《斯坦福：2025斯坦福新興技術評論：十項關鍵技術及其政策影響分析報告（英文版）（191頁）.pdf》由會員分享，可在線閱讀，更多相關《斯坦福：2025斯坦福新興技術評論：十項關鍵技術及其政策影響分析報告（英文版）（191頁）.pdf（191頁珍藏版）》請在三個皮匠報告上搜索。

1、THE STANFORD EMERGING TECHNOLOGY REVIEW 2025A Report on Ten Key Technologies and Their Policy ImplicationsCO-CHAIRSCondoleezza Rice John B.Taylor Jennifer Widom Amy ZegartDIRECTOR AND EDITOR IN CHIEFHerbert S.LinStanford University Stanford,CaliforniaMANAGING EDITORMartin GilesCONTENTSFOREWORD4EXECU

2、TIVE SUMMARY10INTRODUCTION18The Role of Science and Technology in Advancing National Interests18 Policy for Science and Technology19Ten Science and Technology Fields1901Artificial Intelligence2102Biotechnology and Synthetic Biology3903Cryptography5304Lasers6505Materials Science7706Neuroscience9107Ro

3、botics10308Semiconductors11509Space12710Sustainable Energy Technologies14111Crosscutting Themes and Commonalities156Key Observations About How Technologies Evolve over Time157Common Innovation Enablers and Inhibitors16412Technology Applications by Policy Area174Economic Growth174National Security176

4、Environmental and Energy Sustainability177Health and Medicine179Civil Society180CONCLUSION182LEADERSHIP184ACKNOWLEDGMENTS1884government offices are likely to set trajectories for the United States and the world for years to come.Now more than ever,understanding the landscape of discovery and how to

5、harness technology to forge a better future requires working across sectors,fields,and generations.Universities like Stanford have a vital role to play in this effort.In 2023,we launched the Stanford Emerging Technology Review(SETR),the first-ever collaboration between Stanford Universitys School of

6、 Engineering and the Hoover Institution.Our goal is ambitious:transforming tech-nology education for decision makers in both the public and private sectors so that the United States can seize opportunities,mitigate risks,and ensure the American innovation ecosystem continues to thrive.This is our la

7、test report surveying the state of ten key emerging technologies and their implications.It harnesses the expertise of leading faculty in science and engineering fields,economics,international relations,and history to identify key technological developments,assess potential implications,and highlight

8、 what policymakers should know.This report is our flagship product,but it is just one element of our continuous technology education campaign for policymakers that now involves nearly one hundred Stanford scholars across forty depart-ments and research institutes.In the past year,SETR experts have b

9、riefed senior leaders across the US government in Congress and in the White House,Commerce Department,Defense Department,and US intelligence community.We have organized and participated in fifteen Stanford programs,including multiday AI and biotechnology boot camps for con-gressional staff;SETR roun

10、dtables for national media and officials from European partners and allies;and In every era,technological discoveries bring both promise and risk.Rarely,however,has the world experienced technological change at the speed and scale we see today.From nanomaterials that are fifty thousand times smaller

11、 than the width of a human hair to commercial satellites and other private-sector technologies deployed in outer space,break-throughs are rapidly reshaping markets,societies,and geopolitics.Whats more,US technology policy isnt the unique province of government like it used to be.Instead,inventors an

12、d investors are making decisions with enormous policy consequences,even if they may not always realize it.Artificial intelligence(AI)algorithms are imbued with policy choices about which outcomes are desired and which are not.Nearly every new technology,from bioengineer-ing new medicines to building

13、 underwater research drones,has both commercial and military applica-tions.Private-sector investment,too,simultaneously generates both national advantages and vulnerabil-ities by developing new capabilities,supply chains,and dependencies and by pursuing commercial opportunities that may not serve lo

14、ng-term national interests.While engineers and executives need to better understand the policy world,government lead-ers need to better understand the engineering and business worlds.Otherwise,public policies intended to protect against societal harms may end up accelerating them,and efforts to alig

15、n innova-tion with the national interest could end up harming that interest by dampening Americas innovation leadership and the geopolitical advantages that come with it.In these complex times,the only certainties are that uncertainty is rampant and the stakes are high:Decisions made today in boardr

16、ooms,labs,and FOREWORDSTANFORD EMERGING TECHNOLOGY REVIEW52.Academias role in American innovation is essential and at risk.The US innovation ecosystem rests on three pillars:the government,the private sector,and the acad-emy.Success requires robust research and develop-ment(R&D)in all three.But they

17、 are not the same,and evidence increasingly suggests that universities role as the engines of innovation is at a growing risk.Universities,along with the US National Laboratories,are the only institutions that conduct research on the frontiers of knowledge without regard for poten-tial profit or for

18、eseeable commercial application.This kind of research is called basic or fundamental research.It takes years,sometimes decades,to bear fruit.But without it,future commercial innovations would not be possible.Radar,the Global Positioning System(GPS),and the internet all stemmed from basic research do

19、ne in universities.So did the recent“overnight success”of the COVID-19 mRNA vac-cines,which relied on decades of university research that discovered mRNA could activate and block protein cells and figured out how to deliver mRNA to human cells to provoke an immune response.Similarly,the cryptographi

20、c algorithms protecting data on the internet today would not have been possible without decades of academic research in pure math.And many of the advances in AI,from ChatGPT to image recognition,build on pioneering work done in university computer science depart-ments that also trained legions of st

21、udents who have gone on to found,fund,and lead many of todays most important tech companies.In many ways and in nearly every field,Americas innovation supply chain starts with research universities.Yet evidence suggests that the engine of inno-vation in US research universities is not running worksh

22、ops convening leaders across sectors in semi-conductors,space technology,and bioengineering.And we are just getting started.Our efforts are guided by three observations:1.Americas global innovation leadership matters.American innovation leadership is not just impor-tant for the nations economy and s

23、ecurity.It is the linchpin for maintaining a dynamic global technol-ogy innovation ecosystem and securing its benefits.International scientific collaboration has long been pivotal to fostering global peace,progress,and prosperity,even in times of intense geopolitical competition.During the Cold War,

24、American and Soviet nuclear scientists and policymakers worked together to reduce the risk of accidental nuclear war through arms control agreements and safety measures.Today,Chinas rise poses many new chal-lenges.Yet maintaining a robust global ecosystem of scientific cooperation remains essential

25、and it does not happen by magic.It takes work,leader-ship,and a fundamental commitment to freedom to sustain the openness essential for scientific dis-covery.Freedom is the fertile soil of innovation,and it takes many forms:the freedom to criticize a government;to admit failure in a research program

26、 as a step toward future progress;to share findings openly with others;to collaborate across geograph-ical and technical borders with reciprocal access to talent,knowledge,and resources;and to work without fear of repression or persecution.In short,it matters whether the innovation ecosystem is led

27、by democracies or autocracies.The United States has its flaws and challenges,but this country remains the best guarantor of scientific freedom in the world.FOREWORD6STANFORD EMERGING TECHNOLOGY REVIEWToday,only a handful of the worlds largest compa-nies have both the talent and the enormous com-pute

28、 power necessary for developing sophisticated large language models(LLMs)like ChatGPT.No uni-versity comes close.In 2024,for example,Princeton University announced that it would use endowment funds to purchase 300 advanced Nvidia chips to use for research,costing about$9 million,while Meta announced

29、 plans to purchase 350,000 of the same chips by years end,at an estimated cost of$10 billion.7These trends have several concerning implications.8 A very significant one is that research in the field is likely to be skewed to applications driven by com-mercial rather than public interests.The ability

30、 for universities or anyone outside of the leading AI companies to conduct independent analysis of the weaknesses,risks,and vulnerabilities of AI(espe-cially LLMs recently in the news)will become more important and simultaneously more difficult.Further,the more that industry offers unparalleled tale

31、nt con-centrations,computing power,training data,and the most sophisticated models,the more likely it is that future generations of the best AI minds will continue to flock there(see figureF.1)potentially eroding the nations ability to conduct broad-ranging foun-dational research in the field.3.The

32、view from Stanford is unique,important and needed now more than ever.Stanford University has a unique vantage point when it comes to technological innovation.It is not an accident that Silicon Valley surrounds Stanford;tech-nology developed at Stanford in the 1930s served as the foundation for the p

33、ioneering companies like Varian Associates and Hewlett-Packard that first shaped industry in the region.Since then,the univer-sity has continued to fuel that innovation ecosystem.Stanford faculty,researchers,and former students have founded Alphabet,Cisco Systems,Instagram,LinkedIn,Nvidia,Sun Micros

34、ystems,Yahoo!,and many other companies,together generating more annual revenues than most of the worlds economies.as well as it could,posing long-term risks to the nation.In 2024,for the first time,the number of Chinese contributions surpassed those of the United States in the closely watched Nature

35、 Index,which tracks eighty-two of the worlds premier science journals.1 Funding trends are also head-ing in the wrong direction.The US government is the only funder capable of making large and risky investments in the basic science conducted at uni-versities(as well as at national laboratories)that

36、is essential for future applications.Yet federal R&D funding has plummeted in percentage terms since the 1960s,from 1.86 percent of GDP in 1964 to just 0.66 percent of GDP in 2016.2 The Creating Helpful Incentives to Produce Semiconductors(CHIPS)and Science Act of 2022 was supposed to turn the tide

37、by dramatically raising funding for basic research,but major increases were subsequently scrapped in budget negotiations.The United States still funds more basic research than China does,but Chinese investment is rising six times faster and is expected to overtake US spending within a decade.3Althou

38、gh private-sector investment in technol-ogy companies and associated university research has increased substantially,it is not a substitute for federal funding,which supports university R&D directed at national and public issues,not commer-cial viability.4To be sure,the rising dominance of private i

39、ndus-try in innovation brings significant benefits.But it is also generating serious and more hidden risks to the health of the entire American innovation ecosystem.Technology and talent are migrating from academia to the private sector,accelerating the development of commercial products while erodi

40、ng the founda-tion for the future.We are already reaching a tip-ping point in AI.In 2022,more than 70 percent of students who received PhDs in artificial intelligence at US universities took industry jobs,leaving fewer faculty to teach the next generation.5 As the bipar-tisan National Security Commi

41、ssion on Artificial Intelligence put it,“Talent follows talent.”67 FOREWORDdriving Google;to optogenetics,a technique pio-neered in 2005 that uses light to control neurons,enabling precise studies of brain function.In this moment of rapid technological change,we must do even more to connect emerging

42、 technologies with policy.We are proud and excited to highlight this collaboration between Stanfords Hoover Institution and the School of Engineering to bring policy analy-sis,social science,science,medicine,and engineer-ing together in new ways.Today,technology policy and education efforts are ofte

