《斯坦福大学：2025年人工智能指数（AI Index）报告（英文版）（456页）.pdf》由会员分享，可在线阅读，更多相关《斯坦福大学：2025年人工智能指数（AI Index）报告（英文版）（456页）.pdf（456页珍藏版）》请在三个皮匠报告上搜索。

1、Artificial IntelligenceIndex Report 2025Artificial IntelligenceIndex Report 20251Welcome to the eighth edition of the AI Index report.The 2025 Index is our most comprehensive to date and arrives at an important moment,as AIs influence across society,the economy,and global governance continues to int

2、ensify.New in this years report are in-depth analyses of the evolving landscape of AI hardware,novel estimates of inference costs,and new analyses of AI publication and patenting trends.We also introduce fresh data on corporate adoption of responsible AI practices,along with expanded coverage of AIs

3、 growing role in science and medicine.Since its founding in 2017 as an offshoot of the One Hundred Year Study of Artificial Intelligence,the AI Index has been committed to equipping policymakers,journalists,executives,researchers,and the public with accurate,rigorously validated,and globally sourced

4、 data.Our mission has always been to help these stakeholders make better-informed decisions about the development and deployment of AI.In a world where AI is discussed everywherefrom boardrooms to kitchen tablesthis mission has never been more essential.The AI Index continues to lead in tracking and

5、 interpreting the most critical trends shaping the fieldfrom the shifting geopolitical landscape and the rapid evolution of underlying technologies,to AIs expanding role in business,policymaking,and public life.Longitudinal tracking remains at the heart of our mission.In a domain advancing at breakn

6、eck speed,the Index provides essential contexthelping us understand where AI stands today,how it got here,and where it may be headed next.Recognized globally as one of the most authoritative resources on artificial intelligence,the AI Index has been cited in major media outlets such as The New York

7、Times,Bloomberg,and The Guardian;referenced in hundreds of academic papers;and used by policymakers and government agencies around the world.We have briefed companies like Accenture,IBM,Wells Fargo,and Fidelity on the state of AI,and we continue to serve as an independent source of insights for the

8、global AI ecosystem.Introduction to the AI Index Report 2025Artificial IntelligenceIndex Report 20252As AI continues to reshape our lives,the corporate world,and public discourse,the AI Index continues to track its progressoffering an independent,data-driven perspective on AIs development,adoption,a

9、nd impact,across time and geography.What a year 2024 has been for AI.The recognition of AIs role in advancing humanitys knowledge is reflected in Nobel prizes in physics and chemistry,and the Turing award for foundational work in reinforcement learning.The once-formidable Turing Test is no longer co

10、nsidered an ambitious goal,having been surpassed by todays sophisticated systems.Meanwhile,AI adoption has accelerated at an unprecedented rate,as millions of people are now using AI on a regular basis both for their professional work and leisure activities.As high-performing,low-cost,and openly ava

11、ilable models proliferate,AIs accessibility and impact are set to expand even further.After a brief slowdown,corporate investment in AI rebounded.The number of newly funded generative AI startups nearly tripled,and after years of sluggish uptake,business adoption accelerated significantly in 2024.AI

12、 has moved from the margins to become a central driver of business value.Governments,too,are ramping up their involvement.Policymakers are no longer just debating AItheyre investing in it.Several countries launched billion-dollar national AI infrastructure initiatives,including major efforts to expa

13、nd energy capacity to support AI development.Global coordination is increasing,even as local initiatives take shape.Yet trust remains a major challenge.Fewer people believe AI companies will safeguard their data,and concerns about fairness and bias persist.Misinformation continues to pose risks,part

14、icularly in elections and the proliferation of deepfakes.In response,governments are advancing new regulatory frameworks aimed at promoting transparency,accountability,and fairness.Public attitudes are also shifting.While skepticism remains,a global survey in 2024 showed a notable rise in optimism a

15、bout AIs potential to deliver broad societal benefits.AI is no longer just a story of whats possibleits a story of whats happening now and how we are collectively shaping the future of humanity.Explore this years AI Index report and see for yourself.Yolanda Gil and Raymond PerraultCo-directors,AI In

16、dex ReportMessage From the Co-directorsArtificial IntelligenceIndex Report 20253Top Takeaways1.AI performance on demanding benchmarks continues to improve.In 2023,researchers introduced new benchmarksMMMU,GPQA,and SWE-benchto test the limits of advanced AI systems.Just a year later,performance sharp

17、ly increased:scores rose by 18.8,48.9,and 67.3 percentage points on MMMU,GPQA,and SWE-bench,respectively.Beyond benchmarks,AI systems made major strides in generating high-quality video,and in some settings,language model agents even outperformed humans in programming tasks with limited time budgets

18、.2.AI is increasingly embedded in everyday life.From healthcare to transportation,AI is rapidly moving from the lab to daily life.In 2023,the FDA approved 223 AI-enabled medical devices,up from just six in 2015.On the roads,self-driving cars are no longer experimental:Waymo,one of the largest U.S.op

19、erators,provides over 150,000 autonomous rides each week,while Baidus affordable Apollo Go robotaxi fleet now serves numerous cities across China.3.Business is all in on AI,fueling record investment and usage,as research continues to show strong productivity impacts.In 2024,U.S.private AI investment

20、 grew to$109.1 billionnearly 12 times Chinas$9.3 billion and 24 times the U.K.s$4.5 billion.Generative AI saw particularly strong momentum,attracting$33.9 billion globally in private investmentan 18.7%increase from 2023.AI business usage is also accelerating:78%of organizations reported using AI in

21、2024,up from 55%the year before.Meanwhile,a growing body of research confirms that AI boosts productivity and,in most cases,helps narrow skill gaps across the workforce.4.The U.S.still leads in producing top AI modelsbut China is closing the performance gap.In 2024,U.S.-based institutions produced 4

22、0 notable AI models,compared to Chinas 15 and Europes three.While the U.S.maintains its lead in quantity,Chinese models have rapidly closed the quality gap:performance differences on major benchmarks such as MMLU and HumanEval shrank from double digits in 2023 to near parity in 2024.China continues

23、to lead in AI publications and patents.Model development is increasingly global,with notable launches from the Middle East,Latin America,and Southeast Asia.5.The responsible AI ecosystem evolvesunevenly.AI-related incidents are rising sharply,yet standardized RAI evaluations remain rare among major

24、industrial model developers.However,new benchmarks like HELM Safety,AIR-Bench,and FACTS offer promising tools for assessing factuality and safety.Among companies,a gap persists between recognizing RAI risks and taking meaningful action.In contrast,governments are showing increased urgency:In 2024,gl

25、obal cooperation on AI governance intensified,with organizations including the OECD,EU,U.N.,and African Union releasing frameworks focused on transparency,trustworthiness,and other core responsible AI principles.Artificial IntelligenceIndex Report 20254Top Takeaways(contd)6.Global AI optimism is ris

26、ingbut deep regional divides remain.In countries like China(83%),Indonesia(80%),and Thailand(77%),strong majorities see AI products and services as more beneficial than harmful.In contrast,optimism remains far lower in places like Canada(40%),the United States(39%),and the Netherlands(36%).Still,sen

27、timent is shifting:Since 2022,optimism has grown significantly in several previously skeptical countries,including Germany(+10%),France(+10%),Canada(+8%),Great Britain(+8%),and the United States(+4%).7.AI becomes more efficient,affordable,and accessible.Driven by increasingly capable small models,th

28、e inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024.At the hardware level,costs have declined by 30%annually,while energy efficiency has improved by 40%each year.Open-weight models are closing the gap with closed models,reduci

29、ng the performance difference from 8%to just 1.7%on some benchmarks in a single year.Together,these trends are rapidly lowering the barriers to advanced AI.8.Governments are stepping up on AIwith regulation and investment.In 2024,U.S.federal agencies introduced 59 AI-related regulationsmore than dou

30、ble the number in 2023and issued by twice as many agencies.Globally,legislative mentions of AI rose 21.3%across 75 countries since 2023,marking a ninefold increase since 2016.Alongside growing attention,governments are investing at scale:Canada pledged$2.4 billion,China launched a$47.5 billion semic

31、onductor fund,France committed 109 billion,India pledged$1.25 billion,and Saudi Arabias Project Transcendence represents a$100 billion initiative.9.AI and computer science education is expandingbut gaps in access and readiness persist.Two-thirds of countries now offer or plan to offer K12 CS educati

32、ontwice as many as in 2019with Africa and Latin America making the most progress.In the U.S.,the number of graduates with bachelors degrees in computing has increased 22%over the last 10 years.Yet access remains limited in many African countries due to basic infrastructure gaps like electricity.In t

33、he U.S.,81%of K12 CS teachers say AI should be part of foundational CS education,but less than half feel equipped to teach it.10.Industry is racing ahead in AIbut the frontier is tightening.Nearly 90%of notable AI models in 2024 came from industry,up from 60%in 2023,while academia remains the top so

34、urce of highly cited research.Model scale continues to grow rapidlytraining compute doubles every five months,datasets every eight,and power use annually.Yet performance gaps are shrinking:the Elo skill score difference between the top and 10th-ranked models fell from 11.9%to 5.4%in a year,and the t

35、op two are now separated by just 0.7%.The frontier is increasingly competitiveand increasingly crowded.Artificial IntelligenceIndex Report 20255Top Takeaways(contd)11.AI earns top honors for its impact on science.AIs growing importance is reflected in major scientific awards:Two Nobel Prizes recogni

36、zed work that led to deep learning(physics)and to its application to protein folding(chemistry),while the Turing Award honored groundbreaking contributions to reinforcement learning.12.Complex reasoning remains a challenge.AI models excel at tasks like International Mathematical Olympiad problems bu

37、t still struggle with complex reasoning benchmarks like PlanBench.They often fail to reliably solve logic tasks even when provably correct solutions exist,limiting their effectiveness in high-stakes settings where precision is critical.Artificial IntelligenceIndex Report 20256ChairMembersRaymond Per

38、raultSRI InternationalChair-electYolanda GilUniversity of Southern California,Information Sciences InstituteResearch Manager and Editor-in-Chief Nestor Maslej,Stanford UniversityResearch AssociateLoredana Fattorini,Stanford UniversityAffiliated ResearchersElif Kiesow Cortez,Stanford Law School Resea

39、rch FellowJulia Betts Lotufo,ResearcherAnka Reuel,Stanford UniversityAlexandra Rome,ResearcherAngelo Salatino,Knowledge Media Institute,The Open UniversityLapo Santarlasci,IMT School for Advanced Studies LuccaErik BrynjolfssonStanford UniversityJack ClarkAnthropic,OECDJohn EtchemendyStanford Univers

40、ityKatrina LigettHebrew UniversityTerah LyonsJPMorgan Chase&Co.James ManyikaGoogle,University of OxfordJuan Carlos NieblesStanford University,SalesforceSteering CommitteeStaff and ResearchersVanessa ParliStanford UniversityYoav Shoham Stanford University,AI21 LabsRussell WaldStanford University Tobi

41、 WalshUNSW SydneyGraduate ResearchersEmily Capstick,Stanford UniversityMalou van Draanen Glismann,Stanford UniversityNjenga Kariuki,Stanford UniversityUndergraduate ResearchersArmin Hamrah,Claremont McKenna CollegeSukrut Oak,Stanford UniversityNgorli Fiifi Paintsil,Stanford UniversityAndrew Shi,Stan

42、ford UniversityArtificial IntelligenceIndex Report 20257The AI Index was conceived within the One Hundred Year Study on Artificial Intelligence(AI100).The AI Index welcomes feedback and new ideas for next year.Contact us at nmaslejstanford.edu.The AI Index acknowledges that while authored by a team

