or services.Used in e-commerce,streaming platforms,and digital advertising.REINFORCEMENT LEARNING FROM HUMAN FEEDBACK(RHLF)A training method where AI models learn through iterative feedback from human evaluators,improving their accu

242、racy,ethical alignment,and usability in real-world applications.NATURAL LANGUAGE PROCESSING(NLP)AI-driven processes that enable machines to understand,interpret,and generate human language.NLP powers chatbots,translation services,sentiment analysis,and automated content moderation.NEURAL ARCHITECTUR

243、E SEARCH(NAS)An AI-driven method for automatically optimizing neural network structures,improving performance while reducing the need for manual tuning by researchers.PARAMETERAn internal variable of an AI model that is fine-tuned during training to improve accuracy and efficiency.Large AI models co

244、ntain billions of parameters,making their training computationally intensive.PROMPT ENGINEERING The practice of designing effective inputs(prompts)to guide AI models in generating desired outputs.Prompt engineering is crucial for optimizing generative AI performance in business and creative applicat

245、ions.GPU(GRAPHICS PROCESSING UNIT)Specialized hardware optimized for parallel computing,accelerating AI model training,deep learning,and high-performance computing tasks.GPUs are essential for running large-scale AI models and data-intensive simulations.MODELA trained AI system that analyzes data to

246、 make predictions,generate insights,or automate decision-making.Models vary in complexity,from simple regression models to advanced deep learning architectures.MULTIMODAL AI AI systems that process and integrate multiple types of datasuch as text,images,video,and audioto improve contextual understan

247、ding and decision-making.Multimodal AI powers advanced chatbots,virtual assistants,and medical diagnostics.59 2025 Future Today Strategy Group.All Rights Reserved.IMPORTANT TERMSARTIFICIAL INTELLIGENCESELF-SUPERVISED LEARNING A machine learning approach where AI models learn from raw,unlabeled data

248、by identifying patterns and relationships within the dataset.This method reduces dependency on human-labeled training data.SUPERVISED LEARNINGA training method where AI models learn from labeled datasets,using known input-output pairs to improve predictive accuracy in new data.SYMBOLIC AIAn AI appro

249、ach that represents knowledge using human-readable symbols and logical rules,enabling reasoning and problem-solving.Often used in expert systems and explainable AI models.SYNTHETIC DATA Artificially generated data used to train AI models when real-world data is scarce,biased,or privacy-sensitive.Syn

250、thetic data enhances AI performance while mitigating data collection risks.TRAINING DATAThe dataset used to train AI models by identifying patterns,making decisions,or generating predictions.The quality and diversity of training data significantly impact model accuracy and fairness.TRUSTWORTHY AI AI

251、 systems designed with transparency,fairness,accountability,and security to foster public trust and regulatory compliance.Trustworthy AI is a key focus for government and enterprise AI strategies.UNSUPERVISED LEARNINGA machine learning approach where AI models detect patterns and structures in data

252、without labeled outputs,enabling tasks like clustering and anomaly detection.60IMPORTANT TERMSXAI(EXPLAINABLE AI)AI systems designed to provide transparent,human-interpretable explanations for their decision-making processes,increasing accountability and trust in high-stakes applications like health

253、 care and finance.ZERO-SHOT LEARNING(ZSL)An AI technique where models generalize knowledge from previously learned concepts to perform tasks without direct prior training on those tasks.Used in applications like language translation and image recognition.2025 Future Today Strategy Group.All Rights R

254、eserved.ARTIFICIAL INTELLIGENCEARTIFICIAL INTELLIGENCETRENDS61 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE62MODELS,TECHNIQUES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE63AI models require massive data and computing resource

255、s to unlock their transformative potential(or so we thought).Generative AI Modalities ExpandHumans dont just learn by readingwe observe,listen,and synthesize information from multiple sources.AI is now following suit,integrating inputs like text,images,and sound to bridge the gap between what we des