43、n led by policy experts with limited technolog-ical expertise.The Stanford Emerging Technology Review flips the script,enlisting ten of the brightest scientific and engineering minds at the university to share their knowledge of their respective fields by working alongside social scientists to trans

44、late their work to nonexpert audiences.We start with science and technology,not policy.And we go from there to emphasize the important interaction between sci-ence and all aspects of policy.Start-ups take flight in our dorm rooms,classrooms,laboratories,and kitchens.Technological innovation is lived

45、 every day and up close on our campus with all its benefits and downsides.This ecosystem and its culture,ideas,and perspectives often seem a world apart from the needs and norms of Washington,DC.Bridging the divide between the locus of American policy and the heart of American technological inno-vat

46、ion has never been more important.Stanford has a rich history of policy engagement,with individuals who serve at the highest levels of government as well as institutional initiatives that bring together policymakers and researchers to tackle the worlds toughest policy problems.And as Stanfords Schoo

47、l of Engineering celebrates its one hundredth anniversary in 2025,we are reminded of the profound impact that generations of Stanford faculty,students,and staff have had through their discoveries from the klystron,a microwave ampli-fier developed in the 1930s that enabled radar and early satellite c

48、ommunications;to the algorithms Employment of new AI PhDs(%of total)in the United States and Canada by sector,201022IndustryAcademiaNew AI PhD graduates(%of total)201020112012201320142015201620172018201920202021202270%60%50%40%30%20%10%0%Government70.71%19.95%0.76%Source:Adapted from Nestor Maslej,L

49、oredana Fattorini,Raymond Perrault,et al.,The AI Index 2024 Annual Report,AI Index Steering Committee,Institute for Human-Centered AI,Stanford University,Stanford,CA,April 2024.Data from CRA Taulbee Survey,2023FIGURE F.1Most new AI PhDs hired in North America are flocking to industry 8STANFORD EMERG

50、ING TECHNOLOGY REVIEWdiscussion.The themes include broad trends,like the tendency for technological breakthroughs to come in fits,starts,and lengthy plateaus that are extremely difficult even for leaders in those fields to predict.(AI leaders have experienced several so-called AI winters over decade

51、s as well as moments of profound and sudden progress like the 2022 release of ChatGPT.)They include enduring and widespread technolog-ical challenges like cybersecurity.And they include cognitive blind spots like frontier bias the natural but mistaken assumption that the only transforma-tional techn

52、ologies sit on the frontiers of a field.For each of the ten technology chapters,reviews of the field were led by world-renowned tenured Stanford faculty members who also delivered sem-inars with other faculty discussants within and out-side their areas of expertise.(SETR contributors and their field

53、s are listed at the end of each chapter.)The SETR team also involved more than a dozen postdoctoral scholars and undergraduate research assistants who interviewed faculty across Stanford and drafted background materials.Each technology chapter begins with an overview of the basics the major technica

54、l subfields,concepts,and terms needed to understand how a technol-ogy works and could affect society.Next,we out-line important developments and advances in the field.Finally,each chapter concludes by offering an over-the-horizon outlook that covers the most cru-cial considerations for policymakers

55、over the next few years including technical as well as policy,legal,and regulatory issues.The report ends with a chapter that looks across the ten technologies,offer-ing analysis of implications for economic growth,national security,environmental and energy sustain-ability,human health,and civil soc

56、iety.Three points bear highlighting.First,we offer no spe-cific policy recommendations in this report.That is by design.Washington is littered with reports offering policy recommendations that were long forgotten,overtaken by events,or both.Opinions are plentiful.Expert insights based on leading res

57、earch are not.How to Use This Report:One Primer,Ten Major Technology AreasThis report is intended to be a one-stop-shopping primer that covers developments and implications in ten major emerging technology areas:AI,biotech-nology and synthetic biology,cryptography,lasers,materials science,neuroscien

58、ce,robotics,semicon-ductors,space,and sustainable energy technologies.The list is broad by design,and it includes fields that are widely regarded as pivotal to shaping soci-ety,economics,and geopolitics today and into the future.That said,the ten major technology areas covered in this report are now

59、here near an exhaustive catalogue of technology research areas at Stanford.And the list may change year to year not because a particular technology sputtered or because we got it wrong,but because categorizing technologies is inherently dynamic.Limiting this report to ten areas imposes discipline on

60、 what we cover and how deeply we go.We seek to highlight relationships among technolo-gies in ways that may not be obvious:Quantum com-puting,for example,is an important field but does not have its own chapter.Instead,it is covered within the semiconductor chapter because we wanted to emphasize that

61、 even if quantum breakthroughs are realized,they will not address many important com-puting needs and challenges.Of note,nine of the ten technology chapters appearing in this edition are on the same subjects as in our previous report.In this report,we have combined nuclear energy and sustainable ene

62、rgy technologies into a single chap-ter and added a chapter on lasers.Many of the most important issues cut across tech-nological fields.We have expanded our previous reports crosscutting themes chapter to highlight four-teen of these themes and offer more examples and 9 FOREWORDWe aim to provide a

63、reference resource that is both timeless and timely,an annual state-of-the-art guide that can inform successive generations of policymak-ers about how to think about evolving technological fields and their implications.Individual SETR faculty may well have views about what should be done.Some of us

64、engage in policy writing and advising.But the mission of this collective report is informing,not advocating.We encourage readers interested in learning more about specific fields and policy ideas to contact our team at SETReview2025stanford.edu.Second,SETR offers a view from Stanford,not the view fr

65、om Stanford.There is no single view of any-thing in a university.Faculty involved in this report may not agree with everything in it.Their colleagues would probably offer a different lay of the technology landscape with varying assessments about impor-tant developments and over-the-horizon issues.Th

66、e report is intended to reflect the best collective judg-ment about the state of these ten fields guided by leading experts in them.Third,this report is intended to be the introductory product that translates a broad swatch of tech-nological research for nontechnical readers.Other SETR offerings pro

67、vide deeper dives into specific technological areas that should be of interest for subject-matter experts.Ensuring continued American leadership in science and technology is essential,and its a team effort.We hope this edition of the Stanford Emerging Technology Review continues to spark meaning-ful

68、 dialogue,better policy,and lasting impact.The promise of emerging technology is boundless if we have the foresight to understand it and the fortitude to embrace the challenges.Condoleezza RiceJohn B.TaylorJennifer WidomAmy ZegartCo-chairs,Stanford Emerging Technology ReviewNOTES1.Simon Baker,“China

69、 Overtakes United States on Contribution to Research in Nature Index,”Nature,May 19,2023,https:/doi.org/10.1038/d41586-023-01705-7.2.Council on Foreign Relations,Innovation and National Security:Keeping Our Edge,Independent Task Force Report No.77,James Manyika and William H.McRaven,chairs,2019,10,h

70、ttps:/www.cfr.org/keeping-our-edge/pdf/TFR_Innovation_Strategy.pdf.3.OECD,“OECD Data Explorer:Gross Domestic Expenditure on R&D by Sector of Performance and Type of R&D,”OECD.org,2024,https:/data-viewer.oecd.org/?chartId=b08256f3-6000-42a5-ad81-b1ab5bd8ac6a.Analysis was conducted by comparing two me

71、thods of measuring annual increases of US and China spending on basic research and projecting rates into future.4.Council on Foreign Relations,Innovation,21.5.Nestor Maslej,Loredana Fattorini,Raymond Perrault,et al.,The AI Index 2024 Annual Report,AI Index Steering Commit-tee,Institute for Human-Cen

72、tered AI,Stanford University,Stan-ford,CA,April 2024,https:/aiindex.stanford.edu/wp-content/uploads/2024/05/HAI_AI-Index-Report-2024.pdf.6.Eric Schmidt and Robert Work,NSCAI Interim Report,National Security Commission on Artificial Intelligence,2020,30,https:/apps.dtic.mil/sti/pdfs/AD1112059.pdf.7.P

73、oornima Apte,“Princeton Invests in New 300-GPU Cluster for Academic AI Research,”AI at Princeton,March 15,2024,https:/ai.princeton.edu/news/2024/princeton-invests-new-300-gpu-cluster-academic-ai-research;Michael Kan,“Zuckerbergs Meta Is Spending Billions to Buy 350,000 Nvidia H100 GPUs,”PCMAG,Jan-ua

74、ry 18,2024,https:/ Leswing,“Nvidias Latest AI Chip Will Cost More than$30,000,CEO Says,”CNBC,March 19,2024,https:/ Jurowetzki,Daniel Hain,Juan Mateos-Garcia,and Kon-stantinos Stathoulopoulos,“The Privatization of AI Research(-ers):Causes and Potential Consequences From University-Industry Interactio

75、n to Public Research Brain-Drain?,”arXiv,2021,https:/arxiv.org/ftp/arxiv/papers/2102/2102.01648.pdf.10US government departments.However,SETR focus technologies are likely to change over time,not because we were incorrect,but because science and technology never sleep,the borders between fields are p

76、orous,and different people categorize similar research in different ways.Report DesignThis report is organized principally by technology,with each area covered in a standalone chapter that gives an overview of the field,highlights key devel-opments,and offers an over-the-horizon view of important te

77、chnological and policy considerations.Although these chapters can be read individually,one of the most important and unusual hallmarks of this moment is convergence:Emerging technologies are intersecting and interacting in a host of ways,with important implications for policy.We examine these broade

78、r dynamics in chapters11 and 12.In chapter11,we describe a number of themes and commonalities that cut across many of the technolo-gies we describe earlier in the report.In chapter12,we consolidate technological developments across all ten areas and discuss how they apply to five policy domains:econ

79、omic growth,national security,environmental and energy sustainability,health and medicine,and civil society.Three tensions run throughout and are worth keep-ing in mind:1.Timeliness and timelessness Each chapter seeks to strike a balance between covering recent developments in science and in the hea

80、dlines and providing essential knowledge about how a field Emerging technologies have never been more important or difficult to understand.Breakthrough advances seem to be everywhere,from ChatGPT to the COVID-19 mRNA vaccines to constellations of cheap commercial shoebox-size satellites that can tra

81、ck events on Earth in nearreal time.This is a pivotal technological moment offering both tre-mendous promise and unprecedented challenges.Policymakers need better expert resources to help them understand the burgeoning and complex array of technological developments more easily and more continuously

82、.The Stanford Emerging Technology Review is designed to meet this need,offering an easy-to-use reference tool that harnesses the expertise of Stanford Universitys leading science and engineer-ing faculty in ten major technological areas.SETR 2025 FOCUS TECHNOLOGIESArtificial IntelligenceBiotechnolog