43、of human researchers,its writing process was aided by AI tools.Specifically,the authors used ChatGPT and Claude to help tighten and copy edit initial drafts.The workflow involved authors writing the original copy and utilizing AI tools as part of the editing process.Nestor Maslej,Loredana Fattorini,

44、Raymond Perrault,Yolanda Gil,Vanessa Parli,Njenga Kariuki,Emily Capstick,Anka Reuel,Erik Brynjolfsson,John Etchemendy,Katrina Ligett,Terah Lyons,James Manyika,Juan Carlos Niebles,Yoav Shoham,Russell Wald,Tobi Walsh,Armin Hamrah,Lapo Santarlasci,Julia Betts Lotufo,Alexandra Rome,Andrew Shi,Sukrut Oak

45、.“The AI Index 2025 Annual Report,”AI Index Steering Committee,Institute for Human-Centered AI,Stanford University,Stanford,CA,April 2025.The AI Index 2025 Annual Report by Stanford University is licensed under Attribution-NoDerivatives 4.0 International.The AI Index 2025 Report is supplemented by r

46、aw data and an interactive tool.We invite each reader to use the data and the tool in a way most relevant to their work and interests.Raw data and charts:The public data and high-resolution images of all the charts in the report are available on Google Drive.Global AI Vibrancy Tool:Compare the AI ec

47、osystems of over 30 countries.The Global AI Vibrancy tool will be updated in the summer of 2025.The AI Index is an independent initiative at the Stanford Institute for Human-Centered Artificial Intelligence(HAI).How to Cite This ReportPublic Data and ToolsAI Index and Stanford HAIArtificial Intellig

48、enceIndex Report 20258Supporting PartnersAnalytics and Research PartnersArtificial IntelligenceIndex Report 20259IntroductionLoredana Fattorini,Yolanda Gil,Nestor Maslej,Vanessa Parli,Ray PerraultChapter 1:Research and DevelopmentNancy Amato,Andrea Brown,Ben Cottier,Luca Ronchi Darr,Virginia Dignum,

49、Meredith Ellison,Robin Evans,Loredana Fattorini,Yolanda Gil,Armin Hamrah,Katrina Ligett,Nestor Maslej,Maurice Pagnucco,Ngorli Fiifi Paintsil,Vanessa Parli,Ray Perrault,Robi Rahman,Christine Raval,Vesna Sabljakovic-Fritz,Angelo Salatino,Lapo Santarlasci,Andrew Shi,Nathan Sturtevant,Daniel Weld,Kevin

50、Xu,Meg YoungChapter 2:Technical PerformanceRishi Bommasani,Erik Brynjolfsson,Loredana Fattorini,Tobi Gertsenberg,Yolanda Gil,Noah Goodman,Nicholas Haber,Armin Hamrah,Sanmi Koyejo,Percy Liang,Katrina Ligett,Nestor Maslej,Juan Carlos Niebles,Sukrut Oak,Vanessa Parli,Marco Pavone,Ray Perrault,Anka Reue

51、l,Andrew Shi,Yoav Shoham,Toby WalshChapter 3:Responsible AIMedha Bankhwal,Emily Capstick,Dmytro Chumachenko,Patrick Connolly,Natalia Dorogi,Loredana Fattorini,Ann Fitz-Gerald,Yolanda Gil,Armin Hamrah,Ariel Lee,Katrina Ligett,Shayne Longpre,Nestor Maslej,Katherine Ottenbreit,Halyna Padalko,Vanessa Pa

52、rli,Ray Perrault,Brittany Presten,Anka Reuel,Roger Roberts,Andrew Shi,Georgio Stoev,Shekhar Tewari,Dikshita Venkatesh,Cayla Volandes,Jakub WiatrakChapter 4:EconomyMedha Bankhwal,Erik Brynjolfsson,Cara Christopher,Michael Chui,Natalia Dorogi,Heather English,Murat Erer,Loredana Fattorini,Yolanda Gil,H

53、eather Hanselman,Vishy Kamalapuram,Njenga Kariuki,Akash Kaura,Elena Magrini,Nestor Maslej,Katherine Ottenbreit,Vanessa Parli,Ray Perrault,Brittany Presten,Roger Roberts,Cayla Volandes,Casey Weston,Hansen YangChapter 5:Science and MedicineRuss Altman,Kameron Black,Jonathan Chen,Jean-Benoit Delbrouck,

54、Joshua Edrich,Loredana Fattorini,Alejandro Lozano,Yolanda Gil,Ethan Goh,Armin Hamrah,Fateme Nateghi Haredasht,Tina Hernandez-Boussard,Yeon Mi Hwang,Rohan Koodli,Arman Koul,Curt Langlotz,Ashley Lewis,Chase Ludwig,Stephen P.Ma,Abdoul Jalil Djiberou Mahamadou,David Magnus,James Manyika,Nestor Maslej,Go

55、wri Nayar,Madelena Ng,Sophie Ostmeier,Vanessa Parli,Ray Perrault,Malkiva Pillai,Ossian Karl-Johan Ferdinand Rabow,Sean Riordan,Brennan Geti Simon,Kotoha Togami,Artem Trotsyuk,Maya Varma,Quinn WaeissChapter 6:Policy Elif Kiesow Cortez,Loredana Fattorini,Yolanda Gil,Julia Betts Lotufo,Vanessa Parli,Ra

56、y Perrault,Alexandra Rome,Lapo Santarlasci,Georgio Stoev,Russell Wald,Daniel ZhangThe AI Index would like to acknowledge the following individuals by chapter and section for their contributions of data,analysis,advice,and expert commentary included in the AI Index Report 2025:ContributorsArtificial

57、IntelligenceIndex Report 202510Chapter 7:Education John Etchemendy,Loredana Fattorini,Lili Gangas,Yolanda Gil,Rachel Goins,Laura Hinton,Sonia Koshy,Kirsten Lundgren,Nestor Maslej,Lisa Cruz Novohatski,Vanessa Parli,Ray Perrault,Allison Scott,Andreen Soley,Bryan Twarek,Laurens VehmeijerChapter 8:Publi

58、c OpinionEmily Capstick,John Etchemendy,Loredana Fattorini,Yolanda Gil,Njenga Kariuki,Nestor Maslej,Vanessa Parli,Ray PerraultOrganizationsContributors(contd)The AI Index would like to acknowledge the following individuals by chapter and section for their contributions of data,analysis,advice,and ex

59、pert commentary included in the AI Index Report 2025:Accenture Arnab Chakraborty,Patrick Connolly,Shekhar Tewari,Dikshita Venkatesh,Jakub WiatrakEpoch AIBen Cottier,Robi RahmanGitHubLuca Ronchi Darr,Kevin XuLightcastCara Christopher,Elena MagriniLinkedIn Mar Carpanelli,Akash Kaura Kory Kantenga,Rosi

60、e Hood,Casey WestonMcKinsey&CompanyMedha Bankhwal,Natalia Dorogi,Katherine Ottenbreit,Brittany Presten,Roger Roberts,Cayla VolandesQuidHeather English,Hansen YangThe AI Index also thanks Jeanina Matias,Nancy King,Carolyn Lehman,Shana Lynch,Jonathan Mindes,and Michi Turner for their help in preparing

61、 this report;Christopher Ellis for his help in maintaining the AI Index website;and Annie Benisch,Stacey Sickels Boyce,Marc Gough,Caroline Meinhardt,Drew Spence,Casey Weston,Madeleine Wright,and Daniel Zhang for their work in helping promote the report.Artificial IntelligenceIndex Report 202511Repor

62、t Highlights 12Chapter 1 Research and Development 24Chapter 2 Technical Performance 81Chapter 3 Responsible AI 160Chapter 4 Economy 214Chapter 5 Science and Medicine 280Chapter 6 Policy and Governance 323Chapter 7 Education 364Chapter 8 Public Opinion 394Appendix 414ACCESS THE PUBLIC DATATable of Co

63、ntentsArtificial IntelligenceIndex Report 2025Artificial IntelligenceIndex Report 202512Report Highlights1.Industry continues to make significant investments in AI and leads in notable AI model development,while academia leads in highly cited research.Industrys lead in notable model development,high

64、lighted in the two previous AI Index reports,has only grown more pronounced,with nearly 90%of notable models in 2024(compared to 60%in 2023)originating from industry.Academia has remained the single leading institutional producer of highly cited(top 100)publications over the past three years.2.China

65、 leads in AI research publication totals,while the United States leads in highly influential research.In 2023,China produced more AI publications(23.2%)and citations(22.6%)than any other country.Over the past three years,U.S.institutions have contributed the most top-100-cited AI publications.3.AI p

66、ublication totals continue to grow and increasingly dominate computer science.Between 2013 and 2023,the total number of AI publications in venues related to computer science and other scientific disciplines nearly tripled,increasing from approximately 102,000 to over 242,000.Proportionally,AIs share

67、 of computer science publications has risen from 21.6%in 2013 to 41.8%in 2023.4.The United States continues to be the leading source of notable AI models.In 2024,U.S.-based institutions produced 40 notable AI models,significantly surpassing Chinas 15 and Europes combined total of three.In the past d

68、ecade,more notable machine learning models have originated from the United States than any other country.5.AI models get increasingly bigger,more computationally demanding,and more energy intensive.New research finds that the training compute for notable AI models doubles approximately every five mo

69、nths,dataset sizes for training LLMs every eight months,and the power required for training annually.Large-scale industry investment continues to drive model scaling and performance gains.6.AI models become increasingly cheaper to use.The cost of querying an AI model that scores the equivalent of GP

70、T-3.5(64.8)on MMLU,a popular benchmark for assessing language model performance,dropped from$20.00 per million tokens in November 2022 to just$0.07 per million tokens by October 2024(Gemini-1.5-Flash-8B)a more than 280-fold reduction in approximately 18 months.Depending on the task,LLM inference pri

71、ces have fallen anywhere from 9 to 900 times per year.CHAPTER 1:Research and DevelopmentArtificial IntelligenceIndex Report 202513Report Highlights7.AI patenting is on the rise.Between 2010 and 2023,the number of AI patents has grown steadily and significantly,ballooning from 3,833 to 122,511.In jus

72、t the last year,the number of AI patents has risen 29.6%.As of 2023,China leads in total AI patents,accounting for 69.7%of all grants,while South Korea and Luxembourg stand out as top AI patent producers on a per capita basis.8.AI hardware gets faster,cheaper,and more energy efficient.New research s

73、uggests that machine learning hardware performance,measured in 16-bit floating-point operations,has grown 43%annually,doubling every 1.9 years.Price performance has improved,with costs dropping 30%per year,while energy efficiency has increased by 40%annually.9.Carbon emissions from AI training are s

74、teadily increasing.Training early AI models,such as AlexNet(2012),had modest amounts of carbon emissions at 0.01 tons.More recent models have significantly higher emissions for training:GPT-3(2020)at 588 tons,GPT-4(2023)at 5,184 tons,and Llama 3.1 405B(2024)at 8,930 tons.For perspective,the average

75、American emits 18 tons of carbon per year.1.AI masters new benchmarks faster than ever.In 2023,AI researchers introduced several challenging new benchmarks,including MMMU,GPQA,and SWE-bench,aimed at testing the limits of increasingly capable AI systems.By 2024,AI performance on these benchmarks saw

76、remarkable improvements,with gains of 18.8 and 48.9 percentage points on MMMU and GPQA,respectively.On SWE-bench,AI systems could solve just 4.4%of coding problems in 2023a figure that jumped to 71.7%in 2024.2.Open-weight models catch up.Last years AI Index revealed that leading open-weight models l

