Introduction to the Al Index Report 2024
2024 年 Al 指数报告简介

Message From the Co-directors
联合主任致辞

Message From the Co-directors (cont'd)
联合主任们的消息（续）

Ray Perrault and Jack Clark
Ray Perrault 和 Jack Clark

Top 10 Takeaways 十大要点

1. Al beats humans on some tasks, but not on all. Al has surpassed human performance on several benchmarks, including some in image classification, visual reasoning, and English understanding. Yet it trails behind on more complex tasks like competition-level mathematics, visual commonsense reasoning and planning.
   在某些任务上，人工智能超过人类，但并非所有任务都是如此。人工智能在一些基准测试中已经超过了人类的表现，包括图像分类、视觉推理和英语理解等方面。然而，在更复杂的任务上，如竞赛级数学、视觉常识推理和规划方面，人工智能仍然落后。
2. Industry continues to dominate frontier Al research. In 2023, industry produced 51 notable machine learning models, while academia contributed only 15 . There were also 21 notable models resulting from industry-academia collaborations in 2023, a new high.
   行业继续主导前沿的人工智能研究。2023 年，行业发布了 51 个显著的机器学习模型，而学术界仅贡献了 15 个。2023 年，行业与学术界合作产生了 21 个显著模型，创下新高。
3. Frontier models get way more expensive. According to AI Index estimates, the training costs of state-of-the-art Al models have reached unprecedented levels. For example, OpenAl's GPT-4 used an estimated million worth of compute to train, while Google's Gemini Ultra cost million for compute.
   边境模型变得更加昂贵。根据 AI 指数估计，现代 AI 模型的训练成本已经达到了前所未有的水平。例如，OpenAI 的 GPT-4 使用了估计价值 百万的计算资源进行训练，而谷歌的 Gemini Ultra 则需要 百万计算成本。
4. The United States leads China, the EU, and the U.K. as the leading source of top AI models. In 2023, 61 notable Al models originated from U.S.-based institutions, far outpacing the European Union's 21 and China's 15.
   美国领先于中国、欧盟和英国，成为顶尖人工智能模型的主要来源。2023 年，61 个知名的人工智能模型源自美国机构，远远超过欧盟的 21 个和中国的 15 个。

5. Robust and standardized evaluations for LLM responsibility are seriously lacking.
5. 对LLM责任的强大和标准化评估严重匮乏。

New research from the AI Index reveals a significant lack of standardization in responsible Al reporting. Leading developers, including OpenAI, Google, and Anthropic, primarily test their models against different responsible Al benchmarks. This practice complicates efforts to systematically compare the risks and limitations of top Al models.

6. Generative Al investment skyrockets. Despite a decline in overall Al private investment last year, funding for generative AI surged, nearly octupling from 2022 to reach billion. Major players in the generative Al space, including OpenAI, Anthropic, Hugging Face, and Inflection, reported substantial fundraising rounds.
   生成性人工智能投资大幅增长。尽管去年整体人工智能私人投资出现下滑，但生成性人工智能的资金激增，从 2022 年增长近 8 倍，达到 亿美元。生成性人工智能领域的主要参与者，包括 OpenAI、Anthropic、Hugging Face 和 Inflection，都宣布了重大的募资轮次。
7. The data is in: Al makes workers more productive and leads to higher quality work. In 2023, several studies assessed Al's impact on labor, suggesting that Al enables workers to complete tasks more quickly and to improve the quality of their output. These studies also demonstrated Al's potential to bridge the skill gap between low- and high-skilled workers. Still, other studies caution that using Al without proper oversight can lead to diminished performance.
   数据显示：人工智能提高了工人的生产力，促进了更高质量的工作。2023 年，几项研究评估了人工智能对劳动力的影响，表明人工智能使工人能够更快地完成任务并提高其产出质量。这些研究还展示了人工智能在弥合低技能和高技能工人之间的技能差距方面的潜力。然而，其他研究警告称，没有适当监督的情况下使用人工智能可能会导致绩效下降。