83、y and Synthetic BiologyCryptographyLasersMaterials ScienceNeuroscienceRoboticsSemiconductorsSpaceSustainable Energy TechnologiesThese particular fields were chosen for this report because they leverage areas of deep expertise at Stanford and cover many critical and emerging technologies identified b

84、y the Office of Science and Technology Policy in the White House and other EXECUTIVE SUMMARYSTANFORD EMERGING TECHNOLOGY REVIEW11learning,interacting,problem-solving,and even exercising creativity.In the past year,the main AI-related headlines have been the rise of large language models(LLMs)like GP

85、T-4,on which some versions of the chatbot ChatGPT are based,and the recognition of AIs significance through the award-ing of two Nobel Prizes in physics and chemistry for AI-related work.KEY CHAPTER TAKEAWAYS Artificial intelligence(AI)is a foundational tech-nology that is supercharging other scient

86、ific fields and,like electricity and the internet,has the potential to transform societies,economies,and politics worldwide.Despite rapid progress in the past several years,even the most advanced AI still has many failure modes that are unpredictable,not widely appre-ciated,not easily fixed,not expl

87、ainable,and capable of leading to unintended consequences.Mandatory governance regimes for AI,even those to stave off catastrophic risks,will face stiff opposition from AI researchers and companies,but voluntary regimes calling for self-governance are more likely to gain support.Biotechnology and Sy

88、nthetic BiologyBiotechnology is the use of cellular and biomo-lecular processes to develop products or services.Synthetic biology is a subset of biotechnology that involves using engineering tools to modify or create biological functions like creating a bacterium that can glow in the presence of exp

89、losives.Synthetic biology is what created the COVID-19 mRNA vac-cine in record time although it relied on decades works,what is important within it,and what chal-lenges lie ahead.2.Technical depth and breadth This report inten-tionally skews toward breadth,offering a 30,000-foot view of a vast techn

90、ological landscape in one compendium.Readers should consider it an intro-ductory course.Other products and/or educational tools will be released in the months ahead that will offer additional insights into each field.3.Technical and nontechnical aspects of inno-vation We start with the science but d

91、o not end with the science.Technological breakthroughs are necessary but not sufficient conditions for successful innovation.Economic,political,and societal factors play enormous and often hidden roles.Johannes Gutenberg invented the printing press in 1452,but it took more than 150 years before the

92、Dutch invented the first successful newspapers not because they perfected the mechanics of movable type,but because they decided to use less paper,making newspapers sustainably profitable for the first time.1 Each chapter in this report was written with an eye toward highlighting important economic,

93、political,policy,legal,and societal factors likely to impede,shape,or accelerate progress.Technologies and Takeaways at a GlanceArtificial Intelligence Artificial intelligence(AI)is a computers ability to perform some of the functions associated with the human brain,including perceiving,reasoning,Ex

94、ECUTIVE SUMMARY12STANFORD EMERGING TECHNOLOGY REVIEWprovenance of information,identity management,supply chain management,and cryptocurrencies.KEY CHAPTER TAKEAWAYS Cryptography is essential for protecting infor-mation,but alone it cannot secure cyberspace against all threats.Cryptography is the ena

95、bling technology of blockchain,which is the enabling technology of cryptocurrencies.Central bank digital currencies(CBDCs)are a particular type of cryptography-based digital currency supported by states and one that could enhance financial inclusion.Although the United States lags some countries in

96、experimenting with a CBDC,it may benefit from a cautious,well-timed approach by learning from other nationsefforts.LasersImprovements in laser technology since its inven-tion in 1960 have enabled light to be manipulated and used in previously unimaginable ways.It is now applied so broadly that it ca

97、n be considered an enabling technology one whose existence and characteristics enable other applications that would not be feasible and/or affordable without it.New lasers are being developed by researchers and companies across a wider range of light wave-lengths,which will make the devices even mor

98、e useful.KEY CHAPTER TAKEAWAYS Laser technology has become essential for a wide range of applications,including communications,high-end chip production,defense,manufactur-ing,and medicine.Because advances in laser technology tend to occur in the context of specific applications,laser of earlier rese

99、arch.Just as rockets enabled humans to overcome the constraints of gravity to explore the universe,synthetic biology is enabling humans to overcome the constraints of lineage to develop new living organisms.KEY CHAPTER TAKEAWAYS Biotechnology is poised to emerge as a general-purpose technology by wh

100、ich anything bioen-gineers learn to encode in DNA can be grown whenever and wherever needed essentially enabling the production of a wide range of prod-ucts through biological processes across multiple sectors.The US government is still working to grasp the scale of this bio-opportunity and has reli

101、ed too heavily on private-sector investment to support the foundational technology innovation needed to unlock and sustain progress.Biotechnology is one of the most important areas of technological competition between the United States and China,and China is investing consid-erably more resources.La

102、cking equivalent efforts domestically,the United States runs the risk of Sputnik-like strategic surprises in biotechnology.CryptographyThe word cryptography originates from Greek words that mean“secret writing.”In ancient times,cryp-tography involved the use of ciphers and secret codes.Today,it reli

103、es on sophisticated mathemat-ical models to protect data from being altered or accessed inappropriately.Cryptography is often invisible,but it is essential for most internet activi-ties,such as messaging,e-commerce,and banking.In recent years,a type of cryptographic technol-ogy called blockchain whi

104、ch records transac-tions in distributed ledgers in the computing cloud that cannot be altered retroactively without being detected has been used for a variety of applica-tions,including time-stamping and ensuring the 13 ExECUTIVE SUMMARYdevelopment to degeneration in later years.The brain is perhaps

105、 the least understood and yet most important organ in the human body.Three major research subfields of neuroscience are neuroengi-neering(e.g.,brain-machine interfaces),neurohealth(e.g.,brain degeneration and aging),and neurodis-covery(e.g.,the science of addiction).KEY CHAPTER TAKEAWAYS Popular int

106、erest in neuroscience vastly exceeds the actual current scientific understanding of the brain,giving rise to overhyped claims in the public domain that revolutionary advances are just around the corner.Advances in human genetics and experimental neuroscience,along with computing and neuro-science th

107、eory,have led to some progress in sev-eral areas,including understanding and treating addiction and neurodegenerative diseases and designing brain-machine interfaces for restoring vision.American leadership is essential for establishing and upholding global norms about ethics and human subjects rese

108、arch in neuroscience,but this leadership is slipping with decreased strate-gic planning and increased foreign investments in the field.RoboticsRobotics is an integrative field that draws on advances in multiple technologies rather than a single disci-pline.The question“What is a robot?”is harder to

109、answer than it appears.At a minimum,the emerg-ing consensus among researchers is that a robot is a physical entity that has ways of sensing itself and the world around it and can create physical effects on that world.Robots are already used across a range of sectors in a variety of ways including as

110、sembly-line manufacturing,space exploration,autonomous vehi-cles,tele-operated surgery,military reconnaissance,and disaster assistance.technology research and development is widely dispersed among different types of laboratories and facilities.Broad investment in next-generation lasers holds the pot

111、ential to improve progress in nuclear fusion energy technology,weapons develop-ment,and quantum communication.Materials ScienceMaterials science studies the structure and proper-ties of materials from those visible to the naked eye to microscopic features and how they can be engineered to change per

112、formance.Contributions to the field have led to better semiconductors,“smart bandages”with integrated sensors and sim-ulators to accelerate healing,more easily recycla-ble plastics,and more energy-efficient solar cells.Materials science has also been key to the devel-opment of additive manufacturing

113、,often known as 3-D printing.KEY CHAPTER TAKEAWAYS Materials science is a foundational technology that underlies advances in many other fields,including robotics,space,energy,and synthetic biology.Materials science will exploit artificial intelligence as another promising tool to predict new mate-ri

114、als with new properties and identify novel uses for known materials.Future progress in materials science requires new funding mechanisms to more effectively tran-sition from innovation to implementation and access to more computational power.NeuroscienceNeuroscience is the study of the human brain a

115、nd the nervous system its structure,function,healthy and diseased states,and life cycle from embryonic 14STANFORD EMERGING TECHNOLOGY REVIEWKEY CHAPTER TAKEAWAYS The growing demand for artificial intelligence and machine learning is driving innovations in chip fabrication that are essential for enha

116、ncing com-putational power and managing energy efficiency.Advances in memory technologies and high-bandwidth interconnects,including photonic links,are critical for meeting the increasing data needs of modern applications.Even if quantum computing advancements are realized,the United States will sti

117、ll need compre-hensive innovation across the technology stack to continue to scale the power of information technology.SpaceSpace technologies include any technology devel-oped to conduct or support activities approximately sixty miles or more beyond Earths atmosphere.A single space mission is a sys

118、tem of systems including everything from the spacecraft itself to propulsion,data storage and processing,electrical power gener-ation and distribution,thermal control to ensure that components are within their operational and sur-vival limits,and ground stations.While in the past,space was the exclu

119、sive province of government spy satellites and discovery missions,the number and capabilities of commercial satellites have increased dramatically in recent years.Today,around ten thou-sand working satellites,many no larger than a loaf of bread,circle the planet.Some operate in constel-lations that

120、can revisit the same location multiple times a day and offer image resolutions so sharp they can identify different car models driving on a road.KEY CHAPTER TAKEAWAYS A burgeoning“NewSpace”economy driven by private innovation and investment is transforming KEY CHAPTER TAKEAWAYS Future robots may be

121、useful for improving the US manufacturing base,reducing supply chain vulnerabilities,delivering eldercare,enhancing food production,tackling the housing shortage,improving energy sustainability,and performing almost any task involving physical presence.Progress in artificial intelligence holds the p

122、oten-tial to advance robotics significantly but also raises ethical concerns that are essential to address,including the privacy of data used to train robots,data bias that could lead to physical harm by robots,and other safety issues.Achieving the full potential of robots will require a major push

123、from the federal government and the private sector to improve robotics adoption and research across the nation.SemiconductorsSemiconductors,or chips,are crucial and ubiqui-tous components used in everything from refrig-erators and toys to smartphones,cars,computers,and fighter jets.Chip production i

124、nvolves two distinct steps:(1)design,which requires talented engineers to design complex integrated circuits involving millions of components;and(2)fabri-cation,which is the task of manufacturing chips in large,specially designed factories called“fabs.”Because fabs involve highly specialized equipme

125、nt and facilities,they can cost billions of dollars.US companies still play a leading role in semiconduc-tor design,but US semiconductor-manufacturing capacity has plummeted,leaving the country heav-ily dependent on foreign chips,most notably from Taiwan.The Creating Helpful Incentives to Produce Se