256、cribe and what machines can fully understand.2024 marked the year when multimodal AI capabilities not only ma-tured but began transforming real-world applications.OpenAIs GPT-4o builds on the multimod-al progress of 2023 by integrating text,vision,and voice into one robust model.Its real-time conver

257、sational capabilities open doors for fluid,human-like interactions.GPT-4os ability to analyze and generate insights from combined text,audio,and image inputs allows it to solve complex tasks with remarkable depth and accuracy.Anthropics Claude 3 brings sophisticated visual interpretation to enterpri

258、se settings,where roughly half of knowledge bases are image-based.This capability is particularly transformative in health care,where AI can now connect medical imagery with patient records for enhanced diagnostics.On the consumer side,multimodal AI is starting to become second nature.Instead of typ

259、ing out queries,users now share photos of recipes to adjust portions or upload images of rashes for medical suggestions.While this democratization makes expertise more accessible,it also raises questions about accuracy and ethical boundaries when AI tools replace professional judgment.MIT and Micros

260、ofts Large Language Model for Mixed Reality(LLMR)is pushing the multimodal boundaries even further.LLMR uses AI to simplify the creation and modification of virtual environments.Instead of needing complex coding,LLMR enables users to describe their vision in plain language,and the system transforms

261、those words into interactive mixed reality experiences in real time.For example,a user might say,“Place a green bench in the park next to the fountain,”and the system executes it instantly.While 2024s breakthroughs in multimodal AI mark a technological leap forward,their true significance lies in ho

262、w theyre reshap-ing the fundamental relationship between humans and machines,moving us from giving commands to having conversations.Fine TuningFine-tuningthe process of refining LLMs on specialized datasetsis improving our ability to customize and control AI systems.In 2023,the University of Washing

263、tons QLoRA breakthrough marked a turning point,enabling the fine-tuning of massive 65-billion-parameter models on a single GPU with just 48GB of memorya 16-fold efficiency improvement over traditional methods.Building on this foundation,in 2024,Answer.AI integrated QLoRA with Fully Sharded Data Para

264、llel processing,making it possible to train 70-billion-pa-rameter models on consumer-grade hard-ware.This democratization has profound implications:Researchers and developers can now experiment with large-scale lan-guage models without access to expensive data center infrastructure.The impact extend

265、s beyond accessibility.In the biological sciences,researchers have adapted fine-tuning techniques to protein language models(PLMs),which are trained on extensive datasets of protein sequenc-es.These models are now being fine-tuned to predict protein stability,functions,and interactions with remarkab

266、le accuracy.Fine-tuned PLMs outperform their non-tuned counterparts across multiple bench-marks,showcasing enhanced predictive capabilities.Fine-tuning isnt just a technical capabil-ityits a strategic business advantage.In the enterprise context,companies can MODELS,TECHNIQUES,&RESEARCH 2025 Future

267、Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE64use fine-tuning to customize powerful AI models like GPT and Claude for their spe-cific needs without building from scratch.Health care providers can enhance diag-nostic capabilities by training models on anonymized patient records wh

268、ile main-taining HIPAA compliance,while financial institutions can embed regulatory require-mentsfrom GDPR to PCI DSSdirectly into their AI workflows.This dramatically reduces development costs and time-to-market while ensuring AI systems speak the organizations language,understand industry-specific

269、 contexts,and operate within required compliance frameworks.Automated Reinforcement LearningTraditional AI training using Reinforcement Learning from Human Feedback(RLHF)involves people rating AI responses to help improve the system.While this method works well,its expensive and time-con-suming.Deep

270、Seek found a clever alterna-tivethe startup developed a way to train AI systems using automated computer feedback instead of human ratings.While more subjective tasks(like creative writing or open-ended questions)still need some human input,DeepSeeks automated meth-od works especially well for tasks

271、 with clear right/wrong answers,like math and coding problems.To make this automated training even more efficient,DeepSeek cre-ated a special method called GRPO(Group Relative Policy Optimization)and tested it first with its math-focused model.The company isnt alone.Microsoft Asia devel-oped a math