77、agged significantly behind their closed-weight counterparts.By 2024,this gap had nearly disappeared.In early January 2024,the leading closed-weight model outperformed the top open-weight model by 8.0%on the Chatbot Arena Leaderboard.By February 2025,this gap had narrowed to 1.7%.CHAPTER 1:Research a

78、nd Development(contd)CHAPTER 2:Technical PerformanceArtificial IntelligenceIndex Report 2025143.The gap closes between Chinese and U.S.models.In 2023,leading American models significantly outperformed their Chinese counterpartsa trend that no longer holds.At the end of 2023,performance gaps on bench

79、marks such as MMLU,MMMU,MATH,and HumanEval were 17.5,13.5,24.3,and 31.6 percentage points,respectively.By the end of 2024,these margins had narrowed substantially to 0.3,8.1,1.6,and 3.7 percentage points.4.AI model performance converges at the frontier.According to last years AI Index,the Elo score

80、difference between the top and 10th-ranked model on the Chatbot Arena Leaderboard was 11.9%.By early 2025,this gap had narrowed to 5.4%.Likewise,the difference between the top two models shrank from 4.9%in 2023 to just 0.7%in 2024.The AI landscape is becoming increasingly competitive,with high-quali

81、ty models now available from a growing number of developers.5.New reasoning paradigms like test-time compute improve model performance.In 2024,OpenAI introduced models like o1 and o3 that are designed to iteratively reason through their outputs.This test-time compute approach dramatically improved p

82、erformance,with o1 scoring 74.4%on an International Mathematical Olympiad qualifying exam,compared to GPT-4os 9.3%.However,this enhanced reasoning comes at a cost:o1 is nearly six times more expensive and 30 times slower than GPT-4o.6.More challenging benchmarks are continually being proposed.The sa

83、turation of traditional AI benchmarks like MMLU,GSM8K,and HumanEval,coupled with improved performance on newer,more challenging benchmarks such as MMMU and GPQA,has pushed researchers to explore additional evaluation methods for leading AI systems.Notable among these are Humanitys Last Exam,a rigoro

84、us academic test where the top system scores just 8.80%;FrontierMath,a complex mathematics benchmark where AI systems solve only 2%of problems;and BigCodeBench,a coding benchmark where AI systems achieve a 35.5%success ratewell below the human standard of 97%.7.High-quality AI video generators demon

85、strate significant improvement.In 2024,several advanced AI models capable of generating high-quality videos from text inputs were launched.Notable releases include OpenAIs SORA,Stable Video Diffusion 3D and 4D,Metas Movie Gen,and Google DeepMinds Veo 2.These models produce videos of significantly hi

86、gher quality compared to those from 2023.CHAPTER 2:Technical Performance(contd)Report HighlightsArtificial IntelligenceIndex Report 2025158.Smaller models drive stronger performance.In 2022,the smallest model registering a score higher than 60%on MMLU was PaLM,with 540 billion parameters.By 2024,Mic

87、rosofts Phi-3-mini,with just 3.8 billion parameters,achieved the same thresholdthe equivalent of a 142-fold reduction in two years.9.Complex reasoning remains a problem.Even though the addition of mechanisms such as chain-of-thought reasoning has significantly improved the performance of LLMs,these

88、systems still cannot reliably solve problems for which provably correct solutions can be found using logical reasoning,such as arithmetic and planning,especially on instances larger than those they were trained on.This has a significant impact on the trustworthiness of these systems and their suitab

89、ility in high-risk applications.10.AI agents show early promise.The launch of RE-Bench in 2024 introduced a rigorous benchmark for evaluating complex tasks for AI agents.In short time-horizon settings(two-hour budget),top AI systems score four times higher than human experts,but as the time budget i

90、ncreases,human performance surpasses AIoutscoring it two to one at 32 hours.AI agents already match human expertise in select tasks,such as writing Triton kernels,while delivering results faster and at lower costs.1.Evaluating AI systems with responsible AI(RAI)criteria is still uncommon,but new ben

91、chmarks are beginning to emerge.Last years AI Index highlighted the lack of standardized RAI benchmarks for LLMs.While this issue persists,new benchmarks such as HELM Safety and AIR-Bench help to fill this gap.2.The number of AI incident reports continues to increase.According to the AI Incidents Da

92、tabase,the number of reported AI-related incidents rose to 233 in 2024a record high and a 56.4%increase over 2023.CHAPTER 2:Technical Performance(contd)CHAPTER 3:Responsible AIReport HighlightsArtificial IntelligenceIndex Report 2025163.Organizations acknowledge RAI risks,but mitigation efforts lag.

93、A McKinsey survey on organizations RAI engagement shows that while many identify key RAI risks,not all are taking active steps to address them.Risks including inaccuracy,regulatory compliance,and cybersecurity were top of mind for leaders with only 64%,63%,and 60%of respondents,respectively,citing t

94、hem as concerns.4.Across the globe,policymakers demonstrate a significant interest in RAI.In 2024,global cooperation on AI governance intensified,with a focus on articulating agreed-upon principles for responsible AI.Several major organizationsincluding the OECD,European Union,United Nations,and Afr

95、ican Unionpublished frameworks to articulate key RAI concerns such as transparency and explainability,and trustworthiness.5.The data commons is rapidly shrinking.AI models rely on massive amounts of publicly available web data for training.A recent study found that data use restrictions increased si

96、gnificantly from 2023 to 2024,as many websites implemented new protocols to curb data scraping for AI training.In actively maintained domains in the C4 common crawl dataset,the proportion of restricted tokens jumped from 57%to 2033%.This decline has consequences for data diversity,model alignment,an

97、d scalability,and may also lead to new approaches to learning with data constraints.6.Foundation model research transparency improves,yet more work remains.The updated Foundation Model Transparency Indexa project tracking transparency in the foundation model ecosystemrevealed that the average transp

98、arency score among major model developers increased from 37%in October 2023 to 58%in May 2024.While these gains are promising,there is still considerable room for improvement.7.Better benchmarks for factuality and truthfulness.Earlier benchmarks like HaluEval and TruthfulQA,aimed at evaluating the f

99、actuality and truthfulness of AI models,have failed to gain widespread adoption within the AI community.In response,newer and more comprehensive evaluations have emerged,such as the updated Hughes Hallucination Evaluation Model leaderboard,FACTS,and SimpleQA.8.AI-related election misinformation spre

100、ad globally,but its impact remains unclear.In 2024,numerous examples of AI-related election misinformation emerged in more than a dozen countries and across over 10 social media platforms,including during the U.S.presidential election.However,questions remain about the measurable impacts of this pro

101、blem,with many expecting misinformation campaigns to have affected elections more profoundly than they did.CHAPTER 3:Responsible AI(contd)Report HighlightsArtificial IntelligenceIndex Report 2025179.LLMs trained to be explicitly unbiased continue to demonstrate implicit bias.Many advanced LLMsinclud

102、ing GPT-4 and Claude 3 Sonnetwere designed with measures to curb explicit biases,but they continue to exhibit implicit ones.The models disproportionately associate negative terms with Black individuals,more often associate women with humanities instead of STEM fields,and favor men for leadership rol

103、es,reinforcing racial and gender biases in decision making.Although bias metrics have improved on standard benchmarks,AI model bias remains a pervasive issue.10.RAI gains attention from academic researchers.The number of RAI papers accepted at leading AI conferences increased by 28.8%,from 992 in 20

104、23 to 1,278 in 2024,continuing a steady annual rise since 2019.This upward trend highlights the growing importance of RAI within the AI research community.CHAPTER 3:Responsible AI(contd)Report Highlights1.Global private AI investment hits record high with 26%growth.Corporate AI investment reached$25

105、2.3 billion in 2024,with private investment climbing 44.5%and mergers and acquisitions up 12.1%from the previous year.The sector has experienced dramatic expansion over the past decade,with total investment growing more than thirteenfold since 2014.2.Generative AI funding soars.Private investment in

106、 generative AI reached$33.9 billion in 2024,up 18.7%from 2023 and over 8.5 times higher than 2022 levels.The sector now represents more than 20%of all AI-related private investment.3.The U.S.widens its lead in global AI private investment.U.S.private AI investment hit$109.1 billion in 2024,nearly 12

107、 times higher than Chinas$9.3 billion and 24 times the U.K.s$4.5 billion.The gap is even more pronounced in generative AI,where U.S.investment exceeded the combined total of China and the European Union plus the U.K.by$25.4 billion,expanding on its$21.8 billion gap in 2023.4.Use of AI climbs to unpr

108、ecedented levels.In 2024,the proportion of survey respondents reporting AI use by their organizations jumped to 78%from 55%in 2023.Similarly,the number of respondents who reported using generative AI in at least one business function more than doubledfrom 33%in 2023 to 71%last year.CHAPTER 4:Economy

109、Artificial IntelligenceIndex Report 2025185.AI is beginning to deliver financial impact across business functions,but most companies are early in their journeys.Most companies that report financial impacts from using AI within a business function estimate the benefits as being at low levels.49%of re

110、spondents whose organizations use AI in service operations report cost savings,followed by supply chain management(43%)and software engineering(41%),but most of them report cost savings of less than 10%.With regard to revenue,71%of respondents using AI in marketing and sales report revenue gains,63%

111、in supply chain management,and 57%in service operations,but the most common level of revenue increases is less than 5%.6.Use of AI shows dramatic shifts by region,with Greater China gaining ground.While North America maintains its leadership in organizations use of AI,Greater China demonstrated one

112、of the most significant year-over-year growth rates,with a 27 percentage point increase in organizational AI use.Europe followed with a 23 percentage point increase,suggesting a rapidly evolving global AI landscape and intensifying international competition in AI implementation.7.Chinas dominance in

113、 industrial robotics continues despite slight moderation.In 2023,China installed 276,300 industrial robots,six times more than Japan and 7.3 times more than the United States.Since surpassing Japan in 2013,when China accounted for 20.8%of global installations,its share has risen to 51.1%.While China

114、 continues to install more robots than the rest of the world combined,this margin narrowed slightly in 2023,marking a modest moderation in its dramatic expansion.8.Collaborative and interactive robot installations become more common.In 2017,collaborative robots represented a mere 2.8%of all new indu

115、strial robot installations,a figure that climbed to 10.5%by 2023.Similarly,2023 saw a rise in service robot installations across all application categories except medical robotics.This trend indicates not just an overall increase in robot installations but also a growing emphasis on deploying robots

116、 for human-facing roles.9.AI is driving significant shifts in energy sources,attracting interest in nuclear energy.Microsoft announced a$1.6 billion deal to revive the Three Mile Island nuclear reactor to power AI,while Google and Amazon have also secured nuclear energy agreements to support AI oper

117、ations.10.AI boosts productivity and bridges skill gaps.Last years AI Index was among the first reports to highlight research showing AIs positive impact on productivity.This year,additional studies reinforced those findings,confirming that AI boosts productivity and,in most cases,helps narrow the g

118、ap between low-and high-skilled workers.CHAPTER 4:Economy(contd)Report HighlightsArtificial IntelligenceIndex Report 2025191.Bigger and better protein sequencing models emerge.In 2024,several large-scale,high-performance protein sequencing models,including ESM3 and AlphaFold 3,were launched.Over tim

119、e,these models have grown significantly in size,leading to continuous improvements in protein prediction accuracy.2.AI continues to drive rapid advances in scientific discovery.AIs role in scientific progress continues to expand.While 2022 and 2023 marked the early stages of AI-driven breakthroughs,

120、2024 brought even greater advancements,including Aviary,which trains LLM agents for biological tasks,and FireSat,which significantly enhances wildfire prediction.3.The clinical knowledge of leading LLMs continues to improve.OpenAIs recently released o1 set a new state-of-the-art 96.0%on the MedQA be

121、nchmarka 5.8 percentage point gain over the best score posted in 2023.Since late 2022,performance has improved 28.4 percentage points.MedQA,a key benchmark for assessing clinical knowledge,may be approaching saturation,signaling the need for more challenging evaluations.4.AI outperforms doctors on k

122、ey clinical tasks.A new study found that GPT-4 alone outperformed doctorsboth with and without AIin diagnosing complex clinical cases.Other recent studies show AI surpassing doctors in cancer detection and identifying high-mortality-risk patients.However,some early research suggests that AI-doctor c

123、ollaboration yields the best results,making it a fruitful area of further research.5.The number of FDA-approved,AI-enabled medical devices skyrockets.The FDA authorized its first AI-enabled medical device in 1995.By 2015,only six such devices had been approved,but the number spiked to 223 by 2023.6.