126、miconductors(CHIPS)and Science Act of 2022 was intended to help the US semiconductor indus-try regain a foothold in fabrication,but progress will take years,if not decades.15 ExECUTIVE SUMMARYvehicles,and transitioning residential and com-mercial heating and industrial energy.In the long term,techno

127、logies for decarboniz-ing buses and long-haul trucks,decarbonizing carbon-intensive industries,and reducing green-house gases from refrigerants and agriculture will play key roles in a net-zero,emissions-free energy infrastructure.Important Crosscutting ThemesChapter11 discusses fourteen themes that

128、 cut across the technological areas.We split these themes into two categories.Category 1:Key Observations About How Technologies Evolve over Time1.The Goldilocks challenge:moving too quickly,moving too slowly.Innovation that emerges too fast threatens to disrupt the status quo around which many nati

129、onal,organizational,and personal inter-ests have coalesced.It is also more likely to lead to unintended consequences and give short shrift to security,safety,ethics,and geopolitics.Innovation that moves too slowly increases the likelihood that a nation will lose the technical,economic,and national s

130、ecurity advantages that often accrue to first movers in a field.2.There is a trend toward increasing access to new technologies worldwide.Even innovations that are US born are unlikely to remain in the exclu-sive control of American actors for long periods.3.The synergies between different technolog

131、ies are large and growing.Advances in one tech-nology domain often support advances in other technologies.space launch,vehicles,communications,and key space actors in a domain that has until now been dominated by superpower governments.Space is a finite planetary resource.Because of dramatic increas

132、es in satellites,debris,and geo-political space competition,new technologies and new international policy frameworks will be needed to prevent and manage international conflict in space and ensure responsible steward-ship of this global commons.A race to establish a permanent human presence on the M

133、oon is underway,with serious concerns that,despite Outer Space Treaty prohibitions against it,the first nation to reach the Moon may be in a strong position to prevent others from establishing their own lunar presences.Sustainable Energy TechnologiesThis vital strategic resource for nations typicall

134、y involves generation,transmission,and storage.In recent years it has also come to include carbon capture and carbons removal from the atmosphere.Energy mix and innovation are key to efforts to address climate change.Success will also depend on tackling challenges such as decentralizing and moderniz

135、ing electricity grids and achieving greater national consensus about energy goals to enable strategic and effective R&D programs and funding.KEY CHAPTER TAKEAWAYS Although many clean energy technologies are now available and increasingly affordable,scaling them to a meaningful degree and building th

136、e massive infrastructure needed to deploy them will take decades.The largest impact on reducing emissions in the near to medium term will come from building a no-to very-low-emission electricity grid,elec-trifying passenger cars and small commercial 16STANFORD EMERGING TECHNOLOGY REVIEWCategory 2:Co

137、mmon Innovation Enablers and Inhibitors1.Ideas and human talent play a central role in scientific discovery and cannot be manufactured at will.They must be either domestically nurtured or imported from abroad.Today,both paths for generating ideas and human talent face serious and rising challenges.2

138、.A policy bias toward science or technology at the frontiers of knowledge tends to overestimate the benefits accruing from such advances,at least in the short term.Many technologies with transfor-mational potential are not necessarily on the techni-cal frontier,and frontier bias carries with it the

139、risk of overlooking older technologies that can be used in novel and impactful ways.3.Good public policy anticipates wide variations in perspectives on any given technology.When everyone in a decision-making organization shares similar perspectives on technology creating ana-lytical blind spots and

140、potentially groupthink,or unwarranted conformity in beliefs the risks associ-ated with innovation can be underestimated.4.US universities play a pivotal role in the inno-vation ecosystem that is increasingly at risk.Although the US government frequently talks about the importance of public-private p

141、artnerships in emerging technology,universities also play a pivotal and often underappreciated role.They are the only organizations with the mission of pursuing high-risk research that may not pay off commercially for a long time,if ever.That high-risk focus has yielded high-benefit payoffs in a wid

142、e range of fields.5.Sustaining American innovation requires long-term government R&D.Investments with clear strategies and sustained priorities not the increas-ingly common wild swings in research funding from year to year are crucial.4.The path from research to application is often not linear.Many

143、believe that technological break-throughs arise from a step-by-step linear progres-sion where basic research leads to applied research,which then leads to development and prototyping and finally to a marketable product.Yet innovation often does not work this way.Many scientific devel-opments enhance

144、 understanding but never advance to the marketplace.Many marketable products emerge in a nonlinear fashion,after many rounds of feedback between phases.Other products emerge only when several different technologies reach some level of maturity.5.The speed of change is hard even for leading researche

145、rs to anticipate.Technology often pro-gresses in fits and starts,with long periods of incre-mental results followed by sudden breakthroughs.6.Nontechnical factors often determine whether new technologies succeed or fail.Adoption of novel technologies hinges on economic viability and societal accepta

146、bility,not just scientific proof of con-cept and engineering feasibility.7.The US government is no longer the primary driver of technological innovation or funder of research and development.Historically,technolog-ical advances(including semiconductors,the inter-net,and jet engines)were funded and a

147、dvocated for by the US government.Today,private-sector R&D investment is playing a much larger role,rais-ing important concerns about how to ensure that US national interests are properly taken into account and that basic science which is an important foun-dation for future innovation remains strong

148、.8.Technological innovation occurs in both democ-racies and autocracies,but different regime types enjoy different advantages and challenges.Democracies provide greater freedom for explo-ration,while authoritarian regimes can direct sus-tained funding and focus toward the technologies they believe a

149、re most important.17 ExECUTIVE SUMMARY6.Cybersecurity is an enduring concern for every aspect of emerging technology research.State and nonstate actors will continue to threaten the confidentiality,integrity,and availability of informa-tion that is crucial for emerging technology R&D.Finally,each of

150、 the ten technology fields covered in this report bears on five policy areas that are of interest to policymakers:economic growth,national security,environmental and energy sustainability,health and medicine,and civil society.Chapter12 identifies applications and consequences of each field as they a

151、pply to these policy areas.NOTES1.Andrew Pettegree and Arthur der Weduwen,The Bookshop of the World:Making and Trading Books in the Dutch Golden Age(New Haven,CT:Yale University Press,2019),7072.18diseases,energy supply and demand,and sustain-able development.3S&T is one important battleground for s

152、eeking ad vantage in geopolitical competition,as advances in scientific and technical fields can contribute to national interests,including a stronger national security posture,greater national pride and self-confidence,economic influence,and diplo matic leverage.But four other points about S&T are

153、equally important:Advances in S&T must be leveraged alongside strong public policy if those advances are to serve the national interest.Coupling advanced technology with poor policy to influence that technology rarely ends well.Advantages gained from S&T advances are tran-sient in the long run.Attem

154、pting to restrict the transfer of scientific and technical knowledge to other nations may delay its spread,but the first successful demonstration of a technological advance on the part of the United States is often the impetus for other nations to launch their own efforts to catch up.Internationally

155、,S&T is not always a zero-sum game,as advances originating in one nation often benefit others.For example,the internet and what most people know simply as GPS navi-gation are US-born innovations whose uses have spread around the world and the United States itself has gained from that spread.Internat

156、ional competition does not occur only with adversaries.Our allies and partners also compete in the S&T space,developing technol-ogy or deploying policy that can leave the United States at a disadvantage.The Role of Science and Technology in Advancing National InterestsVannevar Bush,an engineer and p

157、olicymaker who oversaw the development of the Manhattan Project,was the nations first presidential science advisor.In 1945,he wrote,“Advances in science when put to practical use mean more jobs,higher wages,shorter hours,more abundant crops,more leisure for rec-reation,for study,for learning how to

158、live without the deadening drudgery which has been the burden of the common man for ages past.Advances in science will also bring higher standards of living,will lead to the prevention or cure of diseases,will promote conservation of our limited national resources,and will assure means of defense ag

159、ainst aggression.”1Science and technology(S&T)remain essential to our national interests.Advances in S&T are closely tied to national needs in transportation,agricul-ture,communication,energy,education,environ-ment,health,and defenseas well as to millions of American jobs.S&T also underpins and driv

160、es many strategic objectives in foreign policy,such as reducing the proliferation of weapons of mass destruction,strengthening relationships with allies and partners,improving humanitarian assistance,and promoting growth in developing and transi-tional economies.2 Research and development in S&T fie

161、lds such as information technology,biotech-nology,materials science,and lasers will impact both“hard power”issues defense,arms control,nonproliferation and“soft power”concerns,such as climate change,infectious and chronic INTRODUCTIONSTANFORD EMERGING TECHNOLOGY REVIEW19academia and think tanks,as w

162、ell as discussions with science and engineering colleagues at Stanford University and other research universities.We do not claim that any one of these ten is more important than the others,and the discussion below addresses subjects in alphabetical order.Indeed,one of the unexpected aspects of this

163、 technological moment is convergence:New technologies are intersecting,overlapping,and driving each other in all sorts of ways some obvious,some more hidden.The description of each field is divided into three parts.The first part is an overview of the field.The second part addresses noteworthy key d

164、evelop-ments in the domain that are relevant to under-standing the field from a policy perspective.The last part,providing an over-the-horizon perspective,is itself subdivided into three sections:the poten-tial impact of the field in the future(i.e.,the fields potential over-the-horizon impact);the

165、likely chal-lenges facing innovation and implementation;and relevant policy,legal,and regulatory issues.Policy for Science and TechnologyPolicymakers have a wide variety of tools to influence the conduct of S&T research and development.Many of these are obvious,such as research funding,tax incentive

166、s to firms,intellectual property rights,export controls,classification authority,4 regulation,public procurement,funding and other aid to strategic sec-tors,as well as labor force training and education.On the other hand,policy need not be directed at S&T to have a meaningful impact.For example,immi

167、gration policy is not primarily directed at the S&T workforce,but it can have profound effects on the talent available to academic and industry research.Policy oriented in one direction attracts talent to the United States,while policy oriented in another diminishes such talent.Or consider the natio

168、nal economic environment.Stable fiscal and monetary policies make it easier for private-sector decision makers to plan and invest for the long term a criti-cal consideration when many S&T advances must be nurtured along an extended path from conception to maturity.Ten Science and TechnologyFieldsCha

169、pters 1 through 10 describe in more detail ten S&T fields important to the national agenda.Our selection of these fields was driven by several fac-tors:inclusion on common lists of key technologies developed by government,the private sector,and NOTES1.Vannevar Bush,Science:The Endless Frontier(Washi