Top 10 Takeaways (cont'd)
十大要点（续）

8. Scientific progress accelerates even further, thanks to Al. In 2022, Al began to advance scientific discovery. 2023, however, saw the launch of even more significant science-related Al applicationsfrom AlphaDev, which makes algorithmic sorting more efficient, to GNoME, which facilitates the process of materials discovery.
   科学进步得到进一步加速，这要归功于人工智能。2022 年，人工智能开始推动科学发现。然而，2023 年，更多重要的与科学相关的人工智能应用程序推出，从 AlphaDev 推出的使算法排序更加高效，到 GNoME 简化材料发现过程。
9. The number of AI regulations in the United States sharply increases. The number of AIrelated regulations in the U.S. has risen significantly in the past year and over the last five years. In 2023, there were 25 Al-related regulations, up from just one in 2016. Last year alone, the total number of Al-related regulations grew by .
   美国的人工智能法规数量急剧增加。美国与人工智能相关的法规数量在过去一年和过去五年显著增加。2023 年，与人工智能相关的法规数量达到 25 项，远高于 2016 年的 1 项。仅去年一年，与人工智能相关的法规总数增长了 。
10. People across the globe are more cognizant of Al's potential impact-and more nervous. A survey from Ipsos shows that, over the last year, the proportion of those who think Al will dramatically affect their lives in the next three to five years has increased from to . Moreover, express nervousness toward Al products and services, marking a 13 percentage point rise from 2022. In America, Pew data suggests that of Americans report feeling more concerned than excited about Al, rising from 37% in 2022.
    全球各地的人们更加意识到人工智能的潜在影响，并感到更加紧张。Ipsos 的一项调查显示，过去一年中，认为人工智能将在未来三到五年内对他们的生活产生巨大影响的人所占比例已从 增加到 。此外， 对人工智能产品和服务表现出紧张情绪，较 2022 年增加了 13 个百分点。根据皮尤数据，美国有 的美国人表示对人工智能感到更担忧而不是兴奋，较 2022 年的 37%上升。

Steering Committee 指导委员会

Co-directors 联合导演

Members 成员

Staff and Researchers 员工和研究人员

Research Manager and Editor in Chief
研究经理和总编辑

Research Associate 研究助理

Affiliated Researchers 附属研究员

Graduate Researchers 研究生研究员

Undergraduate Researchers
本科研究员

How to Cite This Report
如何引用本报告

Public Data and Tools
公共数据和工具

• Raw data and charts: The public data and high-resolution images of all the charts in the report are available on Google Drive.
  原始数据和图表：报告中所有图表的公共数据和高分辨率图像可在 Google Drive 上获得。
• Global AI Vibrancy Tool: Compare the Al ecosystems of over 30 countries. The Global Al Vibrancy tool will be updated in the summer of 2024.
  全球 AI 活力工具：比较 30 多个国家的 AI 生态系统。全球 AI 活力工具将于 2024 年夏季更新。

