April 8, 2025 Recently, Stanford HAI's "AI Index Report 2025" served as a comprehensive "annual health check" for the artificial intelligence field, unprecedentedly revealing the latest dynamics of global AI research, development, deployment, and societal impact. The report clearly depicts the astonishing speed of global AI development, particularly in the leaps in model performance, significant improvements in hardware efficiency, and the widespread adoption of AI across various industries. This report is not only a crucial reference for AI professionals but also provides key insights for all sectors of society to understand the future trajectory of AI.
Reshaping the Landscape of Scientific Research: East Asia Leads in Publication Output, China's Patent "Acceleration" is Noteworthy
The first chapter of the report, "Research and Development," showcases the new global landscape of AI research. In 2023, East Asia and the Pacific region dominated global AI publication output with a 34.5% share. This data highlights the region's strong vitality in fundamental AI research. Meanwhile, Europe and Central Asia (18.2%) and North America (10.3%) also contributed significant research findings. More noteworthy is China's significant growth in AI patent grants, particularly its percentage change per 100,000 residents. This indicates China's continued efforts not only in AI research but also in rapidly translating research results into practical applications. Furthermore, this year's report includes, for the first time, an analysis of AI model training energy consumption, environmental impact, and the changing costs of model inference over time. This provides crucial data for a more comprehensive assessment of the sustainability of AI development.
A Surge of Star Models: Parameter Scale Continues to Expand, Industry's "Financial Muscle" is Evident
Since the early 2010s, the number of parameters in machine learning models has experienced exponential growth. This is not merely a numerical change but reflects increasingly sophisticated model architectures, the availability of massive datasets, rapid advancements in hardware performance, and the superior effectiveness of larger models in solving complex problems. The report particularly emphasizes the widespread use of high-parameter models in the industry. This clearly shows that industry giants possess the substantial resources to bear the enormous computational costs required for large-scale data training. In terms of model accessibility, in 2024, the number of well-known AI models with open weights (for non-commercial use) and accessible via APIs remained significant. This suggests that the dissemination and democratization of AI technology is continuing.
Hardware Engine Efficiency Leaps: Astonishing Energy Efficiency Improvements, Green AI Takes Center Stage
The hardware ecosystem, the infrastructure supporting AI development, is constantly evolving, and energy efficiency is becoming a key indicator of its advancement. The report cites an encouraging example: the NVIDIA B100, released in March 2024, boasts an energy efficiency of 2.5 trillion floating-point operations per watt. Compared to the NVIDIA P100, released in April 2016, with 74 billion floating-point operations per watt, the energy efficiency has improved by a staggering 33.8 times. This significant leap not only signifies more powerful computing capabilities but also offers hope for reducing AI's energy consumption and environmental footprint. The report assesses the energy needs and environmental impact of AI training and presents carbon emission estimates for some AI models, suggesting that "green AI" will become a crucial direction for future AI development.
Flourishing Open-Source Power: GitHub Stars Shine Bright, US Leads Community Contributions
The open-source AI software ecosystem remains vibrant. As of 2024, the United States contributed the highest percentage to GitHub AI projects, reaching 23.42%, followed by Europe (19.15%) and India (19.91%). The number of "stars" on GitHub visually reflects the developer community's recognition and support for open-source projects. TensorFlow, OpenCV, Keras, and PyTorch, due to their widespread popularity among developers, are among the most prominent projects. In 2024, the United States garnered a total of 21.1 million GitHub stars, further demonstrating its leading position in the open-source AI field.
Continuous Technological Breakthroughs: Stunningly Realistic Video Generation, Autonomous Driving Safety Potentially Exceeding Human Capabilities
In terms of technological performance, the report covers several cutting-edge areas, including language, vision, video, code, and robotics. Open evaluation platforms such as Chatbot Arena are widely used to measure the capabilities of top AI systems. The rapid progress in video generation is particularly noteworthy. The report mentions that the quality of Pika's 2025 video of "Will Smith eating spaghetti" represents a qualitative leap compared to the 2023 version. In code generation, benchmarks like SWE-bench are used to evaluate the ability of AI models to solve real-world software engineering problems. Autonomous driving technology is also rapidly evolving, with companies like Waymo achieving significant milestones in autonomous driving mileage. A preliminary study in the report even suggests that autonomous vehicles may be safer than human-driven cars, showing that the number of accidents reported per million miles for Waymo vehicles is significantly lower than the human driving benchmark. China's development in autonomous driving is also rapidly progressing.
Responsible AI is Increasingly Important: Joint Efforts from Academia and Industry, AI Incidents Sound a Warning Bell
“Responsible AI” (RAI) has become a crucial issue in AI development. Academia's emphasis on RAI continues to rise, with a significant increase in the number of RAI-related papers accepted at top AI conferences, reaching 1278 in 2024, a 28.8% increase from 992 in 2023. The United States leads in the number of RAI papers published. Privacy and data governance, fairness, transparency and explainability, and security are core areas of RAI research. The industry is also increasing its investment in RAI, although investment size is correlated with company revenue. However, a survey shows that a significant proportion of organizations have reported experiencing AI incidents, serving as a warning for responsible AI development. Cybersecurity, regulatory compliance, and personal privacy are considered the most relevant AI risks by businesses.
AI Empowers the Economy and Research: Continued Surge in Talent Demand, Unlimited Potential in Biomedicine
The driving role of AI in the economy is becoming increasingly apparent. The report analyzes the global and US AI workforce demand, and the distribution of AI talent across different skills and industries. Global investment in AI remains high, with global private investment in generative AI alone reaching $3.94 billion in 2024. The number of newly established AI companies worldwide is also growing. In science and medicine, AI demonstrates immense potential. AlphaFold3 can more accurately predict the interactions between proteins and biomolecules such as DNA and RNA, promising to accelerate drug development. Performance evaluations of large language models in healthcare tasks are also attracting significant attention.
Proactive Policies and Governance: Increased Public Sector Investment, Global Collaboration on the Agenda
Public sector investment in AI is a crucial driver of AI development. The report analyzes public spending on AI-related contracts in the US and Europe, as well as AI-related funding in the US. Globally, international organizations, including the G7, are actively exploring AI governance and regulatory frameworks.
Looking Ahead
The "AI Index Report 2025" not only documents the significant progress in the global AI field over the past year but also foreshadows the continued profound impact of AI technology on our lives and work in the future. In the face of the opportunities and challenges presented by AI, strengthening international cooperation and promoting responsible AI development will be key to ensuring that AI technology benefits all of humanity.
Report Download Link: https://hai-production.s3.amazonaws.com/files/hai_ai_index_report_2025.pdf