170、ngton,DC:US Government Printing Office,1945),https:/www.nsf.gov/od/lpa/nsf50/vbush1945.htm.2.National Research Council,The Pervasive Role of Science,Technology,and Health in Foreign Policy:Imperatives for the Department of State(Washington,DC:National Academies Press,1999),https:/doi.org/10.17226/96

171、88.3.National Intelligence Council,Global Trends 2015:A Dia-logue About the Future with Nongovernment Officials,Decem-ber 2000,https:/www.dni.gov/files/documents/Global%20Trends _2015%20Report.pdf;National Intelligence Council,Mapping the Global Future:Report of the National Intelligence Councils 20

172、20 Project,December 2004,https:/www.dni.gov/files/documents/Global%20Trends_Mapping%20the%20Global%20Future%202020%20Project.pdf.4.Under some circumstances(such as when federal funding is involved),the US government may have the authority to classify research even if that research was performed with

173、out access to classified information.INTRODUCTION21OverviewArtificial intelligence(AI),a term coined by computer scientist and Stanford professor John McCarthy in 1955,was originally defined as“the science and engineering of making intelligent machines.”In turn,intelligence might be defined as the a

174、bility to learn and perform suitable techniques to solve prob-lems and achieve goals,appropriate to the context in an uncertain,ever-varying world.1 AI could be said to refer to a computers ability to display this type of intelligence.The emphasis today in AI is on machines that can learn as well as

175、 humans can learn,or at least some-what comparably so.However,because machines are not limited by the constraints of human biology,AI systems may be able to run at much higher speeds and digest larger volumes and types of information than are possible with human capabilities.KEY TAKEAWAYS Artificial

176、 intelligence(AI)is a foundational tech-nology that is supercharging other scientific fields and,like electricity and the internet,has the potential to transform societies,economies,and politics worldwide.Despite rapid progress in the past several years,even the most advanced AI still has many failu

177、re modes that are unpredictable,not widely appre-ciated,not easily fixed,not explainable,and capable of leading to unintended consequences.Mandatory governance regimes for AI,even those to stave off catastrophic risks,will face stiff opposition from AI researchers and companies,but voluntary regimes

178、 calling for self-governance are more likely to gain support.ARTIFICIAL INTELLIGENCE0122STANFORD EMERGING TECHNOLOGY REVIEWproductivity growth by 1.5percent over a ten-year period if it is adopted widely.10 Private funding for generative AI start-ups surged to$25.2 billion in 2023,a nearly ninefold

179、increase from 2022,and accounted for around a quarter of all private invest-ments related to AI in 2023.11The question of what subfields are considered part of AI is a matter of ongoing debate,and the bound-aries between these fields are often fluid.Some of the core subfields are the following:Compu

180、ter vision,enabling machines to recog-nize and understand visual information from the world,convert it into digital data,and make deci-sions based on these data Machine learning(ML),enabling computers to perform tasks without explicit instructions,often by generalizing from patterns in data.This inc

181、ludes deep learning that relies on multilayered artificial neural networks which process infor-mation in a way inspired by the human brain to model and understand complex relationships within data.Natural language processing,equipping machines with capabilities to understand,interpret,and produce sp

182、oken words and written textsMost of todays AI is based on ML,though it draws on other subfields as well.ML requires data and computing power often called compute12 and much of todays AI research requires access to these on an enormous scale.In October 2024,the Royal Swedish Academy of Sciences award

183、ed the Nobel Prize in Physics for 2024 to John Hopfield and Geoffrey Hinton for their work in applying tools and concepts from sta-tistical mechanics to develop“foundational discov-eries and inventions that enable machine learning with artificial neural networks”13(further discussed below).Underscor

184、ing the importance of AI-based techniques in advancing science,it also awarded the Today,AI promises to be a fundamental enabler of technological advancement in many fields,arguably of comparable importance to electricity in an earlier era or the internet in more recent years.The science of computin

185、g,worldwide availability of networks,and civilization-scale data all that collectively underlies the AI of today and tomorrow are poised to have similar impact on technological progress in the future.Moreover,the users of AI will not be lim-ited to those with specialized training;instead,the average

186、 person on the street will increasingly inter-act directly with sophisticated AI applications for a multitude of everyday activities.The global AI market was worth$196.63billion in 2023,with North America receiving 30.9percent of total AI revenues.2 The Stanford Institute for Human-Centered Artifici

187、al Intelligence(HAI)AI Index 2024 Annual Report found that private investment in all AI start-ups totaled$95.99 billion in 2023,marking the second consecutive year of decline since a record high of over$120 billion in 2021.3 Amid a 42 per-cent fall in overall global venture funding across all sector

188、s in 2023,4 AI start-ups raised$42.5 billion in venture capital that year,marking only a 10 percent decrease5 from 2022.6 Many tech companies are significantly ramping up investments in AI infrastructure,such as larger and more powerful computing clusters to meet the grow-ing demand for AI capabilit

189、ies.Companies such as Amazon and Meta have begun revamping their data centers,7 and BlackRock,Microsoft,and the technol-ogy investor MGX,which is backed by the United Arab Emirates,announced in September 2024 the new Global AI Infrastructure Investment Partnership fund,which seeks to raise$30 billio

190、n in private equity cap-ital to finance data centers and other projects that span the AI infrastructure ecosystem.8 The fund may ultimately invest up to$100 billion over time.9One estimate forecasts that generative AI which can create novel text,images,and audio output and is discussed in more detai

191、l later in this chapter could raise global GDP by$7 trillion and raise 2301ARTIFICIAL INTELLIGENCEhardware components were likely also needed suggests the overall hardware costs for GPT-4 were at least a few hundred million dollars.And the chips underlying this hardware are specialty chips often fab

192、ricated offshore.16(Chapter8 on semiconductors discusses this point at greater length.)Lastly,AI models consume a lot of energy.Consider first the training phase:One estimate of the elec-tricity required to train a foundation model such as GPT-4 pegs the figure at about fifty million kilowatt-hours(

193、kWh).17 The average American house-hold uses about 11,000 kWh per year,meaning the energy needed to train GPT-4 was approximately the same as that used by 4,500 average homes in a year.Paying for this energy adds significant cost,even before a single person actually uses a model.Then,once a model is

194、 up and running,the cost of energy used to power queries can add up fast.This is known as the inference phase.For ChatGPT,the energy used per query is around 0.002 of a kilowatt-hour,or 2 watt-hours.18(For comparison,a single Google search requires about 0.3 watt-hours,19 and an alkaline AAA battery

195、 contains about 2 watt-hours of energy.)Given hundreds of millions of queries per day,the operating energy require-ment of ChatGPT might be a few hundred thousand kilowatt-hours per day,at a cost of several tens of thousands of dollars.AI can automate a wide range of tasks.But it also has particular

196、 promise in augmenting human capa-bilities and further enabling people to do what they are best at doing.20 AI systems can work alongside humans,complementing and assisting their work rather than replacing them.Some present-day examples are discussed below.Healthcare Medical diagnostics An AI system

197、 that can pre-dict and detect the onset of strokes qualified for Medicare reimbursement in 2020.21Nobel Prize in Chemistry for 2024 to Demis Hassabis and John M.Jumper for AI-based protein structure prediction,14 an important and long-standing prob-lem in biology and chemistry involving the predic-t

198、ion of the three-dimensional shape a protein would assume given only the DNA sequence associated with it.Machine learning also requires large amounts of data from which it can learn.These data can take various forms,including text,images,videos,sensor readings,and more.Learning from these data is ca

199、lled training the AI model.The quality and quantity of data play a crucial role in determining the performance and capabilities of AI systems.Without sufficient and high-quality data,models may generate inaccurate or biased outcomes.(Roughly speaking,a traditional ML model is developed to solve a pa

200、rticular problem different problems call for different models;for prob-lems sufficiently different from each other,entirely new models need to be developed.Foundation models,discussed below,break this tradition to some extent.)Research continues on how to train systems incrementally,starting from ex

201、isting models and using a much smaller amount of specially curated data to refine those models performance for special-ized purposes.For a sense of scale,estimates of the data required to train GPT-4,OpenAIs large language model(LLM)released in March 2023 and the base on which some versions of ChatG

202、PT were built,suggest that its training database consisted of the textual equiv-alent of around 100 million books,or about 10 tril-lion words,drawn from billions of web pages and scanned books.(LLMs are discussed further below.)The hardware requirements for computing power are also substantial.The c

203、osts to compute the training of GPT-4,for example,were enormous.Reports indi-cate that the training took about twenty-five thousand Nvidia A100 GPU deep-learning chips at a cost of$10,000 each running for about one hundred days.15 Doing the math and noting that other 24STANFORD EMERGING TECHNOLOGY R

204、EVIEW Autonomous trucking Multiple companies col-laborated in a consortium that arranged for trucks carrying tires to drive autonomously for over fifty thousand long-haul trucking miles in the period from January to August 2024.28 If this and other demonstrations continue to be successful,it is poss

205、ible that long-haul drives the most boring and time-consuming aspect of a truck drivers job can be automated;at the same time,aspects of such jobs requiring human-centered interactions,including navigating the first miles out of the factory and the last miles of delivering goods to customers,could b

206、e retained.Law Legal transcription AI enables the real-time transcription of legal proceedings and client meetings with reasonably high accuracy,and some of these services are free of charge.29 Legal review AI-based systems can reduce the time lawyers spend on contract review by as much as 60percent

207、.Further,such systems can enable lawyers to search case databases more rapidly than online human searches and even write case summaries.30Key DevelopmentsFoundation ModelsFoundation models dominated the conversation about AI in both 2023 and 2024.These models are large-scale systems trained on vast

208、amounts of diverse data that can handle a variety of tasks.31 They often contain billions or trillions of parame-ters,32 and their massive size allows them to capture more complex patterns and relationships.Trained on these datasets,foundation models can develop broad capabilities33 and are thus som

209、etimes called general-purpose models.They excel at transfer Drug discovery An AI-enabled search iden-tified a compound that inhibits the growth of a bacterium responsible for many drug-resistant infections,such as pneumonia and meningitis,by sifting through a library of seven thousand poten-tial dru

210、g compounds for an appropriate chemical structure.22 Patient safety Smart AI sensors and cameras can improve patient safety in intensive care units,operating rooms,and even at home by improv-ing healthcare providers and caregivers ability to monitor and react to patient health develop-ments,includin

211、g falls and injuries.23 Robotic assistants Mobile robots using AI can carry out healthcare-related tasks such as making specialized deliveries,disinfecting hospital wards,and assisting physical therapists,thus supporting nurses and enabling them to spend more time having face-to-face human interacti