124、Synthetic data shows significant promise in medicine.Studies released in 2024 suggest that AI-generated synthetic data can help models better identify social determinants of health,enhance privacy-preserving clinical risk prediction,and facilitate the discovery of new drug compounds.7.Medical AI eth

125、ics publications are increasing year over year.The number of publications on ethics in medical AI nearly quadrupled from 2020 to 2024,rising from 288 in 2020 to 1,031 in 2024.CHAPTER 5:Science and MedicineReport HighlightsArtificial IntelligenceIndex Report 2025208.Foundation models come to medicine

126、.In 2024,a wave of large-scale medical foundation models were released,ranging from general-purpose multimodal models like Med-Gemini to specialized models such as EchoCLIP for echocardiology,VisionFM for ophthalmology,and ChexAgent for radiology.9.Publicly available protein databases grow in size.S

127、ince 2021,the number of entries in major public protein science databases has grown significantly,including UniProt(31%),PDB(23%),and AlphaFold(585%).This expansion has important implications for scientific discovery.10.AI research recognized by two Nobel Prizes.In 2024,AI-driven research received t

128、op honors,with two Nobel Prizes awarded for AI-related breakthroughs.Google DeepMinds Demis Hassabis and John Jumper won the Nobel Prize in Chemistry for their pioneering work on protein folding with AlphaFold.Meanwhile,John Hopfield and Geoffrey Hinton received the Nobel Prize in Physics for their

129、foundational contributions to neural networks.1.U.S.states are leading the way on AI legislation amid slow progress at the federal level.In 2016,only one state-level AI-related law was passed,increasing to 49 by 2023.In the past year alone,that number more than doubled to 131.While proposed AI bills

130、 at the federal level have also increased,the number passed remains low.2.Governments across the world invest in AI infrastructure.Canada announced a$2.4 billion AI infrastructure package,while China launched a$47.5 billion fund to boost semiconductor production.France committed$117 billion to AI in

131、frastructure,India pledged$1.25 billion,and Saudi Arabias Project Transcendence includes a$100 billion investment in AI.3.Across the world,mentions of AI in legislative proceedings keep rising.Across 75 countries,AI mentions in legislative proceedings increased by 21.3%in 2024,rising to 1,889 from 1

132、,557 in 2023.Since 2016,the total number of AI mentions has grown more than ninefold.CHAPTER 5:Science and Medicine(contd)CHAPTER 6:Policy and GovernanceReport HighlightsArtificial IntelligenceIndex Report 2025214.AI safety institutes expand and coordinate across the globe.In 2024,countries worldwid

133、e launched international AI safety institutes.The first emerged in November 2023 in the U.S.and the U.K.following the inaugural AI Safety Summit.At the AI Seoul Summit in May 2024,additional institutes were pledged in Japan,France,Germany,Italy,Singapore,South Korea,Australia,Canada,and the European

134、 Union.5.The number of U.S.AI-related federal regulations skyrockets.In 2024,59 AI-related regulations were introducedmore than double the 25 recorded in 2023.These regulations came from 42 unique agencies,twice the 21 agencies that issued them in 2023.6.U.S.states expand deepfake regulations.Before

135、 2024,only five statesCalifornia,Michigan,Washington,Texas,and Minnesotahad enacted laws regulating deepfakes in elections.In 2024,15 more states,including Oregon,New Mexico,and New York,introduced similar measures.Additionally,by 2024,24 states had passed regulations targeting deepfakes.1.Access to

136、 and enrollment in high school computer science(CS)courses in the U.S.has increased slightly from the previous school year,but gaps remain.Student participation varies by state,race and ethnicity,school size,geography,income,gender,and disability.2.CS teachers in the U.S.want to teach AI but do not

137、feel equipped to do so.Despite the 81%of CS teachers who agree that using AI and learning about AI should be included in a foundational CS learning experience,fewer than half of high school CS teachers feel equipped to teach AI.3.Two-thirds of countries worldwide offer or plan to offer K12 CS educat

138、ion.This fraction has doubled since 2019,with African and Latin American countries progressing the most.However,students in African countries have the least amount of access to CS education due to schools lack of electricity.CHAPTER 6:Policy and Governance(contd)CHAPTER 7:EducationReport HighlightsA

139、rtificial IntelligenceIndex Report 2025224.Graduates who earned their masters degree in AI in the U.S.nearly doubled between 2022 and 2023.While increased attention on AI will be slower to emerge in the number of bachelors and PhD degrees,the surge in masters degrees could indicate a developing tren

140、d for all degree levels.5.The U.S.continues to be a global leader in producing information,technology,and communications(ICT)graduates at all levels.Spain,Brazil,and the United Kingdom follow the U.S.as top producers at various levels,while Turkey boasts the best gender parity.1.The world grows caut

141、iously optimistic about AI products and services.Among the 26 nations surveyed by Ipsos in both 2022 and 2024,18 saw an increase in the proportion of people who believe AI products and services offer more benefits than drawbacks.Globally,the share of individuals who see AI products and services as m

142、ore beneficial than harmful has risen from 52%in 2022 to 55%in 2024.2.The expectation and acknowledgment of AIs impact on daily life is rising.Around the world,two thirds of people now believe that AI-powered products and services will significantly impact daily life within the next three to five ye

143、arsan increase of 6 percentage points since 2022.Every country except Malaysia,Poland,and India saw an increase in this perception since 2022,with the largest jumps in Canada(17%)and Germany(15%).3.Skepticism about the ethical conduct of AI companies is growing,while trust in the fairness of AI is d

144、eclining.Globally,confidence that AI companies protect personal data fell from 50%in 2023 to 47%in 2024.Likewise,fewer people today believe that AI systems are unbiased and free from discrimination compared to last year.4.Regional differences persist regarding AI optimism.First reported in the 2023

145、AI Index,significant regional differences in AI optimism endure.A large majority of people believe AI-powered products and services offer more benefits than drawbacks in countries like China(83%),Indonesia(80%),and Thailand(77%),while only a minority share this view in Canada(40%),the United States(

146、39%),and the Netherlands(36%).CHAPTER 7:Education(contd)CHAPTER 8:Public OpinionReport HighlightsArtificial IntelligenceIndex Report 2025235.People in the United States remain distrustful of self-driving cars.A recent American Automobile Association survey found that 61%of people in the U.S.fear sel

147、f-driving cars,and only 13%trust them.Although the percentage who expressed fear has declined from its 2023 peak of 68%,it remains higher than in 2021(54%).6.There is broad support for AI regulation among local U.S.policymakers.In 2023,73.7%of local U.S.policymakersspanning township,municipal,and co

148、unty levelsagreed that AI should be regulated,up significantly from 55.7%in 2022.Support was stronger among Democrats(79.2%)than Republicans(55.5%),though both registered notable increases over 2022.7.AI optimism registers sharp increase among countries that previously showed the most skepticism.Glo

149、bally,optimism about AI products and services has increased,with the sharpest gains in countries that were previously the most skeptical.In 2022,Great Britain(38%),Germany(37%),the United States(35%),Canada(32%),and France(31%)were among the least likely to view AI as having more benefits than drawb

150、acks.Since then,optimism has grown in these countries by 8%,10%,4%,8%,and 10%,respectively.8.Workers expect AI to reshape jobs,but fear of replacement remains lower.Globally,60%of respondents agree that AI will change how individuals do their job in the next five years.However,a smaller subset of re

151、spondents,36%,believe that AI will replace their jobs in the next five years.9.Sharp divides exist among local U.S.policymakers on AI policy priorities.While local U.S.policymakers broadly support AI regulation,their priorities vary.The strongest backing is for stricter data privacy rules(80.4%),ret

152、raining for the unemployed(76.2%),and AI deployment regulations(72.5%).However,support drops significantly for a law enforcement facial recognition ban(34.2%),wage subsidies for wage declines(32.9%),and universal basic income(24.6%).10.AI is seen as a time saver and entertainment booster,but doubts

153、remain on its economic impact.Global perspectives on AIs impact vary.While 55%believe it will save time,and 51%expect it will offer better entertainment options,fewer are confident in its health or economic benefits.Only 38%think AI will improve health,whilst 36%think AI will improve the national ec

154、onomy,31%see a positive impact on the job market,and 37%believe it will enhance their own jobs.CHAPTER 8:Public Opinion(contd)Report HighlightsArtificial IntelligenceIndex Report 2025CHAPTER 1:Research and DevelopmentTable of Contents25Overview 26Chapter Highlights 271.1 Publications 29Overview 29 T

155、otal Number of AI Publications 29 By Venue 31 By National Affiliation 32 By Sector 36 By Topic 38Top 100 Publications 39 By National Affiliation 39 By Sector 40 By Organization 411.2 Patents 42 Overview 42 By National Affiliation 431.3 Notable AI Models 46 By National Affiliation 46 By Sector 47 By

156、Organization 49 Model Release 50 Parameter Trends 52 Compute Trends 56 Highlight:Will Models Run Out of Data?59 Inference Cost 64 Training Cost 651.4 Hardware 68Overview 68 Highlight:Energy Efficiency and Environmental Impact 711.5 AI Conferences 75Conference Attendance 751.6 Open-Source AI Software

157、 77Projects 77Stars 79Chapter 1:Research and DevelopmentArtificial IntelligenceIndex Report 2025ACCESS THE PUBLIC DATATable of Contents26Artificial IntelligenceIndex Report 2025Chapter 1 PreviewThis chapter explores trends in AI research and development,beginning with an analysis of AI publications,

158、patents,and notable AI systems.These topics are examined through the lens of the countries,organizations,and sectors producing them.The chapter also covers AI model training costs,AI conference attendance,and open-source AI software.New additions this year include profiles of the evolving AI hardwar

159、e ecosystem,an assessment of AI trainings energy requirements and environmental impact,and a temporal analysis of model inference costs.OverviewCHAPTER 1:Research and DevelopmentArtificial IntelligenceIndex Report 2025Table of Contents27Artificial IntelligenceIndex Report 2025Chapter 1 PreviewChapte

160、r Highlights1.Industry continues to make significant investments in AI and leads in notable AI model development,while academia leads in highly cited research.Industrys lead in notable model development,highlighted in the two previous AI Index reports,has only grown more pronounced,with nearly 90%of

161、 notable models in 2024(compared to 60%in 2023)originating from industry.Academia has remained the single leading institutional producer of highly cited(top 100)publications over the past three years.2.China leads in AI research publication totals,while the United States leads in highly influential

162、research.In 2023,China produced more AI publications(23.2%)and citations(22.6%)than any other country.Over the past three years,U.S.institutions have contributed the most top-100-cited AI publications.3.AI publication totals continue to grow and increasingly dominate computer science.Between 2013 an