Al Index and Stanford HAI
阿尔指数和斯坦福人工智能研究所

Stanford University 斯坦福大学

Supporting Partners 支持合作伙伴

Analytics and Research Partners
分析和研究合作伙伴

accenture 밈은 QCSET 三EPochal Github (govini =

Contributors 贡献者

Introduction 介绍

Chapter 1: Research and Development
第一章：研究与开发

Chapter 2: Technical Performance
第二章：技术表现

Chapter 3: Responsible AI
第三章：负责任的人工智能

Chapter 4: Economy 第四章：经济

Chapter 5: Science and Medicine
第五章：科学与医学

Contributors (cont'd) 贡献者（续）

Chapter 6: Education 第六章：教育

Chapter 7: Policy and Governance
第七章：政策与治理

Chapter 8: Diversity 第八章：多样性

Chapter 9: Public Opinion
第九章：公众舆论

Organizations 组织

Center for Security and Emerging
安全与新兴技术中心

Technology, Georgetown University

Code.org

Computing Research Association
计算机研究协会

Epoch 时代

GitHub

Govini

Informatics Europe

International Federation of Robotics
国际机器人联合会

Lightcast 亮投

Linkedln 领英

McKinsey & Company 麦肯锡公司

Quid

Schwartz Reisman Institute for Technology and Society

Studyportals

Table of Contents 目录

Report Highlights 报告亮点

Chapter 1: Research and Development
第一章：研究与开发

1. Industry continues to dominate frontier Al research. In 2023, industry produced 51 notable machine learning models, while academia contributed only 15 . There were also 21 notable models resulting from industry-academia collaborations in 2023, a new high.
   行业继续主导前沿的人工智能研究。2023 年，行业发布了 51 个显著的机器学习模型，而学术界仅贡献了 15 个。2023 年，行业与学术界合作产生了 21 个显著模型，创下新高。
2. More foundation models and more open foundation models. In 2023 , a total of 149 foundation models were released, more than double the amount released in 2022. Of these newly released models, were open-source, compared to only in 2022 and in 2021.
   发布更多基础模型和开源基础模型。2023 年共发布了 149 个基础模型，比 2022 年发布的数量翻了一倍还多。其中， 个是开源的，而 2022 年只有 个，2021 年更少，只有 个。
3. Frontier models get way more expensive. According to Al Index estimates, the training costs of state-of-the-art AI models have reached unprecedented levels. For example, OpenAl's GPT-4 used an estimated million worth of compute to train, while Google's Gemini Ultra cost million for compute.
   前沿模型变得更加昂贵。根据 Al Index 估计，最先进的人工智能模型的训练成本已经达到了前所未有的水平。例如，OpenAI 的 GPT-4 使用了估计价值 百万的计算资源进行训练，而谷歌的 Gemini Ultra 则花费了 百万用于计算资源。
4. The United States leads China, the EU, and the U.K. as the leading source of top AI models. In 2023, 61 notable Al models originated from U.S.-based institutions, far outpacing the European Union's 21 and China's 15.
   美国领先于中国、欧盟和英国，成为顶尖人工智能模型的主要来源。2023 年，61 个知名的人工智能模型源自美国机构，远远超过欧盟的 21 个和中国的 15 个。
5. The number of Al patents skyrockets. From 2021 to 2022, Al patent grants worldwide increased sharply by . Since 2010 , the number of granted Al patents has increased more than 31 times.
   Al 专利数量激增。从 2021 年到 2022 年，全球 Al 专利授权数量急剧增加 。自 2010 年以来，已授权的 Al 专利数量增加了超过 31 倍。
6. China dominates Al patents. In 2022, China led global Al patent origins with , significantly outpacing the United States, which accounted for of Al patent origins. Since 2010, the U.S. share of AI patents has decreased from .
   中国主导人工智能专利。2022 年，中国在全球人工智能专利来源中处于领先地位，占比 ，远远超过美国，美国占 的人工智能专利来源。自 2010 年以来，美国在人工智能专利中的份额已经从 下降。
7. Open-source Al research explodes. Since 2011, the number of AI-related projects on GitHub has seen a consistent increase, growing from 845 in 2011 to approximately 1.8 million in 2023. Notably, there was a sharp 59.3% rise in the total number of GitHub Al projects in 2023 alone. The total number of stars for Al-related projects on GitHub also significantly increased in 2023, more than tripling from 4.0 million in 2022 to 12.2 million.
   开源人工智能研究蓬勃发展。自 2011 年以来，GitHub 上与人工智能相关项目的数量持续增加，从 2011 年的 845 个增长到 2023 年的约 180 万个。值得注意的是，仅在 2023 年，GitHub 上人工智能项目的总数就急剧增加了 59.3%。2023 年 GitHub 上与人工智能相关项目的星标总数也显著增加，从 2022 年的 400 万增至 1220 万。
8. The number of Al publications continues to rise. Between 2010 and 2022 , the total number of publications nearly tripled, rising from approximately 88,000 in 2010 to more than 240,000 in 2022 . The increase over the last year was a modest .
   人工智能出版物数量持续增长。2010 年至 2022 年间，人工智能出版物总数几乎翻了三番，从 2010 年的约 8.8 万增至 2022 年的超过 24 万。过去一年的增长幅度较为温和，为 。

Report Highlights  报告亮点

Chapter 2: Technical Performance
第二章：技术表现

2. Here comes multimodal Al. Traditionally Al systems have been limited in scope, with language models excelling in text comprehension but faltering in image processing, and vice versa. However, recent advancements have led to the development of strong multimodal models, such as Google's Gemini and OpenAl's GPT-4. These models demonstrate flexibility and are capable of handling images and text and, in some instances, can even process audio.
   传统上，AI 系统在范围上存在局限性，语言模型擅长文本理解，但在图像处理方面表现不佳，反之亦然。然而，最近的进展导致了强大的多模态模型的发展，例如谷歌的 Gemini 和 OpenAI 的 GPT-4。这些模型表现出灵活性，能够处理图像和文本，并且在某些情况下甚至可以处理音频。
3. Harder benchmarks emerge. Al models have reached performance saturation on established benchmarks such as ImageNet, SQuAD, and SuperGLUE, prompting researchers to develop more challenging ones. In 2023, several challenging new benchmarks emerged, including SWE-bench for coding, HEIM for image generation, MMMU for general reasoning, MoCa for moral reasoning, AgentBench for agent-based behavior, and HaluEval for hallucinations.
   更难的基准出现了。AI 模型在 ImageNet、SQuAD 和 SuperGLUE 等已建立的基准上达到了性能饱和，促使研究人员开发出更具挑战性的基准。2023 年，出现了几个具有挑战性的新基准，包括用于编码的 SWE-bench，用于图像生成的 HEIM，用于一般推理的 MMMU，用于道德推理的 MoCa，用于基于代理的行为的 AgentBench，以及用于幻觉的 HaluEval。