272、model using comparable tech-niques,Ai2 created a model called Tulu that combines both automated and human feedback,and Hugging Face is working on recreating DeepSeeks approach to better understand how it works.The key take-away is that DeepSeek showed its possible to create high-performing AI system

273、s with less reliance on expensive human feed-back,particularly for certain types of tasks.This could make AI development more efficient and cost-effective,though human input is still valuable for some applications.Evolutionary CompositionSakana AI is challenging the conventional wisdom that bigger,m

274、ore expensive models are the only path to better AI.Instead of training massive models from scratcha process requiring enormous computational resourcesthe Japanese firm has developed an elegant alternative:using evolutionary algorithms to automatically discover optimal ways to combine existing AI mo

275、dels.This“evolutionary optimization”approach is big;by intelligently merging models from different domainssuch as language processing and visual understandingSakana creates hybrid systems that exceed the capabilities of their individual components.The results are impressive:The experiments produced

276、Japanese language models with enhanced mathematical reasoning and cultural awareness that outperformed larger,more resource-intensive systems.The implications extend far beyond tech-nical achievement.This methodology democratizes advanced AI development by reducing the need for massive computing inf

277、rastructure and specialized expertise.Rather than requiring tens of millions in computing resources,developers can now create sophisticated multi-capable models by intelligently combining existing ones.Most significantly,Sakanas approach suggests a future where AI advancement isnt just about buildin

278、g bigger models but about finding smarter ways to com-bine existing ones.Just as nature creates complexity through the combination and evolution of simpler elements,this new paradigm points to a more sustainable and accessible path forward in AI develop-mentone where innovation comes from intelligen

279、t composition rather than brute-force scaling.Mixture of ExpertsUnlike the previous approach that merges entire trained models,mixture-of-experts(MoE)divides up the work inside a single framework by creating multiple specialized“expert”sub-models.Think of it like having a team of people where one pe

280、rson is great MODELS,TECHNIQUES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE65at math,another at writing,and another at design,with a manager who knows who to call on for each task.This“manager”(the gating mechanism)directs each input to the right expert,so

281、every piece of the job is handled by the specialist best suited to it.By splitting tasks among experts and letting the gating mechanism handle the“who does what,”MoE models can become more efficient and accurate than if one gi-ant,one-size-fits-all model tried to handle everything on its own.Notably

282、,DeepSeeks January 2025 re-lease,R1,uses MoE at its core.As reported,DeepSeek claims to have built a ChatGPT-like system at a fraction of the usual cost by employing MoE(along with other techniques such as knowledge distillation and reinforcement learning).Because MoE breaks a large model into speci

283、alized“experts”and relies on a gating function to route each request to the most appropriate expert,it can be more efficient and poten-tially less expensive to train or run than a single giant monolithic model.DeepSeeks success with this approach has sparked new attention on MoE as a viable alterna-

284、tive for scaling AI without requiring mas-sive,prohibitively expensive hardware.Autonomy-of-ExpertsThough DeepSeek is what put MoE in the news for the general public,others had already made significant breakthroughs in the field.Researchers at Renmin University of China,Tencent,and Southeast Univer-

285、sity released a paper that describes a new“Autonomy-of-Experts”(AoE)approach for mixture-of-experts models.While the typical MoE model relies on a“router”that makes its best guess of which specialist should handle each incoming question or input,AoE doesnt need the router.Instead,each expert peeks a

286、t the input and says,“I can handle this,”or“No thanks,thats not my specialty,”based on how strongly it lights up the experts internal signals(the“activation norms”).The strongest signals win,so those experts step up,and the rest step back.In other words,each expert autonomously decides whether its t

287、he best fit.This cuts out the middleman(the router)entirely.MODELS,TECHNIQUES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE66LLMs as Operating SystemsImagine an operating system fundamen-tally powered by a large language model(LLM),where the LLM is not just a