163、d 2023,the total number of AI publications in venues related to computer science and other scientific disciplines nearly tripled,increasing from approximately 102,000 to over 242,000.Proportionally,AIs share of computer science publications has risen from 21.6%in 2013 to 41.8%in 2023.4.The United St

164、ates continues to be the leading source of notable AI models.In 2024,U.S.-based institutions produced 40 notable AI models,significantly surpassing Chinas 15 and Europes combined total of three.In the past decade,more notable machine learning models have originated from the United States than any ot

165、her country.5.AI models get increasingly bigger,more computationally demanding,and more energy intensive.New research finds that the training compute for notable AI models doubles approximately every five months,dataset sizes for training LLMs every eight months,and the power required for training a

166、nnually.Large-scale industry investment continues to drive model scaling and performance gains.CHAPTER 1:Research and DevelopmentArtificial IntelligenceIndex Report 2025Table of Contents28Artificial IntelligenceIndex Report 2025Chapter 1 PreviewChapter Highlights(contd)CHAPTER 1:Research and Develop

167、mentArtificial IntelligenceIndex Report 20256.AI models become increasingly affordable to use.The cost of querying an AI model that scores the equivalent of GPT-3.5(64.8)on MMLU,a popular benchmark for assessing language model performance,dropped from$20.00 per million tokens in November 2022 to jus

168、t$0.07 per million tokens by October 2024(Gemini-1.5-Flash-8B)a more than 280-fold reduction in approximately 18 months.Depending on the task,LLM inference prices have fallen anywhere from 9 to 900 times per year.7.AI patenting is on the rise.Between 2010 and 2023,the number of AI patents has grown

169、steadily and significantly,ballooning from 3,833 to 122,511.In just the last year,the number of AI patents has risen 29.6%.As of 2023,China leads in total AI patents,accounting for 69.7%of all grants,while South Korea and Luxembourg stand out as top AI patent producers on a per capita basis.8.AI har

170、dware gets faster,cheaper,and more energy efficient.New research suggests that machine learning hardware performance,measured in 16-bit floating-point operations,has grown 43%annually,doubling every 1.9 years.Price performance has improved,with costs dropping 30%per year,while energy efficiency has

171、increased by 40%annually.9.Carbon emissions from AI training are steadily increasing.Training early AI models,such as AlexNet(2012),had modest amounts of carbon emissions at 0.01 tons.More recent models have significantly higher emissions for training:GPT-3(2020)at 588 tons,GPT-4(2023)at 5,184 tons,

172、and Llama 3.1 405B(2024)at 8,930 tons.For perspective,the average American emits 18 tons of carbon per year.Table of Contents29Artificial IntelligenceIndex Report 2025Chapter 1 Preview242.7420132014201520162017201820192020202120222023050100150200250Number of AI publications in CS(in thousands)Number

173、 of AI publications in CS worldwide,201323Source:AI Index,2025|Chart:2025 AI Index report1.1 PublicationsThe figures below show the global count of English-language AI publications from 2010 to 2023,categorized by affiliation type,publication type,and region.New to this years report,the AI Index inc

174、ludes a section analyzing trends among the 100 most-cited AI publications,which can offer insights into particularly high-impact research.This year,the AI Index analyzed AI publication trends using the OpenAlex database.As a result,the numbers in this years report differ slightly from those in previ

175、ous editions.1 Given that there is a significant lag in the collection of publication metadata,and that in some cases it takes until the middle of any given year to fully capture the previous years publications,in this years report,the AI Index team elected to examine publication trends only through

176、 2023.OverviewThe following section reports on trends in the total number of English-language AI publications.Total Number of AI PublicationsFigure 1.1.1 displays the global count of AI publications.These are the publications with a computer science(CS)label in the OpenAlex catalog that were classif

177、ied by the AI Index as being related to AI.2 Between 2013 and 2023,the total number of AI 1.1 PublicationsChapter 1:Research and DevelopmentFigure 1.1.11 OpenAlex is a fully open catalog of scholarly metadata,including scientific papers,authors,institutions,and more.The AI Index used OpenAlex as a b

178、ibliographic database and automatically classified AI-related research using the latest version of the CSO Classifier.In previous years,the Index relied on third-party providers with different underlying data sources and classification methods.As a result,this years findings differ slightly from tho

179、se included in previous reports.Additionally,the AI Index applied the classifier only to papers that OpenAlex categorized under the broad field of computer science.This approach may have led to an undercount of AI-related publications by excluding research from fields like social sciences that emplo

180、y AI methodologies but fall outside the computer sciencedesignated classification.2 The CSO Classifier(v3.3)is an automated text classification system designed to categorize research papers in computer science using a comprehensive ontology of 15,000 topics and 166,000 relationships,including emergi

181、ng fields like GenAI,LLMs,and prompt engineering.It processes metadata(such as title and abstract)through three modules:a syntactic module for exact topic matches,a semantic module leveraging word embeddings to infer related topics,and a post-processing module that refines results by filtering outli

182、ers and adding relevant higher-level areas.Table of Contents30Artificial IntelligenceIndex Report 2025Chapter 1 Previewpublications more than doubled,rising from approximately 102,000 in 2013 to more than 242,000 in 2023.The increase over the last year was a meaningful 19.7%.Many fields within compu

183、ter science,from hardware and software engineering to human-computer interaction,are now contributing to AI.As a result,the observed growth reflects a broader and increased interest in AI across the discipline.201320142015201620172018201920202021202220230%5%10%15%20%25%30%35%40%45%AI publications in

184、 CS(%of total)41.76%AI publications in CS(%of total)worldwide,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.21.1 PublicationsChapter 1:Research and DevelopmentTable of Contents31Artificial IntelligenceIndex Report 2025Chapter 1 PreviewFigure 1.1.2 shows the proportion of computer s

185、cience publications in the OpenAlex corpus classified as AI-related.Figure 1.1.2 features the same data included in Figure 1.1.1 but in a proportional form.The share of AI publications has grown significantly,almost doubling from 2013 to 2023.By VenueAI researchers publish their work across various

186、venues.Figure 1.1.3 visualizes the total number of AI publications by venue type.In 2023,journals accounted for the largest share of AI publications(41.8%),followed by conferences(34.3%).Even though the total number of journal and conference publications has increased since 2013,the share of AI publ

187、ications in journals and conferences has steadily declined,from 52.6%and 36.4%in 2013 to 41.8%and 34.3%,respectively,in 2023.Conversely,AI publications in repositories like arXiv have seen a growing share.1.1 PublicationsChapter 1:Research and Development201320142015201620172018201920202021202220230

188、%5%10%15%20%25%30%35%40%45%AI publications in CS(%of total)41.76%AI publications in CS(%of total)worldwide,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.3Table of Contents32Artificial IntelligenceIndex Report 2025Chapter 1 PreviewBy National AffiliationFigure 1.1.4 visualizes AI pu

189、blications over time by region.3 In 2023,East Asia and the Pacific led AI research output,accounting for 34.5%of all AI publications,followed by Europe and Central Asia(18.2%)and North America(10.3%).4While Figure 1.1.4 examines the geographic distribution of AI publications,identifying which region

190、s produce the most research,Figure 1.1.5 focuses on citations,measuring the share of total AI publication citations attributed to work originating from each region.As of 2023,AI publications from East Asia and the Pacific accounted for the largest share of AI article citations at 37.1%(Figure 1.1.5)

191、.In 2017,citation shares from East Asia and the Pacific and North America were roughly equal,but since then,North American and European citation shares have declined,while East Asia and the Pacifics share has risen sharply.1.1 PublicationsChapter 1:Research and Development201320142015201620172018201

192、920202021202220230%5%10%15%20%25%30%35%AI publications in CS(%of total)0.89%,Sub-Saharan Africa1.66%,Latin America and the Caribbean5.18%,Middle East and North Africa9.98%,South Asia10.31%,North America18.15%,Europe and Central Asia19.37%,Unknown34.46%,East Asia and Paci?cAI publications in CS(%of t

193、otal)by region,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.43 Regions in this chapter are classified according to the World Bank analytical grouping.The AI Index determines the country affiliation of authors using the“countries”field from the authorship data.This field lists all

194、the countries an author is affiliated with,as retrieved from OpenAlex based on institutional affiliations.These affiliations can be explicitly stated in the paper or inferred from the authors most recent publications.When counting publications by country,the AI Index assigns one count to each countr

195、y linked to the publication.For example,if a paper has three authors,two affiliated with institutions in the U.S.and one in China,the publication is counted once for the U.S.and once for China.4 A publication may have an“unknown”country affiliation when the authors institutional affiliation is missi

196、ng or incomplete.This issue arises due to various factors,including unstructured or omitted institution names,platform functional deficiencies,group authorship practices,unstandardized affiliation labeling,document type inconsistencies,or the authors limited publication record.The problem as it rela

197、tes to OpenAlex is addressed in this paper;however,the issue of missing institutions pertains to other bibliographic databases as well.Table of Contents33Artificial IntelligenceIndex Report 2025Chapter 1 Preview1.1 PublicationsChapter 1:Research and Development201320142015201620172018201920202021202

198、220230%5%10%15%20%25%30%35%40%AI publication citations in CS(%of total)0.89%,Sub-Saharan Africa1.35%,Latin America and the Caribbean7.55%,Unknown7.69%,South Asia7.97%,Middle East and North Africa15.59%,North America21.88%,Europe and Central Asia37.07%,East Asia and Paci?cAI publication citations in

199、CS(%of total)by region,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.5Table of Contents34Artificial IntelligenceIndex Report 2025Chapter 1 PreviewIn 2023,China was the global leader in AI article publications,accounting for 23.2%of the total,compared to 15.2%from Europe and 9.2%fro

200、m India(Figure 1.1.6).5 Since 2016,Chinas share has steadily increased,while the proportion attributed to Europe has declined.AI publications attributed to the United States remained relatively stable until 2021 but have shown a slight decline since then.201320142015201620172018201920202021202220230

201、%5%10%15%20%25%AI publications in CS(%of total)9.20%,United States9.22%,India15.22%,Europe20.65%,Unknown22.51%,Rest of the world23.20%,ChinaAI publications in CS(%of total)by select geographic areas,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.661.1 PublicationsChapter 1:Research

202、and Development5 For the“Europe”designation in this and other chapters of the report,the AI Index follows the list of countries defined by the United Nations Statistics Division.6 To maintain concision,the AI Index visualized results for a select group of countries.However,full results for all count

203、ries will be available on the AI Indexs Global Vibrancy Tool,which is set to be updated in summer 2025.For immediate access to country-specific research and development data,please contact the AI Index team.Table of Contents35Artificial IntelligenceIndex Report 2025Chapter 1 PreviewIn 2023,Chinese A

204、I publications accounted for 22.6%of all AI citations,followed by Europe at 20.9%and the United States at 13.0%(Figure 1.1.7).As with total AI publications,the late 2010s marked a turning point when China surpassed Europe and the U.S.as the leading source of AI publication citations.2013201420152016

205、20172018201920202021202220230%5%10%15%20%25%30%35%AI publication citations in CS(%of total)6.10%,India7.54%,Unknown13.03%,United States20.90%,Europe22.60%,China29.83%,Rest of the worldAI publication citations in CS(%of total)by select geographic areas,201323Source:AI Index,2025|Chart:2025 AI Index r

206、eportFigure 1.1.71.1 PublicationsChapter 1:Research and DevelopmentTable of Contents36Artificial IntelligenceIndex Report 2025Chapter 1 PreviewBy SectorAcademic institutions remain the primary source of AI publications worldwide(Figure 1.1.8).In 2013,they accounted for 85.9%of all AI publications,a