212、ons.24Agriculture Production optimization AI-enabled computer vision helps some salmon farmers pick out fish that are the right size to keep,thus off-loading the labor-intensive task of sorting them.25 Crop management Some farmers are using AI to detect and destroy weeds in a targeted manner,signifi

213、cantly decreasing environmental harm by using herbicides only on undesired vegetation rather than entire fields,in some cases reducing herbicide use by as much as 90percent.26Logistics and Transportation Resource allocation AI enables some commer-cial shipping companies to predict ship arrivals five

214、 days into the future with high accuracy,thus allowing real-time allocations of personnel and schedule adjustments.272501ARTIFICIAL INTELLIGENCEmedical advice,and they outscore the median human performance on clinical examination in obstet-rics and gynecology,34 on standardized tests of divergent th

215、inking,35 and on other standardized tests such as the LSAT,sections of the GRE,and various AP exams.36 However,models do not necessarily excel at the actual tasks or skills that these tests are trying to capture and,as discussed below,still pro-duce errors and fail in all sorts of other ways,many of

216、 them unexpected.Well-known closed-source LLMs include OpenAIs GPT models(e.g.,GPT-3,GPT-3.5,and GPT-4),Anthropics Claude,and Googles Gemini.Well-known open-source LLMs include Metas Llama,Big-Sciences BLOOM,EleutherAIs GPT-J,and Googles BERT and T5.Specialized foundation models have also been devel

217、oped in other modalities such as audio,video,and images:Foundation models for images are able to gen-erate new images based on a users text input.Novel methods for handling images,combined with using very large collections of pictures and text for training,have led to models that can turn written de

218、scriptions into images that are quickly becoming comparable to and sometimes indis-tinguishable from real-life photographs and artwork created by humans.Examples include OpenAIs DALL-E 3,the open-source Stable Diffusion,Googles Imagen,Adobe Firefly,and Metas Make-A-Scene.An example of a foundation m

219、odel for audio is UniAudio,which handles all audio types and employs predictive algorithms to generate high-quality speech,sound,and music,sur-passing leading methods in tasks such as text to speech,speech enhancement,and voice conversion.Foundation models in video such as Metas Emu Video represent

220、a significant advancement in learning applying knowledge learned in one con-text to another making them more flexible and effi-cient than traditional task-specific models.A single foundation model is often fine-tuned for various tasks,reducing the need to train separate models from scratch.These mod

221、els are generally classified as closed source or open source.A closed-source model is a proprietary one developed and maintained by a specific organization,usually a for-profit com-pany,with its source code,data,and architecture kept confidential.Access to these models is typi-cally restricted throu

222、gh technically enforced usage permissions,such as application programming interfaces,allowing the developers to control the models distribution,usage,and updates.By con-trast,an open-source model is one whose code,data,and underlying architecture are publicly acces-sible,allowing anyone to use,modif

223、y,and distribute it freely.The most familiar type of foundation model is an LLM a system trained on very large volumes of textual content.LLMs are an example of generative AI,a type of AI that can produce new material based on how it has been trained and the inputs it is given.Models trained on text

224、 can generate new text based on a statistical analysis that makes predictions about what other words are likely to be found immediately after the occurrence of certain words.These models do not think or feel like humans do,even though their responses may make it seem like they do.Instead,LLMs use st

225、atistical analysis based on training data.For example,because the word sequence“thank you”is far more likely to occur than“thank zebras,”a persons query to an LLM asking it to draft a thank-you note to a colleague is unlikely to generate the response“thank zebras.”These models generate linguistic ou

226、tput surpris-ingly similar to that of humans across a wide range of subjects.For example,LLMs can generate useful computer code,poetry,legal case summaries,and 26STANFORD EMERGING TECHNOLOGY REVIEWto various formats and learner types,improving engagement and comprehension.When integrated with virtua

227、l and augmented reality,it can create immersive,highly realistic training environments that are particularly valuable in fields like healthcare.The advent of multimodal AI is also set to further trans-form human-computer interactions,enabling more intuitive communication and expanding the range of t

228、asks that AI systems can handle.Embodied AIEmbodied AI involves integrating AI systems into robots or other physical devices.This approach aims to bridge the gap between the digital and physical realms.Embodied AI has the potential to enhance robotic capabilities and expand the range of inter-action

229、s robots have with the physical world.These robot-plus-AI systems could potentially address knowledge tasks,physical tasks,or combinations of both.(This topic is explored further in chapter 7 on robotics.)As research progresses in AI autonomy and reasoning,embodied AI systems may be able to handle i

230、ncreasingly complex tasks with greater inde-pendence.This could lead to applications in various fields such as logistics and domestic assistance.video generation.Emu first generates an image from text input and then creates a video based on both the text and the generated image.Emu Video has demonst

231、rated superior performance over previous state-of-the-art methods in terms of image quality,faithfulness to text instructions,and evaluations from humans.Multimodal ModelsAI systems that incorporate multiple modalities text,images,and sound within single models are becoming increasingly popular.This

232、 multimodal approach,shown in figure1.1,aims to create more humanlike experiences by leveraging various senses such as sight,speech,and hearing to mirror how humans interact with the world.Multimodal AI systems have diverse applications across sectors.They can enhance accessibility for people with d

233、isabilities through real-time transcrip-tion,sign language translation,and detailed image descriptions.They can also eliminate language barriers via cost-effective,near-real-time transla-tion services.In education,multimodal AI can sup-port personalized learning by adapting content Single-modal AI m

234、odelProcessInputOutputAITextTextMultimodal AI modelProcessInputOutputTextAIImageSoundTextImagesSoundsFIGURE 1.1Multimodal AI systems can transform one type of input into a different type of output2701ARTIFICIAL INTELLIGENCEA National AI Research ResourceLLMs such as GPT-4,Claude,Gemini,and Llama can

235、 be developed only by large companies with the resources to build and operate very large data and compute centers.For a sense of scale,Princeton University announced in March 2024 that it would dip into its endowment to purchase 300 advanced Nvidia chips to use for research at a total estimated cost

236、 of about$9 million.40 By contrast,Meta announced at the start of 2024 that it intended to purchase 350,000 such chips by the end of the year41 over one thousand times as many chips as Princeton and with a likely price tag of nearly$10 billion.Traditionally,academics and others in civil society have

237、 undertaken research to understand the poten-tial societal ramifications of AI,but with large com-panies controlling access to these AI systems,they can no longer do so independently.In July 2023,a bipartisan bill(S.2714,the CREATE AI Act of 2023)42 was proposed to establish the National Artificial

238、Intelligence Research Resource(NAIRR)as a shared national research infrastructure that would provide civil society researchers greater access to the com-plex resources,data,and tools needed to support research on safe and trustworthy AI.The bills text did not mention funding levels,but the final NAI

239、RR task force report,released in January 2023,indicated that NAIRR should be funded at a level of$2.6 billion over its initial six-year span.43 In January 2024,the National Science Foundation established the NAIRR pilot to establish proof of concept for the full-scale NAIRR.Existential Concerns Abou

240、t AILLMs have generated considerable attention because of their apparent sophistication.Indeed,their capa-bilities have led some to suggest that they are the initial sparks of artificial general intelligence(AGI).37 AGI is AI that is capable of performing any intellec-tual task that a human can perf

241、orm,including learn-ing.But,according to this argument,because an electronic AGI would run on electronic circuits rather than biological ones,it is likely to learn much faster than biological human intelligences rapidly out-stripping their capabilities.The belief in some quarters that AGI will soon

242、be achieved has led to substantial debate about its risks.Scholars have continued to argue over the past year about whether current models present initial sparks of AGI,38 although there hasnt been substan-tial evidence presented that proves they possess such capabilities.Others suggest that focusin

243、g on low-probability doomsday scenarios distracts from the real and imme-diate risks AI poses today.39 Instead,society should be prioritizing efforts to address the harms that AI systems are already causing,like biased decision-making,hallucinations(error-ridden responses that appear to provide accu

244、rate information),and job displacement.Those who support this view argue that these problems are the ones on which govern-ments and regulators should be concentrating their efforts.The advent of multimodal AI is.set to further transform human-computer interactions,enabling more intuitive communicati

245、on and expanding the range of tasks that AI systems can handle.28STANFORD EMERGING TECHNOLOGY REVIEWalways relevant,but in certain cases,such as medical decision-making,they may be critical so that users can have confidence in an AI systems output.Bias and fairness Because ML models are trained on e

246、xisting datasets,they are likely to encode any biases present in these datasets.(Bias should be understood here as a property of the data that is commonly regarded as societally undesirable.)For example,if a facial recognition system is primarily trained on images of individuals from one ethnic grou

247、p,its accuracy at identifying people from other ethnic groups may be reduced.45 Use of such a system could well lead to disproportionate singling out of individuals in those other groups.To the extent that these datasets reflect historical approaches,they will also reflect the biases embedded in tha

248、t history,and an ML model based on such datasets will also reflect these biases.Vulnerability to spoofing It is possible to tweak data inputs to fool many AI models into drawing false conclusions.For example,in figure1.2,chang-ing a small number of pixels in a visual image of a traffic stop sign can

249、 lead to its being classified as As a point of comparison to the fledgling NAIRR effort,investments from high-tech companies for AI exceeded$27 billion in 2023 alone.44Over the HorizonImpact of New AI TechnologiesPotential positive impacts of new AI technologies are most likely to be seen in the app

250、lications they enable for societal use,as described in detail above.On the other hand,no technology is an unalloyed good.Potential negative impacts from AI will likely emerge from known problems with current state-of-the-art AI and from technical advances in the future.Some of the known issues with

251、todays leading AI models include the following:Explainability This is the ability to explain the rea-soning behind and describe the data underlying an AI systems conclusions.Todays AI is largely incapable of explaining the basis on which it arrives at any particular conclusion.Explanations are not C

252、hange a fewpixelsNew AI classifcation:“yield”Original AI classifcation:”stop”Source:Derived from figure 1 in Fabio Carrara,Fabrizio Falchi,Giuseppe Amato,Rudy Becarelli,and Roberto Caldelli,“Detecting Adversarial Inputs by Looking in the Black Box,”in“Transparency in Algorithmic Decision Making,”spe

253、cial issue,ERCIM News 116(January 2019):1619.FIGURE 1.2Changing a few pixels can fool AI into thinking a picture of a stop sign is a picture of a yield sign2901ARTIFICIAL INTELLIGENCEelection by buying votes.47 In January 2024,voters in New Hampshire received robocalls that used a voice sounding lik

254、e President Bidens telling them not to vote in the states presidential primary.48 In elections in India in early 2024,deepfake videos were used to depict deceased politicians as though they were still alive(see figure1.3).49 All of these deepfakes are much more sophisticated than attempts such as th