288、n add-on but the core kernel of the OS.This OS could automate routine tasks with un-precedented sophistication,eliminating the need for manual intervention.It would move beyond traditional graphical user interfaces and command-line interactions,embracing a more intuitive,natural lan-guagebased appro

289、ach.Users could inter-act with their computers through conversa-tional commands,inquiries,or requests for specific tasks,and the LLM would interpret these inputs,executing a series of actions to deliver the desired outcomes.One such project,AIOS,envisions an LLM as the“brain”of the OS.AIOS optimizes

290、 re-source allocation,manages context switch-ing,facilitates concurrent agent execution,provides tools for agents,and maintains access control.The LLM handles complex decision-making,turning the OS into a more intelligent,adaptive system.Another project,MemGPT,focuses on enhancing LLM-driven systems

291、 by integrating long-term memory and improving reasoning capabilities.Traditional LLMs are limited by small context windows,restricting how much information they can process at once.MemGPT addresses this by introducing a multilevel memory architecture,inspired by traditional OS memory management tec

292、h-niques like virtual memory,to enable more complex and contextually aware process-ing over time.Together,these projects rep-resent the future of LLM-centric operating systems,enabling more efficient,natural,and powerful interactions.LLMs:Bigger and More ExpensiveLLMs have grown exponentially in siz

293、e and cost over the past decade,driven by the“bigger is better”paradigm.This approach emerged from scaling laws,first introduced by Prasanth Kolachina in 2012 and later validated by Kaplan et al.in 2020,which demonstrated a strong correlation between model size and performance.Following these insigh

294、ts,the industry has pursued increasingly larger systems,progressing from GPT-2s 1.5 billion parameters in 2019 to models with trillions of parameters like GPT-4 and PaLM 2 in 2023.The financial impact of this growth is significant:Stanfords 2024 AI Index Report estimates place the training costs of

295、top-tier models at unprecedented levels,with OpenAIs GPT-4 requiring approximately$78 million in compute and Googles Gemini Ultra costing an estimated$191 million.The benefits of scaling have been substan-tial and well-documented.Larger models have demonstrated remarkable capabil-ities in handling c

296、omplex tasks,showing improved accuracy and efficiency across a wide range of applications.These achieve-ments have validated,at least partially,that bigger is indeed better.However,this progress has come with significant costs and challenges.Training GPT-4 required approximately 10,000 times more co

297、mputa-tional resources than its predecessor GPT-2,necessitating enormous investments in infrastructure and specialized hardware.Perhaps most significantly,the relation-ship between model size and performance has proven more complex than initially assumed.Many tasks exhibit diminishing returns as mod

298、els grow larger,calling into question the long-term viability of this approach.This observation has particular relevance for businesses,which are discov-ering that larger models dont automatically translate to better solutions for their spe-cific needs.The combination of rising costs,environmental c

299、oncerns,and uncertain per-formance benefits has prompted a critical examination of the scaling paradigm.As the field matures,a more nuanced approach to AI development is emerging.Rather than pursuing size alone,research-ers are increasingly focusing on efficiency improvements and the development of

300、smaller,specialized models.Chain-of-Thought Models As larger models reach practical and financial limits,researchers are shift-ing attention to new approaches such as Chain-of-Thought(CoT),which emphasizes MODELS,TECHNIQUES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL IN

301、TELLIGENCE67deeper real-time reasoning rather than raw parameter counts.For more than a decade,AI progress has been largely driven by the pretraining scaling law,which emerged with AlexNet in 2012 and gained momentum with the Transformer architecture in 2017.This law states that increasing the amoun

302、t of training data(now reaching trillions of tokens),expanding model parameters,and using more compute(FLOPS)leads to bet-ter performance across various tasks.Put simply,pretraining scaling laws describe how larger models,with more data and compute,achieve superior performance.Now,there is a new sca

303、ling law in town.Previously,most computational cost was concentrated in pretraining.Once a model was trained,running inferencegenerating responses or completing tasksrequired significantly less compute.Inference scaled in a straightforward way:The more requests a model handled,the more compute it us