207、figure that remained high,at 84.9%,in 2023.Industry contributed 7.1%of AI publications in 2023,followed by government institutions at 4.9%and nonprofit organizations at 1.7%.201320142015201620172018201920202021202220230%10%20%30%40%50%60%70%80%90%AI publications in CS(%of total)1.35%,Other1.70%,Nonp

208、ro?t4.90%,Government7.14%,Industry84.91%,AcademiaAI publications in CS(%of total)by sector,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.871.1 PublicationsChapter 1:Research and Development7 For Figures 1.1.8 and 1.1.9,publications with unknown affiliations were excluded from the f

209、inal visualization.Table of Contents37Artificial IntelligenceIndex Report 2025Chapter 1 PreviewAI publications emerge from various sectors in differing proportions across geographic regions.In the United States,a higher share of AI publications(16.5%)comes from industry compared to China(8.0%)(Figur

210、e 1.1.9).Among major geographic areas,China has the highest percentage of AI publications originating from the education sector(84.5%).75.61%16.49%4.02%3.88%79.49%9.62%6.79%4.09%84.45%8.02%6.96%0.58%0%10%20%30%40%50%60%70%80%90%GovernmentNonpro?tIndustryAcademiaUnited StatesEuropeChinaAI publication

211、s(%of total)AI publications in CS(%of total)by sector and select geographic areas,2023Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.91.1 PublicationsChapter 1:Research and DevelopmentTable of Contents38Artificial IntelligenceIndex Report 2025Chapter 1 PreviewBy TopicMachine learning was

212、the most prevalent research topic in AI publications in 2023,comprising 75.7%of publications,followed by computer vision(47.2%),pattern recognition(25.9%)and natural language processing(17.1%)(Figure 1.1.10).Over the past year,there has been a sharp increase in publications on generative AI.20132014

213、201520162017201820192020202120222023050100150Number of AI publications(in thousands)5.25,Robotics11.28,Multi-agent systems12.00,Logic and reasoning13.07,Generative AI17.34,Evolutionary computation21.82,Knowledge based systems41.40,Natural language processing62.90,Pattern recognition114.61,Computer v

214、ision183.78,Machine learningNumber of AI publications by select top topics,201323Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.1081.1 PublicationsChapter 1:Research and Development8 The AI Index categorized papers using its own topic classifier.It is possible for a single publication to

215、be assigned multiple topic labels.Table of Contents39Artificial IntelligenceIndex Report 2025Chapter 1 Preview5034776654445934764433216433108774311024681012141618 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 62 64 66SingaporeIsraelUnited Arab EmiratesUnited KingdomSouth KoreaCanada

216、Hong KongGermanyChinaUnited States202320222021Number of highly cited publications in top 100Number of highly cited publications in top 100 by select geographic areas,202123Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.11Top 100 PublicationsWhile tracking total AI publications provides a

217、broad view of research activity,focusing on the most-cited papers offers a perspective of the fields most influential work.This analysis sheds light on where some of the most groundbreaking and influential AI research is emerging.This year,the AI Index identified the 100 most-cited AI publications i

218、n 2021,2022,and 2023,using citation data from OpenAlex.This analysis was further supplemented with insights from Google Scholar and Semantic Scholar.9 Some of the most highly cited AI publications in 2023 included OpenAIs GPT-4 technical report,Metas Llama 2 technical report,and Googles PaLM-E techn

219、ical report.It is important to note that due to citation lag,the most-cited papers in this years report may change in future editions.By National AffiliationFigure 1.1.11 illustrates the geographic distribution of the top 100 most-cited AI publications by year.From 2021 to 2023,the U.S.consistently

220、had the highest number of top-cited publications,with 64 in 2021,59 in 2022,and 50 in 2023.10 In each of these years,China ranked second.Since 2021,the U.S.share of top AI publications has gradually declined.1.1 PublicationsChapter 1:Research and Development9 The full methodological guide can be acc

221、essed in the Appendix,along with the list of the top 100 articles.10 A publication can have multiple authors from different countries or organizations.For example,if a paper includes authors from multiple countries,each country is credited once.As a result,the totals in this sections figures exceed

222、100.Table of Contents40Artificial IntelligenceIndex Report 2025Chapter 1 Preview4272524227193517341731171AcademiaIndustryIndustry and academiaMixedOther051015202530354045202320222021SectorNumber of highly cited publications in top 100Number of highly cited publications in top 100 by sector,202123Sou

223、rce:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.1211By SectorAcademia consistently produces the most top-cited AI publications,with 42 in 2023,27 in 2022,and 34 in 2021(Figure 1.1.12).Notably,there was a sharp decline in industry contributions,with the number of top 100 publications dropping

224、from 17 in 2021 and 19 in 2022 to just 7 in 2023.As AI research grows more competitive,many industrial AI labs are publishing less frequently or disclosing fewer details about their research in their publications.1.1 PublicationsChapter 1:Research and Development11 The“mixed”designation includes all

225、 intersector collaborations that are not industry and academia(e.g.,industry and government,academia and nonprofit).Some institutions lack data for 2021 because they did not have papers included in the top 100 that year.Since papers can have multiple authors from different institutions,the total ins

226、titutional tags in Figure 1.1.12 may exceed 100.Also,because two of the papers had authors with an unknown sectoral affiliation,the total sum of publications in Figure 1.1.12 is 98.Table of Contents41Artificial IntelligenceIndex Report 2025Chapter 1 PreviewBy OrganizationFigure 1.1.13 highlights the

227、 organizations that produced the top 100 most-cited AI publications from 2021 to 2023.Some organizations may have empty bars on the chart if they lacked a top 100 publication in a given year.Additionally,Figure 1.1.13 highlights only the top 10 institutions,though many others contribute significant

228、research.Google led each year,but it tied with Tsinghua University in 2023,when both contributed eight publications to the top 100.In 2023,Carnegie Mellon University was the highest-ranked U.S.academic institution.8866555444201099433222151075221GoogleTsinghuaUniversityCarnegie MellonUniversityMicros

229、oftBeijing Academy ofArti?cial IntelligenceHong Kong University ofScience and TechnologyShanghaiAI LaboratoryChinese Academyof SciencesMetaNvidia0246810121416182022202320222021OrganizationNumber of highly cited publications in top 100Number of highly cited publications in top 100 by organization,202

230、123Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.1.131.1 PublicationsChapter 1:Research and DevelopmentTable of Contents42Artificial IntelligenceIndex Report 2025Chapter 1 PreviewThis section examines trends over time in global AI patents,which can reveal important insights into the evolut

231、ion of innovation,research,and development within AI.Additionally,analyzing AI patents can reveal how these advances are distributed globally.Similar to the publications data,there is a noticeable delay in AI patent data availability,with 2023 being the most recent year for which data is accessible.

232、The data in this section is sourced from patent-level bibliographic records in PATSTAT Global,a comprehensive database provided by the European Patent Office(EPO).12122.5120102011201220132014201520162017201820192020202120222023020406080100120Number of AI patents granted(in thousands)Number of AI pat

233、ents granted worldwide,201023Source:AI Index,2025|Chart:2025 AI Index reportFigure 1.2.11.2 PatentsOverviewFigure 1.2.1 examines the global growth in granted AI patents from 2010 to 2023.Over the past dozen years,the number of AI patents has grown steadily and significantly,increasing from 3,833 in

234、2010 to 122,511 in 2023.In the last year,the number of AI patents has risen 29.6%.1.2 PatentsChapter 1:Research and Development12 More details on the methodology behind the patent analysis in this section can be found in the Appendix.Table of Contents43Artificial IntelligenceIndex Report 2025Chapter

235、 1 Preview201020112012201320142015201620172018201920202021202220230%10%20%30%40%50%60%70%80%90%Granted AI patents(%of world total)0.02%,Sub-Saharan Africa0.02%,Middle East and North Africa0.04%,Latin America and the Caribbean0.15%,Rest of the world0.37%,South Asia2.77%,Europe and Central Asia14.23%,

236、North America82.40%,East Asia and Paci?cGranted AI patents(%of world total)by region,201023Source:AI Index,2025|Chart:2025 AI Index reportBy National AffiliationFigure 1.2.2 showcases the regional breakdown of granted AI patents,as in the number of patents filed in different regions across the world

237、.As of 2023,the bulk of the worlds granted AI patents(82.4%)originated from East Asia and the Pacific,with North America being the next largest contributor at 14.2%.Since 2010,the gap in AI patent grants between East Asia and the Pacific and North America has steadily widened.Figure 1.2.2131.2 Paten

238、tsChapter 1:Research and Development13 Patent standards and laws vary across countries and regions,so these charts should be interpreted with caution.More detailed country-level patent information will be released in a subsequent edition of the AI Indexs Global Vibrancy Tool.Table of Contents44Artif

239、icial IntelligenceIndex Report 2025Chapter 1 PreviewDisaggregated by geographic area,the majority of the worlds granted AI patents are from China(69.7%)and the United States(14.2%)(Figure 1.2.3).The share of AI patents originating from the United States has declined from a peak of 42.8%in 2015.Figur

240、e 1.2.3 and Figure 1.2.4 document which countries lead in AI patents per capita.In 2023,the country with the most granted AI patents per 100,000 inhabitants was South Korea(17.3),followed by Luxembourg(15.3)and China(6.1)(Figure 1.2.3).Figure 1.2.5 highlights the change in granted AI patents per cap

241、ita from 2013 to 2023.Luxembourg,China and Sweden experienced the greatest increase in AI patenting per capita during that time period.Figure 1.2.3201020112012201320142015201620172018201920202021202220230%10%20%30%40%50%60%70%Granted AI patents(%of world total)0.37%,India2.77%,Europe13.00%,Rest of t

242、he world14.16%,United States69.70%,ChinaGranted AI patents(%of world total)by select geographic areas,201023Source:AI Index,2025|Chart:2025 AI Index report1.2 PatentsChapter 1:Research and DevelopmentTable of Contents45Artificial IntelligenceIndex Report 2025Chapter 1 PreviewFigure 1.2.4Figure 1.2.5

243、0.270.380.400.430.470.520.740.970.981.224.585.206.0815.3117.270123456789101112131415161718GreeceAustraliaNetherlandsFranceDenmarkUnited KingdomSwedenFinlandSingaporeGermanyJapanUnited StatesChinaLuxembourgSouth KoreaGranted AI patents(per 100,000 inhabitants)Granted AI patents per 100,000 inhabitant

244、s by country,2023Source:AI Index,2025|Chart:2025 AI Index report230%240%365%463%580%730%1,028%1,043%1,097%1,653%2,546%2,851%3,453%6,317%8,216%0%1,000%2,000%3,000%4,000%5,000%6,000%7,000%8,000%DenmarkAustraliaJapanFranceUnited StatesUnited KingdomNetherlandsSouth KoreaGermanyFinlandSingaporeGreeceSwe

245、denChinaLuxembourg%change of granted AI patents(per 100,000 inhabitants)Source:AI Index,2025|Chart:2025 AI Index reportPercentage change of granted AI patents per 100,000 inhabitants by country,2013 vs.20231.2 PatentsChapter 1:Research and DevelopmentTable of Contents46Artificial IntelligenceIndex R

246、eport 2025Chapter 1 Preview1111315400510152025303540South KoreaSaudi ArabiaIsraelCanadaFranceChinaUnited StatesNumber of notable AI modelsNumber of notable AI models by select geographicareas,2024Source:Epoch AI,2025|Chart:2025 AI Index report20032006200920122015201820212024010203040506070Number of