255、e“dumbfake”video of Representative Nancy Pelosi(D-CA)that involved merely slowing down an exist-ing video of her to make her look drunk.50Privacy Many LLMs are trained on data found on the internet rather indiscriminately,and such data may include personal information of individu-als.When incorporat

256、ed into LLMs,this information could be publicly disclosed more often.Overtrust and overreliance If AI systems become commonplace in society,their novelty will inevitably diminish for users.The level of trust in computer out-puts often increases with familiarity.But skepticism about answers received

257、from a system is essential if one is to challenge the correctness of these outputs.As trust in AI grows,reducing skepticism,theres a a yield sign,even though this fuzzing of the image is invisible to the naked eye.That example seems innocuous,but as AI models are used increasingly in applications fr

258、om medical treatment to intelli-gence and military operations,the potential harms could be substantial.It is also possible that an attack targeting one AI model could work against other models performing the same task a phenomenon known as transferability.One study reports that as often as 80 percen

259、t of the time,transferability allows attackers to create an attack on a surrogate model and then apply it to their intended target,too.46 Data poisoning An attacker manipulating the dataset used to train an ML model can damage its performance and even create predictable errors.Deepfakes AI provides

260、the capability for gener-ating highly realistic but entirely inauthentic audio and video imagery.This has obvious implications for evidence presented in courtrooms and for efforts to manipulate political contests.In September 2023,just before elections took place in Slovakia,a deepfake audio was pos

261、ted to Facebook in which a candidate was heard discussing with a journalist how to rig the Photo from the late Indian Congress leaderH.Vasanthakumars funeral in 2020Screenshot from a deepfake video ofH.Vasanthakumar endorsing his sonsparliamentary candidacy in 2024Source:(Left)PTI Photo/R.Senthil Ku

262、mar;(right)“H Vasantha Kumar,”posted April 16,2024,by Vasanth TV,YouTube,https:/ 1.3Deepfake videos of deceased Indian politicians speaking as if they were alive were used in Indias 2024 elections30STANFORD EMERGING TECHNOLOGY REVIEWand metadata in building and offering the prod-ucts Stable Diffusio

263、n(an application that generates images from text)and DreamStudio(the app that serves as a user interface to Stable Diffusion).52 In late 2023,the New York Times sued OpenAI and Microsoft over their alleged use of millions of arti-cles published by the Times to train the companies LLMs.53 In June 202

264、4,music labels Sony Music,Universal Music Group,and Warner Records sued AI start-ups Suno and Udio for copyright infringe-ment,alleging that the companies had trained their music-generation systems on protected content.54AI researchers are cognizant of issues such as these,and in many cases work has

265、 been done or is being done to develop corrective measures.However,in most cases,these defenses dont apply very well to instances beyond the specific problems that they were designed to solve.Challenges of Innovation and Implementation The primary challenge of bringing AI innovation into operation i

266、s risk management.It is often said that AI,and especially ML,brings a new conceptual par-adigm for how systems can exploit information to gain advantage,relying on pattern recognition in the broadest sense rather than on explicit understand-ing of situations that are likely to occur.Because there ha

267、ve been significant recent advances in AI,the people who would make decisions to deploy AI-based systems may not have a good under-standing of the risks that could accompany such deployment.Consider,for example,AI as an important approach for improving the effectiveness of military oper-ations.Despi

268、te broad agreement by the military services and the US Department of Defense(DOD)that AI would be of great benefit,the actual inte-gration of AI-enabled capabilities into military forces has proceeded at a slow pace.Certainly,it is well understood that technical risks of underperformance and error i

269、n new technologies take time to mitigate.higher risk that errors,mishaps,and unforeseen inci-dents will be overlooked.One recent experiment showed that developers with access to an AI-based coding assistant wrote code that was significantly less secure than those without an AI-based assistant even t

270、hough the former were more likely to believe they had written secure code.51Hallucinations As noted earlier,AI hallucina-tions refer to situations where an AI model gen-erates results or answers that are plausible but do not correspond to reality.In other words,models can simply make things up,but h

271、uman users will not be aware they have done this.The results are plausible because they are constructed based on statistical patterns that the model has learned to recognize from its training data.But they may not correspond to reality because the model does not have an understanding of the real wor

272、ld.For exam-ple,in September 2024,a Stanford professor asked an AI model to name ten publications she had writ-ten.The AI responded with five correct publications and five that she had never actually written but the AI results included titles and summaries that made them seem real.When she told the

273、model that“the last two entries are hallucinations,”it simply pro-vided two new results that were also hallucinations.Out-of-distribution inputs All ML systems must be trained on a large volume of data.If the inputs subsequently given to a system are substantially different from the training data a

274、situation known as being out-of-distribution the system may draw conclusions that are more unreliable than if the inputs were similar to the training data.Copyright violations Some AI-based models have been trained on large volumes of data found online.These data have generally been used without the

275、 consent or permission of their owners,thereby raising important questions about appropriately compen-sating and acknowledging those owners.For exam-ple,in January 2023,Getty Images sued Stability AI in an English court for infringing on the copyrights of millions of photographs,their associated cap

276、tions,3101ARTIFICIAL INTELLIGENCEthe humans it replaces.58 In at least some cases,companies are deciding that the cost savings of eliminating human workers outweigh the draw-backs of mediocre AI performance.Training displaced workers to be more compet-itive in an AI-enabled economy does not solve th

277、e problem if new jobs are not available.The nature and extent of new roles resulting from widespread AI deployment are not clear at this point,although historically the introduction of new technologies has not resulted in a long-term net loss of jobs.59GOVERNANCE AND REGULATION OF AIGovernments arou

278、nd the world have been increas-ingly focused on establishing regulations and guide-lines for AI.Research on foundational AI technologies is difficult to regulate across international bound-aries even among like-minded nations,especially when other nations have strong incentives to carry on regardles

279、s of actions taken by US policymakers.It is even more difficult,and may well be impossible,to reach agreement between nations that regard each other as strategic competitors and adversaries.The same applies to voluntary restrictions on research by companies concerned about competition from less cons

280、trained foreign rivals.Regulation of specific applications of AI may be more easily implemented,in part because of existing regulatory frameworks in domains such as healthcare,finance,and law.The most ambitious attempt to regulate AI came into force in August 2024 with the European Unions AI Act.Thi

281、s forbids certain applications of AI,such as individual predictive policing based solely on a persons data profile or tracking of their emotional state in the workplace and educational institutions,unless for medical or safety reasons.60 Additionally,it imposes a number of requirements on what the A

282、I Act calls“high-risk”systems.(The legislation pro-vides a very technical definition of such systems,but generally they include those that could pose a sig-nificant risk to health,safety,or fundamental rights.)But another important reason for the slow pace is that the DOD acquisition system has larg

283、ely been designed to minimize the likelihood of program-matic failure,fraud,unfairness,waste,and abuse in short,to minimize risk.It is therefore not surprising that the incentives at every level of the bureaucracy are aligned in that manner.For new approaches like AI to take root,a greater degree of

284、 programmatic risk acceptance may be necessary,especially in light of the possibility that other nations could adopt the technology faster,achieving military advantages over US forces.Policy,Legal,and Regulatory IssuesTHE FUTURE OF WORKSome researchers expect that,within the next five to ten years,m

285、ore and more workers will have AI added to their workflows to enhance productivity or will even be replaced by AI systems,which may cause significant disruptions to the job market in the near future.55 LLMs have already demonstrated how they can be used in a wide variety of fields,includ-ing law,cus

286、tomer support,coding,and journalism.These demonstrations have led to concerns that the impact of AI on employment will be substan-tial,especially on jobs that involve knowledge work.However,uncertainty abounds.What and how many present-day jobs will disappear?Which tasks could best be handled by AI?

287、And what new jobs might be created by the technology today and in the future?Some broad outlines and trends are clear:Individuals whose jobs entail routine white-collar work may be more affected than those whose jobs require physical labor;some will experience painful shifts in the short term.56 AI

288、is helping some workers to increase their pro-ductivity and job satisfaction.57 At the same time,other workers are already losing their jobs as AI demonstrates adequate competence for business operations,despite potentially underperforming 32STANFORD EMERGING TECHNOLOGY REVIEWin September 2024.The a

289、ct sought to hold the cre-ators of advanced AI models liable in civil court for causing catastrophic harms unless they had taken certain advance measures to forestall such an out-come.Opposition to the bill was based on con-cerns about a technologically deficient definition of advanced AI models,the

290、 burden that the bill would place on small start-ups and academia,and the unfairness of holding model developers responsible for harmful applications that others build using the developers models.Other important developments regarding AI gov-ernance include the AI Safety Summit,held on November 12,2

291、023,at Bletchley Park in the United Kingdom,62 which issued the Bletchley Declaration,and the AI Seoul Summit of May 2024.In the Bletchley Declaration,the European Union and twenty-eight nations collectively endorsed international coopera-tion to manage risks associated with highly capable general-p

292、urpose AI models.Signatories commit-ted to ensuring that AI systems are developed and deployed safely and responsibly.The summit also led to the establishment of the United Kingdoms AI Safety Institute and the US Artificial Intelligence Safety Institute,located within the National Institute of Stand

293、ards and Technology.The Seoul Declaration from the AI Seoul Summit 2024 built on the Bletchley Declaration to acknowl-edge the importance of interoperability between national AI governance frameworks to maximize benefits and minimize risks from advanced AI sys-tems.In addition,sixteen major AI organ

294、izations These requirements address data quality,documen-tation and traceability,transparency and explainabil-ity,human oversight,accuracy,cybersecurity,and robustness.In the United States,the presidents Executive Order on the Safe,Secure,and Trustworthy Development and Use of Artificial Intelligenc

295、e was issued on October 30,2023.61 The order addressed actions to advance AI safety and security;privacy;equity and civil rights;consumer,patient,student,and worker interests;the promotion of innovation and com-petition,as well as American leadership;and gov-ernment use of AI.Of particular note is t

296、he orders requirement that developers of advanced AI systems posing a serious risk to national security,national eco-nomic security,or national public health and safety inform the US government when training them and share with it all results from internal safety testing conducted by red teams.(A re

297、d team is a team of experts that attempts to subvert or break the system it is asked to test.It then reports its findings to the owner of the system so that the owner can take cor-rective action.)The order also requires government actions to develop guidance to help protect against the use of AI to

298、develop biological threats and to advance the use of AI to protect against cyberse-curity threats,to help detect AI-generated content,and to authenticate official content.At the state level in the United States,an attempt to pass an AI regulatory bill in California(SB 1047,the Safe and Secure Innova