304、ed.However,the introduction of CoT models has fundamentally changed this paradigm.OpenAIs o1 model and DeepSeeks R1 have demonstrated that inference compute is no longer strictly proportional to output length.These models generate interme-diate“logic tokens,”acting as an internal scratchpad to break

305、 down problems into structured reasoning steps.This shift means that the more tokens dedicated to this internal process,the better the models output.Essentially,it mimics how humans improve their workdouble-checking,ver-ifying calculations,and cross-referencing solutions to ensure accuracy.Expect a

306、growing emphasis on dynamic inference strategies,where models can flexibly adjust the number of internal logic tokens based on task difficulty or desired accuracy.As a result,inference compute could become much more significant relative to training,driving the need for more efficient hardware soluti

307、ons,better optimization techniques,and new business models around usage-based compute.Overall,AI development will likely shift toward architectures and methods that let models“think out loud,”enabling deeper reasoning and better outcomes at the cost of increased on-the-fly processing.MODELS,TECHNIQU

308、ES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICIAL INTELLIGENCE68Small Language ModelsSmall language models(SLMs)are prov-ing that bigger isnt always better.These compact models can match or exceed the performance of their larger counterparts in specific tasks,while demandi

309、ng far less computational power and resources.SLMs showcased impressive performance in task-specific scenarios,such as zero-shot text classification.Research across mul-tiple datasets revealed that models with fewer parameters could rival larger coun-terparts,emphasizing their potential for efficien

310、cy without compromising effective-ness.Additionally,cost-effective solutions like OpenAIs GPT-4o mini cost more than 60%less compared to previous models,making high-quality AI more accessible to businesses and developers.Microsofts Phi-3-mini stands as another example,achieving superior reasoning an

311、d logic capabilities with just 3.8 billion parametersoutperforming models twice its size.Similarly,Metas Llama 3.2 family demonstrates the viability of smaller mod-els,offering variants from 1 billion to 90 billion parameters that prioritize efficiency without sacrificing effectiveness.This shift to

312、ward smaller models is support-ed by industry experts like Andrej Karpathy,who advocates for distilling models to their essential“cognitive core.”His research suggests that even a 1-billion-parameter model could provide sufficient cognitive capabilities,as much of the additional data in larger model

313、s may not directly enhance performance.This insight has practical im-plications,particularly in consumer technol-ogy.Apples OpenELM models,for instance,enable on-device AI processing,delivering responsive,personalized experiences while maintaining privacy and energy efficien-cy.In 2024,models like t

314、he SlimLM series enabled robust processing on devices such as smartphones,eliminating the need for cloud-based computation.This innovation marked a leap forward in AI accessibility,enabling users to perform tasks directly on their devices while maintaining privacy and reducing latency.The impact of

315、SLMs extends beyond gen-eral applications into specialized domains.At Ignite 2024,Microsoft collaborated with industry leaders like Bayer and Rockwell Automation to develop targeted SLMs for agriculture and manufacturing,demon-strating how these compact models can excel in specific sectors without t

316、he over-head of larger,general-purpose systems.Looking forward,the concept of“compa-nies of LLMs”where multiple specialized models work in parallelcould represent the next evolution,combining the advan-tages of both large and small models in a modular approach.Grounding and Context AugmentationNoteb

317、ookLM from Google Labs transforms AI into a personalized research assistant by“grounding”it in your Google Docs.Unlike traditional chatbots,it ties responses to your specific notes and sources,enabling insights that are highly relevant and trust-worthy.This feature is designed to tackle information

318、overload,making it easier to synthesize and connect ideas from multi-ple sources efficiently.Grounding ensures that AI outputs are anchored in verified data,reducing errors like hallucinations.NotebookLM builds on this by leveraging contextual augmentation,which enriches responses with nuanced under

319、standing tai-lored to your needs.This approach doesnt just deliver answers but delivers the right answers for your unique context.Tools like GenAI Data Fusion extend this principle to businesses,aggregating and contextualizing enterprise-specific data.By rooting outputs in tailored datasets,companie