247、notable AI models3,Europe15,China40,United StatesNumber of notable AI models by select geographicareas,200324Source:Epoch AI,2025|Chart:2025 AI Index report1.3 Notable AI ModelsBy National AffiliationTo illustrate the evolving geopolitical landscape of AI,the AI Index shows the country of origin of

248、notable models.Figure 1.3.1 displays the total number of notable AI models attributed to the location of researchers affiliated institutions.16 In 2024,the United States led with 40 notable AI models,followed by China with 15 and France with three.All major geographic groups,including the United Sta

249、tes,China,the European Union,and the United Kingdom,reported releasing fewer notable models in 2024 than in the previous year(Figure 1.3.2).Since 2003,the United States has produced more models than other major countries such as the United Kingdom,China,and Canada(Figure 1.3.3).It is difficult to pi

250、npoint the exact cause of the decline in total model releases,but it may stem from a combination of factors:increasingly large training runs,the growing complexity of AI technology,and the heightened challenge of Figure 1.3.117Figure 1.3.2This section explores notable AI models.Epoch AI,an AI Index

251、data provider,uses the term“notable machine learning models”to designate particularly influential models within the AI/machine learning ecosystem.Epoch maintains a database of 900 AI models released since the 1950s,selecting entries based on criteria such as state-of-the-art advancements,historical

252、significance,or high citation rates.Since Epoch manually curates the data,some models considered notable by some may not be included.Analyzing these models provides a comprehensive overview of the machine learning landscapes evolution,both in recent years and over the past few decades.Some models ma

253、y be missing from the dataset;however,the dataset can reveal trends in relative terms.Examples of notable AI models include GPT-4o,Claude 3.5,and AlphaGeometry.Within this section,the AI Index explores trends in notable models from various perspectives,including country of origin,originating organiz

254、ation,gradient of model release,parameter count,and compute usage.The analysis concludes with an examination of machine learning training as well as inference costs.1.3 Notable AI ModelsChapter 1:Research and Development14“AI system”refers to a computer program or product based on AI,such as ChatGPT

255、.“AI model”includes a collection of parameters whose values are learned during training,such as GPT-415 New and historic models are continually added to the Epoch AI database,so the total year-by-year counts of models included in this years AI Index might not exactly match those published in last ye

256、ars report.The data is from a snapshot taken on March 17,2025.16 A machine learning model is associated with a specific country if at least one author of the paper introducing it has an affiliation with an institution based in that country.In cases where a models authors come from several countries,

257、double-counting can occur.17 This chart highlights model releases from a select group of geographic areas.More comprehensive data on model releases by country will be available in the upcoming AI Index Global Vibrancy Tool release.Table of Contents47Artificial IntelligenceIndex Report 2025Chapter 1

258、Preview1101120216061100101560Number of notable AI models by geographic area,200324(sum)Source:Epoch AI,2025|Chart:2025 AI Index reportdeveloping new modeling approaches.Epoch AIs curation of notable models may overlook releases from certain countries that receive less coverage.The AI Index,in cooper

259、ation with Epoch,is committed to improving global representation in the AI model ecosystem.If readers believe that models from specific countries are missing,they are encouraged to contact the AI Index team,which will work to address the issue.Figure 1.3.31.3 Notable AI ModelsChapter 1:Research and

260、DevelopmentBy SectorFigure 1.3.4 illustrates the sectoral origin of notable AI releases by the year the models were released.Epoch categorizes models based on their source:Industry includes companies such as Google,Meta,and OpenAI;academia covers universities like Tsinghua,MIT,and Oxford;government

261、refers to state-affiliated research institutes like the UKs Alan Turing Institute for AI and Abu Dhabis Technology Innovation Institute;and research collectives encompass nonprofit AI research organizations such as the Allen Institute for AI and the Fraunhofer Institute.Until 2014,academia led in te

262、rms of releasing machine learning models.Since then,industry has taken the lead.According to Epoch AI,in 2024,industry produced 55 notable AI models.That same year,Epoch AI identified no notable AI models originating from academia(Figure 1.3.5).18 Over time,industry-academia collaborations have cont

263、ributed to a growing number of models.The proportion of notable AI models originating from industry has steadily increased over the past decade,growing to 90.2%in 2024.18 This figure should be interpreted with caution.A count of zero academic models does not mean that no notable models were produced

264、 by academic institutions in 2023,but rather that Epoch AI has not identified any as notable.Additionally,academic publications often take longer to gain recognition,as highly cited papers introducing significant architectures may take years to achieve prominence.Table of Contents48Artificial Intell

265、igenceIndex Report 2025Chapter 1 Preview20032004200520062007200820092010201120122013201420152016201720182019202020212022202320240%20%40%60%80%100%Notable AI models(%of total)0.00%,Academia0.00%,Academiagovernment collaboration0.00%,Academiaresearch collective collaboration0.00%,Research collective0.

266、00%,Industryresearch collective collaboration0.00%,Government1.64%,Industrygovernment collaboration8.20%,Industryacademia collaboration90.16%,IndustryNotable AI models(%of total)by sector,200324Source:Epoch AI,2025|Chart:2025 AI Index report20032004200520062007200820092010201120122013201420152016201

267、720182019202020212022202320240102030405060Number of notable AI models0,Academia0,Academiagovernment collaboration0,Academiaresearch collective collaboration0,Research collective0,Industryresearch collective collaboration0,Government1,Industrygovernment collaboration5,Industryacademia collaboration55

268、,IndustryNumber of notable AI models by sector,200324Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.4Figure 1.3.51.3 Notable AI ModelsChapter 1:Research and DevelopmentTable of Contents49Artificial IntelligenceIndex Report 2025Chapter 1 Preview77644433222222201234567Zhipu AIWriterUC Berke

269、leyTencentMITDeepSeekByteDanceMistral AIAnthropicNvidiaMetaAppleAlibabaOpenAIGoogleAcademiaIndustryNumber of notable AI modelsNumber of notable AI models by organization,2024Source:Epoch AI,2025|Chart:2025 AI Index report1878239362525222217161515151412010203040506070809010011012013014015016017018019

270、0Allen Institute for AIAlibabaUniversity of WashingtonSalesforceMITUniversity of OxfordNvidiaUC BerkeleyTsinghua UniversityStanford UniversityCarnegie Mellon UniversityOpenAIMicrosoftMetaGoogleAcademiaIndustryResearch collectiveNumber of notable AI modelsNumber of notable AI models by organization,2

271、01424(sum)Source:Epoch AI,2025|Chart:2025 AI Index reportBy OrganizationFigure 1.3.6 and Figure 1.3.7 highlight the organizations leading in the production of notable machine learning models in 2024 and over the past decade.In 2024,the top contributors were OpenAI(7 models),Google(6),and Alibaba(4).

272、Since 2014,Google has led with 186 notable models,followed by Meta(82)and Microsoft(39).Among academic institutions,Carnegie Mellon University(25),Stanford University(25),and Tsinghua University(22)have been the most prolific since 2014.Figure 1.3.619Figure 1.3.71.3 Notable AI ModelsChapter 1:Resear

273、ch and Development19 In the organizational tally figures,research published by DeepMind is classified under Google.Table of Contents50Artificial IntelligenceIndex Report 2025Chapter 1 PreviewModel ReleaseMachine learning models are released under various access types,each with varying levels of open

274、ness and usability.API access models,like OpenAIs o1,allow users to interact with models via queries without direct access to their underlying weights.Open weights(restricted use)models,like DeepSeeks-V3,provide access to their weights but impose limitations,such as prohibiting commercial use or red

275、istribution.Hosted access(no API)models,like Gemini 2.0 Pro,refer to models available through a platform interface but without programmatic access.Open weights(unrestricted)models,like AlphaGeometry,are fully open,allowing free use,modification,and redistribution.Open weights(noncommercial)models,li

276、ke Mistral Large 2,share their weights but restrict use to research or noncommercial purposes.Lastly,unreleased models,like ESM3 98B,remain proprietary,accessible only to their developers or select partners.The unknown designation refers to models that have unclear or undisclosed access types.Figure

277、 1.3.8 illustrates the different access types under which models have been released.20 In 2024,API access was the most common release type,with 20 of 61 models made available this way,followed by open weights with restricted use and unreleased models.Figure 1.3.9 visualizes machine learning model ac

278、cess types over time from a proportional perspective.In 2024,most AI models were released via API access(32.8%),which has seen a steady rise since 2020.122091012161110272732202010192336211922141030193638171326303228505851725475861056120142015201620172018201920202021202220232024020406080100120API acc

279、essHosted access(no API)Open weights(noncommercial)Open weights(restricted use)Open weights(unrestricted)UnreleasedUnknownNumber of notable AI modelsNumber of notable AI models by access type,201424Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.8211.3 Notable AI ModelsChapter 1:Research a

280、nd Development20 Hosted access refers to using computing resources or services(such as software,hardware,or storage)provided remotely by a third party,rather than personally owning or managing them.Instead of running software or infrastructure locally,hosted access involves accessing these resources

281、 via the cloud or another remote service,typically over the internet.For example,using GPUs through platforms like AWS,Google Cloud,or Microsoft Azurerather than running them on ones own hardwareis considered hosted access.21 Not all models in the Epoch database are categorized by access type,so the

282、 totals in Figures 1.3.8 through 1.3.10 may not fully align with those reported elsewhere in the chapter.Table of Contents51Artificial IntelligenceIndex Report 2025Chapter 1 Preview1.3 Notable AI ModelsChapter 1:Research and Development201420152016201720182019202020212022202320240%10%20%30%40%50%60%

283、70%80%90%100%Notable AI models(%of total)3.28%,Unknown8.20%,Hosted access(no API)9.84%,Open weights(noncommercial)11.48%,Open weights(unrestricted)16.39%,Unreleased18.03%,Open weights(restricted use)32.79%,API accessNotable AI models(%of total)by access type,201424Source:Epoch AI,2025|Chart:2025 AI

284、Index report16332229161311911152624292837373021374019143848183228505851725475861056120142015201620172018201920202021202220232024020406080100120Open sourceOpen(restricted use)Open(noncommercial)UnreleasedUnknownNumber of notable AI modelsNumber of notable AI models by training code access type,201424

285、Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.9Figure 1.3.10In traditional open-source software releases,all components,including the training code,are typically made available.However,this is often not the case with AI technologies,where even developers who release model weights may wit

286、hhold the training code.Figure 1.3.10 categorizes notable AI models by the openness of their code release.In 2024,the majority60.7%were launched without corresponding training code.Table of Contents52Artificial IntelligenceIndex Report 2025Chapter 1 Preview2003 2004 2005 2006 2007 2008 2009 20102011

287、20122013201420152016201720182019 2020 20212022 2023 202410010K1M100M10B1TAcademiaAcademiagovernmentIndustryIndustryresearch collectiveIndustryacademiaGovernmentResearch collectivePublication dateNumber of parameters(log scale)Number of parameters of notable AI models by sector,200324Source:Epoch AI,

288、2025|Chart:2025 AI Index reportParameter TrendsParameters in machine learning models are numerical values learned during training that determine how a model interprets input data and makes predictions.Models with more parameters require more data to be trained,but they can take on more tasks and typ

289、ically outperform models with fewer parameters.Figure 1.3.11 demonstrates the parameter count of machine learning models in the Epoch dataset,categorized by the sector from which the models originate.Figure 1.3.12 visualizes the same data,but for a smaller selection of notable models.Parameter count

290、s have risen sharply since the early 2010s,reflecting the growing complexity of their architecture,greater availability of data,improvements in hardware,and proven efficacy of larger models.High-parameter models are particularly notable in the industry sector,underscoring the substantial financial r

