Stanford University's Fei-Fei Li team released the "2025 Artificial Intelligence Index Report".
The "2025 Artificial Intelligence Index Report" is the eighth annual report published by the Stanford Human-Centered Artificial Intelligence Institute (Stanford HAI), comprehensively tracking global AI development. The report covers research and development, technical performance,...
2025artificialintelligentThe Index Report is a human-centered report from Stanford University.artificialintelligentStanford InstituteAIThe eighth edition of the annual report, comprehensively tracking global [data/information/etc.], was released.AIDevelopment Updates. The report covers R&D, technical performance, and responsible practices.AIThe eight areas of economic impact, science and medicine, policy governance, education, and public opinion have been added.AIIn-depth analysis of hardware, inference costs, patent trends, and more. The report reveals the trends for 2024. AI The industry's 12 key findings include,AIBenchmark performance improved significantly (e.g., MMMU benchmark score increased by 18.8% within one year), with 90% of the significant improvements.AIThe models come from enterprises, and the performance gap between Chinese and American models has narrowed to near parity globally.AIInvestment reached $252.3 billion (a 26% increase), butAIBias and incident reports continue to increase. The report provides publicly available data tools to assist policymakers, businesses, and researchers in their decision-making.
Get2025artificialintelligentIndex ReportOriginal PDF fileScan the QR code to follow and reply: 20250408
2025artificialintelligentThe core content of the index report
Chapter 1: Research and Development
- Contributions from industry and academia90% of well-known companies in 2024AIThe models come from enterprises (60% in 2023), while academia remains the main source of high-impact research.
- Global Research LandscapeChina inAIThe United States leads in total publications (23.2%) and patents (69.7%), while it excels in top-ranked highly cited research.
- Technology Trends:
- The model size continues to expandTraining computational power requirements double every 5 months, and datasets double every 8 months.
- Hardware efficiency improvementCosts decrease by 30% annually, and energy efficiency improves by 40%.
- Reasoning costs have been significantly reduced.:GPTThe inference cost of the -3.5 level model decreased by 280 times within 18 months.
Chapter Two: Technical Specifications
- Benchmarking Breakthrough2024AISignificant improvements were achieved on new benchmarks such as MMMU (+18.8%) and GPQA (+48.9%), but complex reasoning (such as mathematical proofs) remains a challenge.
- open sourceThe Rise of Models:open sourceThe performance gap with the closed-source model narrowed from 8% to 1.7%.
- US-China competitionThe gap between Chinese models and the US models in tests such as MMLU has narrowed from 17.5% to 0.3%.
- AIAgency potentialIt surpasses humans in short-term tasks (such as programming), but lags behind in long-term tasks.
- Video generation technologyOpenAIModels such as SORA enable high-quality text-to-video generation.
- Key issuesTraditional benchmarks (such as MMLU) are nearing saturation, necessitating the development of more complex evaluation methods (such as the "ultimate human test" benchmark).AI(Only 8.8% scored).
Chapter Three: Taking ResponsibilityAI
- Assess the current situationStandardization is responsibleAI(R)AIAssessments remain scarce, but new benchmarks such as HELM Safety,AIR-Bench begins to fill the gaps.
- Risks and Responses:AIRelated incidents increased by 56.4% year-on-year (reaching 233 incidents in 2024), with only 64% of companies taking actual measures to address the risks. Implicit biases in the model persist (such as gender and racial bias).
- Policy cooperationThe OECD, the EU, and other organizations have released a global framework emphasizing transparency and credibility.
- Data ChallengeThe availability of training data has decreased due to website anti-scraping measures, which may affect the diversity of future models.
- Industry TrendsThe medical and election fields have become RAIFocus, butAIThe actual impact on the election is still unclear.
Chapter Four: Economics
- Investment growth:worldwideAIInvestment reached $252.3 billion (in 2024), generativeAIIt accounted for 33.9 billion (annual increase of 18.7%). The United States (109.1 billion) is 12 times larger than China (930 million).
- Enterprise Applications78% of organizations useAI(55% in 2023), with the fastest growth in adoption rate in China (+27%).
- Productivity impact:AIOn average, it improves enterprise efficiency, but most cost savings are less than 10%.
- Robot marketChina has installed 51.1% of the world's industrial robots, with collaborative robots accounting for 10.5%.
- Regional differencesNorth America leads investment, but Asia (such as China and Indonesia) is leading in application growth.
Chapter 5: Science and Medicine
- Scientific breakthroughModels such as AlphaFold 3 have advanced protein structure prediction, and related research was awarded the 2024 Nobel Prize in Chemistry.AIIt surpasses doctors in cancer detection and high-risk patient identification, but "AIThe "+doctor" collaboration yields the best results.
- Medical applicationsFDA approvedAIThe number of medical devices increased from 6 in 2015 to 223 in 2023.
- Synthetic data:AIThe generated data aids in drug discovery and privacy protection research.
- Ethical concernsMedicalAIThe number of ethics papers has quadrupled in four years.
- Cutting-edge trendsLarge-scale medical basic models (such as Med-Gemini(Accelerate implementation)
Chapter Six: Policy and Governance
- Legislation acceleratesUS state levelAIThe number of laws will increase from 1 in 2016 to 131 in 2024, representing an annual increase of 136% in federal regulations. This applies to legislation in 75 countries.AIThe number of mentions increased ninefold compared to 2016.
- International cooperation: Founded in the United States, Britain, Japan, etc.AISecurity Institute, EU promotes the "AIbill".
- Deep Fake Regulation24 US states pass anti-deepfake laws, focusing on election security.