320、s can achieve highly accurate,task-specific insights for applications rang-ing from research to operational analytics.Grounding and contextual augmentation mark a shift in AI,turning generic systems into precise,adaptive tools that meet indi-vidual and organizational needs.Overcoming the Data Shorta

321、geThe availability of high-quality data is emerging as a bottleneck in the develop-ment of large AI models.According to Ep-och AI,the reservoir of high-quality textual data on the public internet may be exhaust-MODELS,TECHNIQUES,&RESEARCH 2025 Future Today Strategy Group.All Rights Reserved.ARTIFICI

322、AL INTELLIGENCE69ed as early as 2026.Initially,researchers estimated the stock of high-quality lan-guage data could run out by 2024,while low-quality data might last another two de-cades,and image data could face depletion by the late 2030s to mid-2040s.Although the 2024 thresholds have not yet been

323、 reached,the looming scarcity is pushing AI labs to explore alternative strategies for sourcing training data.This prediction has sparked diverse strat-egies among AI labs.Some are pursuing private data sources,purchasing from bro-kers or licensing content from publishers.Others are exploring untapp

324、ed audio and visual data,with video content offering par-ticularly valuable insights into real-world physics and dynamics.Companies like Scale AI and Surge AI are building exten-sive networks of contributors,including Ph.D.-level experts,to create and annotate specialized datasets.These approaches,h

325、owever,come at a high cost,with some estimates suggesting AI labs are spending hundreds of millions of dollars annually on these initiatives.An alternative method involves using one AI model to generate vast amounts of synthetic data to train another,but its risky.Studies have shown that models trai

326、ned predominantly on synthetic data can experience“model collapse,”where their outputs become less diverse and fail to reflect real-world distributions.A related phenomenon,termed Model Autophagy Disorder(MAD),has been observed in gen-erative image models,where reliance on synthetic data leads to a

327、notable decline in output quality.One promising solution lies in techniques such as“self-play,”where models improve through competition or collaboration with themselves.Google DeepMinds AlphaGo,which famously defeated the human world champion in Go after training against itself,exemplifies this appr

328、oach.Today,self-play continues to inform cutting-edge LLM de-velopment,offering a pathway to overcome data limitations while maintaining perfor-mance and innovation.Open-Source AIIn January 2025,Chinese AI company DeepSeek launched R1,an open-source reasoning model designed to compete withand potent

329、ially outperformOpe-nAIs o1 at a fraction of the cost.While R1s reasoning process is slower than that of many general-purpose models,it delivers more nuanced and accurate responses.Alongside its flagship 671-billion-parame-ter version,DeepSeek also introduced six smaller“distilled”models,starting at

330、 1.5 billion parameters and capable of running on local devices.Other open-source models are likewise closing the performance gap with propri-etary alternatives.Metas Llama 3.1 now ri-vals GPT-4 on key benchmarks,and Mistral AIs models offer capabilities on par with top closed-source solutionsso muc

331、h so that Mistral AIs recent$487 million fund-ing round catapulted it to unicorn status.As a growing number of open-source mod-els achieve high-level performance,cor-porations are increasingly taking note.The Brave browser,for example,has integrated Mistral AIs Mixtral 8x7B model into its Leo assist

332、ant,while Wells Fargo has adopted Metas Llama 2 for internal applications.This momentum reflects a broader trend toward open-source AI.GitHub statistics reveal that AI-focused repositories have skyrocketed from just 845 in 2011 to 1.8 million in 2023an impressive 59.3%increase in 2023 alone.During t

333、he same period,community engagement soared,with GitHub stars for AI projects jumping from 4 million to 12.2 million between 2022 and 2023.This surge highlights the grow-ing appetite for collaborative development,signaling that open-source AI will remain a driving force well into the future.Modular AI Rather than building monolithic models,the modular AI approach breaks AI sys-tems into specialized