291、esources available to industry to cover the computational costs of training on vast volumes of data.Several of the figures below use a log scale to reflect the exponential growth in AI model parameters and compute in recent years.Figure 1.3.111.3 Notable AI ModelsChapter 1:Research and DevelopmentTa

292、ble of Contents53Artificial IntelligenceIndex Report 2025Chapter 1 Preview1.3 Notable AI ModelsChapter 1:Research and DevelopmentAlexNetDeepSeek-V3Qwen2.5-72BMistral Large 2Llama 2-70BPaLM(540B)Megatron-Turing NLG 530BGPT-3 175B(davinci)BERT-LargeTransformerERNIE 3.0 TitanRoBERTa Large20122013201420

293、15201620172018201920202021202220232024100M1B10B100B1TAcademiaIndustryIndustryacademiaPublication dateNumber of parameters(log scale)Number of parameters of select notable AI models by sector,201224Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.12Table of Contents54Artificial IntelligenceI

294、ndex Report 2025Chapter 1 Preview20102011201220132014201520162017201820192020202120222023202410K1M100M10B1T100TPublication dateTraining dataset size(tokens-log scale)Training dataset size of notable AI models,201024Source:Epoch AI,2025|Chart:2025 AI Index reportLlama 3.1-405BTransformerGPT-3 175B(da

295、vinci)DeepSeek-V3PaLM(540B)GPT-4AlexNetQwen2.5-72BFigure 1.3.131.3 Notable AI ModelsChapter 1:Research and DevelopmentAs model parameter counts have increased,so has the volume of data used to train AI systems.Figure 1.3.13 illustrates the growth in dataset sizes used to train notable machine learni

296、ng models.The Transformer model,released in 2017 and widely credited with sparking the large language model revolution,was trained on approximately 2 billion tokens.By 2020,GPT-3 175Bone of the models underpinning the original ChatGPTwas trained on an estimated 374 billion tokens.In contrast,Metas f

297、lagship LLM,Llama 3.3,released in the summer of 2024,was trained on roughly 15 trillion tokens.According to Epoch AI,LLM training datasets double in size approximately every eight months.Table of Contents55Artificial IntelligenceIndex Report 2025Chapter 1 PreviewTraining models on increasingly large

298、 datasets has led to significantly longer training times(Figure 1.3.14).Some state-of-the-art models,such as Llama 3.1-405B,required approximately 90 days to traina typical window by todays standards.Googles Gemini 1.0 Ultra,released in late 2023,took around 100 days.This stands in stark contrast to

299、 AlexNet,one of the first models to leverage GPUs for enhanced performance,which trained in just five to six days in 2012.Notably,AlexNet was trained on far less advanced hardware.2010201120122013201420152016201720182019202020212022202320240.1110100Publication dateTraining length(days-log scale)Trai

300、ning length of notable AI models,201024Source:Epoch AI,2025|Chart:2025 AI Index reportAlexNetTransformerBERT-LargeRoBERTa LargeGPT-3 175B(davinci)Megatron-Turing NLG 530BPaLM(540B)GPT-4Llama 3.1-405BFigure 1.3.141.3 Notable AI ModelsChapter 1:Research and DevelopmentTable of Contents56Artificial Int

301、elligenceIndex Report 2025Chapter 1 PreviewCompute TrendsThe term“compute”in AI models denotes the computational resources required to train and operate a machine learning model.Generally,the complexity of the model and the size of the training dataset directly influence the amount of compute needed

302、.The more complex a model is,and the larger the underlying training data,the greater the amount of compute required for training.Before the final training run,researchers conduct numerous test runs throughout the R&D phase.While training a single model is relatively inexpensive,the cumulative cost o

303、f multiple R&D runs and the necessary datasets quickly becomes significant.These figures reflect only the final training run,not the entire R&D process.Figure 1.3.15 visualizes the training compute required for notable machine learning models over the past 22 years.Recently,the compute usage of nota

304、ble AI models has increased exponentially.22 Epoch estimates that the training compute of notable AI models doubles roughly every five months.This trend has been especially pronounced in the last five years.This rapid rise in compute demand has important implications.For instance,models requiring mo

305、re computation often have larger environmental footprints,and companies typically have more access to computational resources than academic institutions.For reference,Chapter 2 of the AI Index analyzes the relationship between improvements in computational resources and model performance.2003 2004 2

306、005 2006 2007 2008 2009 2010201120122013201420152016201720182019 2020 20212022 2023 20241000.01110010K1M100M10BAcademiaIndustryIndustryacademiaAcademiagovernmentIndustryresearch collectiveGovernmentResearch collectivePublication dateTraining compute(petaFLOP-log scale)Training compute of notable AI

307、models by sector,200324Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.152322 FLOP stands for“floating-point operation.”A floating-point operation is a single arithmetic operation involving floating-point numbers,such as addition,subtraction,multiplication,or division.The number of FLOP a

308、processor or computer can perform per second is an indicator of its computational power.The higher the FLOP rate,the more powerful the computer.The number of floating-point operations used to train an AI model reflects its requirement for computational resources during development.23 Estimating trai

309、ning compute is an important aspect of AI model analysis,yet it often requires indirect measurement.When direct reporting is unavailable,Epoch estimates compute by using hardware specifications and usage patterns or by counting arithmetic operations based on model architecture and training data.In c

310、ases where neither approach is feasible,benchmark performance can serve as a proxy to infer training compute by comparing models with known compute values.Full details of Epochs methodology can be found in the documentation section of their website.1.3 Notable AI ModelsChapter 1:Research and Develop

311、mentTable of Contents57Artificial IntelligenceIndex Report 2025Chapter 1 PreviewFigure 1.3.16 highlights the training compute of notable machine learning models since 2012.For example,AlexNet,one of the models that popularized the now standard practice of using GPUs to improve AI models,required an

312、estimated 470 petaFLOP for training.24 The original Transformer,released in 2017,required around 7,400 petaFLOP.OpenAIs GPT-4o,one of the current state-of-the-art foundation models,required 38 billion petaFLOP.Creating cutting-edge AI models now demands a colossal amount of data,computing power,and

313、financial resources that are not available to academia.Most leading AI models are coming from industry,a trend that was first highlighted in last years AI Index.Although the gap has slightly narrowed this year,the trend persists.24 A petaFLOP(PFLOP)is a unit of computing power equal to one quadrilli

314、on(10)floating-point operations per second.1.3 Notable AI ModelsChapter 1:Research and DevelopmentDeepSeek-V3Qwen2.5-72BLlama 2-70BClaude 2PaLM(540B)Megatron-Turing NLG 530BGPT-3 175B(davinci)RoBERTa LargeBERT-LargeTransformerSegment Anything ModelAlexNetGPT-42012201320142015201620172018201920202021

315、202220232024100010K100K1M10M100M1B10B100BLanguageVisionMultimodalPublication dateTraining compute(petaFLOP-log scale)Mistral Large 2Claude 3.5 SonnetGemini 1.5 ProGPT-4oERNIE 3.0 TitanTraining compute of notable AI models by domain,201224Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.16Ta

316、ble of Contents58Artificial IntelligenceIndex Report 2025Chapter 1 Preview1.3 Notable AI ModelsChapter 1:Research and DevelopmentThe launch of DeepSeeks V3 model in December 2024 garnered significant attention,particularly because it achieved exceptionally high performance while requiring far fewer

317、computational resources than many leading LLMs.Figure 1.3.17 compares the training compute of notable machine learning models from the United States and China,highlighting a key trend:Top-tier AI models from the U.S.have generally been far more computationally intensive than Chinese models.According

318、 to Epoch AI,the top 10 Chinese language models by training compute have scaled at a rate of about three times per year since late 2021considerably slower than the five times per year trend observed in the rest of the world since 2018.2018201920202021202220232024100100010K100K1M10M100M1B10B100BUnite

319、d StatesChinaPublication dateTraining compute(petaFLOP log scale)GPT-4GPT-3 175B(davinci)Grok-2Claude 3.5 SonnetDeepSeek-V3Doubao-proERNIE 3.0 TitanQwen2.5-72BTraining compute of select notable AI models in the United States and China,201824Source:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.1

320、7Table of Contents59Artificial IntelligenceIndex Report 2025Chapter 1 PreviewHighlight:Will Models Run Out of Data?One of the key drivers of substantive algorithmic improvements in AI systems has been the scaling of models and their training on ever-larger datasets.However,as the supply of internet

321、training data becomes increasingly depleted,concerns have grown about the sustainability of this scaling approach and the potential for a data bottleneck,where returns to scale diminish.Last years AI Index explored various factors in this debate,including the availability of existing internet data a

322、nd the potential for training models on synthetic data.New research this year suggests that the current stock of data may last longer than previously expected.Epoch AI has updated its previous estimates for when AI researchers might run out of data.In its latest research,the team estimated the total

323、 effective stock of data available for training models according to token count(Figure 1.3.18).Common Crawl,an open repository of web crawl data frequently used in AI training,is estimated to contain a median of 130 trillion tokens.The indexed web holds approximately 510 trillion tokens,while the en

324、tire web contains around 3,100 trillion.Additionally,the total stock of images is estimated at 300 trillion,and video at 1,350 trillion.130T510T3,100T300T1,350TCommon CrawlIndex webWhole web(incl.private data)ImagesVideo300T1000T3000TData sourceNumber of tokens(median-log scale)Estimated median data

325、 stocksSource:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.181.3 Notable AI ModelsChapter 1:Research and DevelopmentTable of Contents60Artificial IntelligenceIndex Report 2025Chapter 1 PreviewThe Epoch AI research team projects,with an 80%confidence interval,that the current stock of training

326、data will be fully utilized between 2026 and 2032(Figure 1.3.19).Several factors influence the point in time when data is likely to run out.One key factor is the historical growth of dataset sizes,which depends on how many people generate and contribute content to the internet.Another important fact

327、or is computer usage.If models are trained in a compute-optimal manner,the available data stock can last longer.However,if models are overtrained to achieve more compute-efficient inference performance,the stock is likely to be depleted sooner.When AI models are overtrained,meaning they are trained

328、for an extended period beyond the typical point of diminishing returns,they may achieve more compute-efficient inferencethat is,they can process prompts(make predictions,generate text,etc.)using less computational power.However,this comes at a cost:The stock(i.e.,data available to train the model)ma

329、y be depleted more quickly.Llama 3.1-405BDBRXFalcon-180BPaLM(540B)FLAN 137BGPT-3 175B(davinci)2020202220242026202820302032203410B100B1T10T100T10 15 Estimated stock of dataMedian date of full stock utilization(5x overtraining)Median date of full stock utilizationPublication dateE?ective stock(number

330、of tokens-log scale)Projections of the stock of public text and data usageSource:Epoch AI,2025|Chart:2025 AI Index reportFigure 1.3.191.3 Notable AI ModelsChapter 1:Research and DevelopmentHighlight:Will Models Run Out of Data?(contd)Table of Contents61Artificial IntelligenceIndex Report 2025Chapter

331、 1 PreviewThese projections differ slightly from Epochs earlier estimates,which predicted that high-quality text data would be depleted by 2024.The revised projections reflect an updated methodology that incorporates new research showing that web data performs better than curated corpora and that mo

332、dels can be trained on the same datasets multiple times.The realization that carefully filtered web data is effective and that repeated training on the same dataset is viable has expanded estimates of the available data stock.As a result,the Epoch researchers pushed back their forecasts of when data

333、 depletion might occur.Using synthetic datadata generated by AI models themselvesto train models has also been suggested as a solution to potential data shortages.The 2024 AI Index suggests there are limitations associated with this approach,namely that models trained this way are likely to lose representation of the tails of distributions when performing repeated training cycles on synthetic data