- National strategyChina establishes a $47.5 billion semiconductor fund; Saudi Arabia launches a $100 billion fund.AIplan.
- challengeThe global coordination framework still needs improvement, and the participation of developing countries is insufficient.
Chapter Seven: Education
- Global popularityTwo-thirds of countries have incorporated computer science into their K-12 curricula (compared to one-third in 2019), with Africa making the fastest progress but lagging behind in infrastructure.
- Inadequate teacher preparation81% of US CS instructors believe it is necessary to teachAIHowever, only half of them possess the ability.
- Degree Trends:USAAIThe number of master's graduates has doubled in two years, and the number of graduates in the ICT field is among the highest in the world.
- Gender differencesTürkiye has the best gender equality in the ICT sector, while the gender gap in STEM in Europe and the United States remains significant.
Chapter 8: Public Opinion
- Optimism rises55% of people worldwide believeAIThe benefits outweigh the drawbacks (52% in 2022), with China (83%) and Indonesia (80%) being the most optimistic, and the United States (39%) and the Netherlands (36%) being the lowest.
- Trust crisisOnly 47% of people believeAIThe company protects its data (50% in 2023), forautomaticDriver trust level is only 13%.
- Policy support73% of US local governments supportAIRegulation exists, but priorities are diverging (data privacy is supported by 80%, while the ban on facial recognition is supported by only 34%).
- Workplace impact60% of people expectAIThey are changing their work habits, but only 36% are worried about losing their jobs.
- Cultural differencesAsia is paying more attentionAIWhile acknowledging the entertainment value, Europe and the United States are more concerned about the economic impact.
2025artificialintelligent12 Key Findings from the Index Report
- AISignificant performance improvement in complex benchmark testsIn 2024,AIIn MMMU (MultimodalPerformance in new benchmarks such as Comprehension, GPQA (Complex Reasoning), and SWE-bench (Code Generation) improved by 18.8, 48.9, and 67.3 percentage points, respectively. Video generation technologies (such as Open...)AIThe quality of SORA has been greatly improved.AIThe agent has surpassed human performance in short-duration programming tasks.
- AIAccelerate integration into daily life:AIIn the fields of healthcare and transportationfastPopular, FDA approvedAIThe number of medical devices increased from 6 in 2015 to 223 in 2023. Waymo provides 150,000 services per week.automaticBaidu Apollo Go, a driver-centric service, has commenced commercial operation in multiple cities across China.
- Enterprise investment and adoption rates hit record highs2024 GlobalAIInvestment reached $252.3 billion, with the US ($109.1 billion) being 12 times that of China. 78% of enterprises have already used [the technology/method/etc.].AI(55% in 2023), generativeAIIt accounts for 20% of the total investment and has significantly boosted productivity.
- China and the United StatesAIThe performance gap between models has been significantly narrowed.In 2023, the US model led China by 17.5 percentage points in tests such as MMLU, and the gap narrowed to 0.3% in 2024.open sourceThe performance gap between the model and the closed-source model has decreased from 8% to 1.7%, indicating that competition in display technology is becoming increasingly fierce.
- ResponsibleAIThe assessment is still imperfect, but new tools are emerging.:AIRelated incidents increased by 56.4% year-on-year (233 incidents in 2024), with only 64% of companies taking actual risk mitigation measures. HELM SafetyAINew benchmarks such as R-Bench are beginning to fill the gap in responsible...AIThe gap in the assessment.
- GlobalAIOptimism is rising, but regional differences are significant.China (83%) and Indonesia (80%)AIThe most optimistic attitude was observed in the United States (39%) and the Netherlands (36%). Optimism in previously pessimistic countries like Germany and France increased by 10 percentage points.
- AICost reduction and energy efficiency improvement:GPTThe inference cost of the -3.5 level model decreased by 280 times in 18 months (from $20/million tokens to $0.07). Hardware costs decreased by 30% annually, and energy efficiency improved by 40%.
- Government acceleratesAIRegulation and infrastructure investmentUS state levelAIThe number of laws increased from 1 in 2016 to 131 in 2024, across 75 countries' legislation.AIThe number of mentions increased ninefold compared to 2016. China established a $47.5 billion semiconductor fund, and Saudi Arabia launched a $100 billion fund.AIplan.
- AIEducation is widespread, but there is a shortage of qualified teachers.Two-thirds of countries worldwide have incorporated computer science into their K-12 curricula (compared to one-third in 2019), but 81% of US CS teachers believe it is necessary to teach it.AIOnly half of them have the capability. Progress in Africa is lagging due to insufficient infrastructure.
- AIDriven scientific breakthroughs and recognized by the Nobel PrizeThe 2024 Nobel Prize in Chemistry was awarded to the AlphaFold team (for protein folding), and the Nobel Prize in Physics was awarded to...Neural NetworksThe founder.AIOutperforming physicians in cancer detection and high-risk patient identification.
- Complex reasoning is stillAIShortcomings:AIIts accuracy rate is only 8.8% in logical reasoning tasks such as the Mathematical Olympiad, and it performs poorly in complex planning tasks such as PlanBench, limiting its application in high-risk scenarios.
- Reduced data sharing may affect model diversity.The availability of publicly available training data is expected to decrease by 20-33% in 2023-2024 (due to anti-scraping measures), which may exacerbate model homogenization and give rise to new learning methods with limited data.
Get2025artificialintelligentIndex ReportOriginal PDF fileScan the QR code to follow and reply: 20250408
Xiamen University DeepSeekLarge ModelEmpowering Universities, Enterprises, and Governments (PDF file) AITutorial materials
How to useAIHow to create heartwarming, grandmother-themed captions and comics? A tutorial for generating viral traffic.