299、tion for Frontier Artificial Intelligence Models Act)was vetoed by the governor The resources needed to train GPT-4 far exceed those available through grants or any other sources to any reasonably sized group of the top US research universities.3301ARTIFICIAL INTELLIGENCEof weapons or in making deci

300、sions about the use of deadly force.Notably,in late 2023,press reports indicated that President Biden and Chinese President Xi Jinping had considered entering into a dialogue about AI in nuclear command and control,but such an arrangement was never formalized.66 TALENTThe United States is eating its

301、 seed corn with respect to the AI talent pool.As noted in the Foreword,fac-ulty at Stanford and other universities report that the number of students studying in AI who are joining the industry,particularly start-ups,is increasing at the expense of those pursuing academic careers and contributing to

302、 foundational AI research.Many factors are contributing to this trend.One is that industry careers come with compensation packages that far outstrip those offered by aca-demia.Academic researchers must also obtain funding to pay for research equipment,computing capability,and personnel like staff sc

303、ientists,tech-nicians,and programmers.This involves searching for government grants,which are typically small compared to what large companies might be will-ing to invest in their own researchers.Consider,for example,that the resources needed to build and train GPT-4 far exceed those available throu

304、gh grants or any other sources to any reasonably sized group of the top US research universities,let alone any single university.Industry often makes decisions more rapidly than government grant makers and imposes fewer reg-ulations on the conduct of research.Large com-panies are at an advantage bec

305、ause they have research-supporting infrastructure in place,such as compute facilities and data warehouses.One important consequence is that academic access to research infrastructure is limited,so US-based stu-dents are unable to train on state-of-the-art systems at least this is the case if their u

306、niversities do not have access to the facilities of the corporate sector.agreed on the Frontier AI Safety Commitments,a set of voluntary guidelines regarding the publication of safety frameworks for frontier AI models and the setting of thresholds for intolerable risks,among other things.NATIONAL SE

307、CURITYAI is expected to have a profound impact on mil-itaries worldwide.63 Weapons systems,command and control,logistics,acquisition,and training will all seek to leverage multiple AI technologies to operate more effectively and efficiently,at lower cost and with less risk to friendly forces.Trying

308、to overcome decades of institutional inertia,the DOD is dedicating billions of dollars to institutional reforms and research advances aimed at integrat-ing AI into its war fighting and war preparation strategies.Senior military officials recognize that failure to adapt to the emerging opportunities

309、and challenges presented by AI would pose significant national security risks,particularly considering that both Russia and China are heavily investing in AI capabilities.In adopting a set of guiding principles that address responsibility,equity,traceability,reliability,and governability in and for

310、AI,64 the DOD has taken an important first step in meeting its obligation to proceed ethically with the development of AI capa-bilities;eventually,these principles will have to be operationalized in specific use cases.An additional important concern,subsumed under these principles but worth calling

311、out,is determining where the use of AI may or may not be appropriate for example,whether AI is appropriate in nuclear command and control.The United States,the United Kingdom,and France have made explicit commitments to maintain human control over nuclear weapons.65 Meanwhile,other countries are als

312、o adopting AI,and nations such as Russia and China are unlikely to make the same operational and ethical decisions as Western countries about the appropriate roles of AI vis-vis humans in controlling the operation 34STANFORD EMERGING TECHNOLOGY REVIEWFigure1.4 shows that most notable ML systems are

313、now released by industry,while very few are released by academic institutions.At the same time,Chinas efforts to recruit top sci-entific talent offer further temptations for scientists to leave the United States.These efforts are often targeted toward ethnic Chinese in the US ranging from well-estab

314、lished researchers to those just fin-ishing graduate degrees and offer recruitment packages that promise benefits comparable to those available from private industry,such as high salaries,lavish research funding,and apparent freedom from bureaucracy.All of these factors are leading to an AI“brain dr

315、ain”that does not favor the US research enterprise.20102011201220132014201520162017200320042005200620072008200920182019202020212022202350403020100Number of notable machine-learning models by sector,200323Number of notablemachine-learning modelsIndustryIndustry-academia collaborationAcademia512115Sou

316、rce:Adapted from Nestor Maslej,Loredana Fattorini,Raymond Perrault,et al.,The AI Index 2024 Annual Report,AI Index Steering Committee,Institute for Human-Centered AI,Stanford University,Stanford,CA,April 2024.Data from Epoch,2023FIGURE 1.4Most notable machine-learning models are now released by indu

317、stryNOTES1.Christopher Manning,“Artificial Intelligence Definitions,”Insti-tute for Human-Centered AI,Stanford University,September 2020,https:/hai.stanford.edu/sites/default/files/2020-09/AI-Definitions-HAI.pdf.2.Grand View Research,“Artificial Intelligence Market Size,Share,Growth Report 20242030,

318、”accessed September 23,2024,https:/ Maslej,Loredana Fattorini,Raymond Perrault,et al.,The AI Index 2024 Annual Report,AI Index Steering Committee,Institute for Human-Centered AI,Stanford University,Stanford,CA,April 2024,https:/aiindex.stanford.edu/wp-content/uploads/2024/05/HAI_AI-Index-Report-2024

319、.pdf.4.CB Insights,“State of Venture 2023 Report,”February 1,2024,https:/ sources disagree as to whether global venture fund-ing for AI start-ups increased or decreased in 2023.We have fol-lowed the majority view.6.CB Insights,“State of AI 2023 Report,”February 1,2024,https:/ Weise,“In Race to Build

320、 A.I.,Tech Plans a Big Plumbing Upgrade,”New York Times,April 27,2024,https:/ INTELLIGENCE8.Jack Pitcher and Connor Hart,“BlackRock,Microsoft,Others Form AI and Energy Infrastructure Investment Partnership,”Wall Street Journal,September 17,2024,https:/ Infrastructure Partners,Micro-soft,and MGX Laun

321、ch New AI Partnership to Invest in Data Cen-ters and Supporting Power Infrastructure,”September 17,2024,https:/ Sachs,“Generative AI Could Raise Global GDP by 7%,”April 5,2023,https:/ et al.,The AI Index 2024 Report.12.Jafar Alzubi,Anand Nayyar,and Akshi Kumar,“Machine Learning from Theory to Algori

322、thms:An Overview,”Journal of Physics:Conference Series 1142,Second National Conference on Computational Intelligence(December 2018),https:/doi.org/10.1088/1742-6596/1142/1/012012.13.The Nobel Prize in Physics 2024,“Summary,”The Nobel Prize,October 12,2024,https:/www.nobelprize.org/prizes/physics/202

323、4/summary/.14.The Nobel Prize in Chemistry 2024,“Summary,”The Nobel Prize,October 12,2024,https:/www.nobelprize.org/prizes/chemistry/2024/summary/.15.Kif Leswing,“Meet the$10,000 Nvidia Chip Powering the Race for A.I.,”CNBC,February 23,2023,https:/ Groes Albin Ludvigsen,“The Carbon Footprint of GPT-

324、4,”Medium,July 18,2023,https:/towardsdata Woods and Adrian Ma,“The Semiconductor Founding Father,”December 21,2021,in The Indicator from Planet Money,podcast produced by NPR,MP3 audio,10:14,https:/www.npr.org/transcripts/1066548023.17.Ludvigsen,“The Carbon Footprint.”18.Kasper Groes Albin Ludvigsen,

325、“ChatGPTs Electricity Con-sumption,”Medium,July 12,2023,https:/ sources provide somewhat different numbers for the energy cost per query,but they all are in the range of a few watt-hours.19.EPRI,“Powering Intelligence:Analyzing Artificial Intelligence and Data Center Energy Consumption,”Technology I

326、nnovation,accessed September 23,2023,https:/ Reese,“A Human-Centered Approach to the AI Revolution,”Institute for Human-Centered AI,Stanford Uni-versity,October 17,2022,https:/hai.stanford.edu/news/human-centered-approach-ai-revolution.21.Viz.ai,“Viz.ai Receives New Technology Add-on Payment(NTAP)Re

327、newal for Stroke AI Software from CMS,”August 4,2021,https:/www.viz.ai/news/ntap-renewal-for-stroke-software.22.Gary Liu,Denise B.Catacutan,Khushi Rathod,et al.,“Deep Learning-Guided Discovery of an Antibiotic Targeting Acineto-bacter baumannii,”Nature Chemical Biology(2023),https:/doi.org/10.1038/s

328、41589-023-01349-8.23.Albert Haque,Arnold Milstein,and Fei-Fei Li,“Illuminating the Dark Spaces of Healthcare with Ambient Intelligence,”Nature 585(2020):193202,https:/doi.org/10.1038/s41586-020-2669-y.24.Khari Johnson,“Hospital Robots Are Helping Combat a Wave of Nurse Burnout,”Wired,April 19,2022,h

329、ttps:/ Site,“Innovasea Launches AI-Powered Biomass Camera for Salmon,”August 17,2023,https:/ Learning in Agriculture:Use Cases and Applications,”February 1,2023,https:/ Maritime,“Vessel Tracking Giving Full Journey Visibility,”accessed August 15,2023,https:/ and Kodiak Surpass 50,000 Autonomous Long

330、-Haul Trucking Miles in Delivery Collabo-ration,”August 7,2024,https:/kodiak.ai/news/jb-hunt-and-kodiak-collaborate.29.JD Supra,“Artificial Intelligence in Law:How AI Can Reshape the Legal Industry,”September 12,2023,https:/ Lohr,“A.I.Is Doing Legal Work.But It Wont Replace Law-yers,Yet,”New York Ti

331、mes,March 19,2017,https:/ Bommasani,Drew A.Hudson,Ehsan Adeli,et al.,“On the Opportunities and Risks of Foundation Models,”arXiv,Stan-ford University,July 12,2022,https:/doi.org/10.48550/arXiv.2108.07258.32.Parameters are like the building blocks of knowledge that make up an AI systems understanding

332、.You can think of them as tiny bits of information the AI uses to make sense of data and gen-erate responses.When we say AI models have billions or trillions of parameters,it means they have an enormous number of these information pieces to work with.This allows them to understand and generate more

333、sophisticated and nuanced content.33.Bommasani et al.,“On the Opportunities and Risks.”34.Sarah W.Li,Matthew W.Kemp,Susan J.S.Logan,Sebastian E.Illanes,and Mahesh A.Choolani,“ChatGPT Outscored Human Candidates in a Virtual Objective Structured Clinical Examination in Obstetrics and Gynecology,”American Journal of Obstetrics&Gynecology 229,no.2(August 2023):172.E1-172.E12,https:/doi.org/10.1016/j.a