4. Better Al means better data which means ... even better Al. New Al models such as
4. 更好的人工智能意味着更好的数据，这意味着...甚至更好的人工智能。新的人工智能模型，如

5. Human evaluation is in. With generative models producing high-quality text, images, and more, benchmarking has slowly started shifting toward incorporating human evaluations like the Chatbot Arena Leaderboard rather than computerized rankings like ImageNet or SQuAD. Public sentiment about AI is becoming an increasingly important consideration in tracking Al progress.
   人类评估正在兴起。随着生成模型生成高质量的文本、图像等，基准测试逐渐开始转向纳入人类评估，如 Chatbot Arena Leaderboard，而不是像 ImageNet 或 SQuAD 这样的计算机排名。公众对人工智能的情绪变得越来越重要，成为追踪人工智能进展的一个日益重要的考虑因素。
6. Thanks to LLMs, robots have become more flexible. The fusion of language modeling with robotics has given rise to more flexible robotic systems like PaLM-E and RT-2. Beyond their improved robotic capabilities, these models can ask questions, which marks a significant step toward robots that can interact more effectively with the real world.
   由于LLMs的出现，机器人变得更加灵活。语言建模与机器人技术的融合催生了更加灵活的机器人系统，如 PaLM-E 和 RT-2。除了改进了机器人的能力，这些模型还可以提问，这是机器人与现实世界更有效互动的重要一步。

Chapter 2: Technical Performance (cont'd)
第二章：技术性能（续）

7. More technical research in agentic Al. Creating Al agents, systems capable of autonomous operation in specific environments, has long challenged computer scientists. However, emerging research suggests that the performance of autonomous agents is improving. Current agents can now master complex games like Minecraft and effectively tackle real-world tasks, such as online shopping and research assistance.
   关于主动型人工智能的更多技术研究。创建自主在特定环境中运行的人工智能代理系统一直是计算机科学家们长期面临的挑战。然而，新兴的研究表明，自主 代理系统的性能正在不断提高。现如今的代理系统已经能够掌握复杂的游戏，如 Minecraft，并且能够有效地处理实际任务，如在线购物和研究辅助。
8. Closed LLMs significantly outperform open ones. On 10 select Al benchmarks, closed models outperformed open ones, with a median performance advantage of . Differences in the performance of closed and open models carry important implications for Al policy debates.
   闭源LLMs明显优于开源。在 10 个选择的 Al 基准测试中，闭源模型的性能优于开源模型，具有 的中位性能优势。闭源模型和开源模型的性能差异对 Al 政策辩论具有重要意义。

Report Highlights 报告亮点

Chapter 3: Responsible Al
第三章：负责任的 Al

1. Robust and standardized evaluations for LLM responsibility are seriously lacking.
1.对于LLM责任的稳健和标准化评估严重缺乏。

New research from the AI Index reveals a significant lack of standardization in responsible AI reporting. Leading developers, including OpenAI, Google, and Anthropic, primarily test their models against different responsible AI benchmarks. This practice complicates efforts to systematically compare the risks and limitations of top Al models.

2. Political deepfakes are easy to generate and difficult to detect. Political deepfakes are already affecting elections across the world, with recent research suggesting that existing Al deepfake methods perform with varying levels of accuracy. In addition, new projects like CounterCloud demonstrate how easily AI can create and disseminate fake content.
   政治 Deepfake 很容易生成，难以检测。最近的研究表明，政治 Deepfake 已经对世界各地的选举产生影响，现有的 Al Deepfake 方法的准确性不同。此外，CounterCloud 等新项目展示了 AI 如何轻松创建和传播假内容。
3. Researchers discover more complex vulnerabilities in LLMs. Previously, most efforts to red team Al models focused on testing adversarial prompts that intuitively made sense to humans. This year, researchers found less obvious strategies to get LLMs to exhibit harmful behavior, like asking the models to infinitely repeat random words.
   研究人员在LLMs中发现了更复杂的漏洞。以前，对 Al 模型进行红队测试的大部分工作都集中在测试对人类直观有意义的对抗性提示上。今年，研究人员发现了一些不太明显的策略，可以使LLMs表现出有害行为，例如要求模型无限重复随机单词。
4. Risks from are becoming a concern for businesses across the globe. A global survey on responsible Al highlights that companies' top Al-related concerns include privacy, data security, and reliability. The survey shows that organizations are beginning to take steps to mitigate these risks. Globally, however, most companies have so far only mitigated a small portion of these risks.
   来自 的风险正成为全球企业关注的焦点。一项关于负责任人工智能的全球调查显示，公司对人工智能相关的主要担忧包括隐私、数据安全和可靠性。调查显示，组织正开始采取措施来减轻这些风险。然而，在全球范围内，大多数公司迄今只减轻了这些风险的一小部分。
5. LLMs can output copyrighted material. Multiple researchers have shown that the generative outputs of popular LLMs may contain copyrighted material, such as excerpts from The New York Times or scenes from movies. Whether such output constitutes copyright violations is becoming a central legal question.
   LLMs可以输出受版权保护的材料。多位研究人员已经证明，流行LLMs的生成输出可能包含受版权保护的材料，例如《纽约时报》的摘录或电影场景。这种输出是否构成侵犯版权的问题正逐渐成为一个核心的法律问题。

6. Al developers score low on transparency, with consequences for research. The newly
6. 人工智能开发者在透明度方面得分较低，对研究产生影响。新近进行的研究显示，人工智能开发者在透明度方面得分较低，这对研究产生了影响。

introduced Foundation Model Transparency Index shows that AI developers lack transparency, especially regarding the disclosure of training data and methodologies. This lack of openness hinders efforts to further understand the robustness and safety of Al systems.

Chapter 3: Responsible AI (cont'd)
第三章:负责任的人工智能(续)

7. Extreme Al risks are difficult to analyze. Over the past year, a substantial debate has emerged among Al scholars and practitioners regarding the focus on immediate model risks, like algorithmic discrimination, versus potential long-term existential threats. It has become challenging to distinguish which claims are scientifically founded and should inform policymaking. This difficulty is compounded by the tangible nature of already present short-term risks in contrast with the theoretical nature of existential threats.
   极端 AI 风险很难分析。在过去的一年中，AI 学者和从业者之间出现了一场关于立即模型风险（如算法歧视）与潜在的长期存在威胁的辩论。区分哪些声明在科学上有根据并应该影响政策制定已变得具有挑战性。这一困难被已经存在的短期风险的具体性质与存在威胁的理论性质所加剧。
8. The number of Al incidents continues to rise. According to the Al Incident Database, which tracks incidents related to the misuse of Al, 123 incidents were reported in 2023, a 32.3 percentage point increase from 2022. Since 2013, Al incidents have grown by over twentyfold. A notable example includes Al-generated, sexually explicit deepfakes of Taylor Swift that were widely shared online.
   AI 事件的数量继续增加。根据 AI 事件数据库的数据，追踪与 AI 滥用相关的事件，2023 年报告了 123 起事件，比 2022 年增长了 32.3 个百分点。自 2013 年以来，AI 事件增长了超过二十倍。一个显著的例子是泰勒·斯威夫特的 AI 生成的深度假视频，在网上广为传播。
9. ChatGPT is politically biased. Researchers find a significant bias in ChatGPT toward Democrats in the United States and the Labour Party in the U.K. This finding raises concerns about the tool's potential to influence users' political views, particularly in a year marked by major global elections.
   ChatGPT 在政治上存在偏见。研究人员发现 ChatGPT 在美国偏向民主党和英国工党。这一发现引发了对该工具可能影响用户政治观点的担忧，尤其是在一个以重大全球选举为特征的年份。

Report Highlights 报告亮点

Chapter 4: Economy 第四章：经济

1. Generative investment skyrockets. Despite a decline in overall Al private investment last year, funding for generative surged, nearly octupling from 2022 to reach billion. Major players in the generative Al space, including OpenAl, Anthropic, Hugging Face, and Inflection, reported substantial fundraising rounds.
   生成 投资激增。尽管去年整体 AI 私人投资有所下降，但生成 的资金激增，从 2022 年几乎增加了八倍，达到 亿美元。生成 AI 领域的主要参与者，包括 OpenAI、Anthropic、Hugging Face 和 Inflection，报告了大量的筹款轮。
2. Already a leader, the United States pulls even further ahead in Al private investment. In 2023, the United States saw Al investments reach billion, nearly 8.7 times more than China, the next highest investor. While private Al investment in China and the European Union, including the United Kingdom, declined by and , respectively, since 2022, the United States experienced a notable increase of in the same time frame.
   美国已经是领导者，在人工智能私人投资方面进一步拉开了差距。2023 年，美国的人工智能投资达到 亿美元，几乎是中国的近 8.7 倍，后者是第二大投资国。尽管自 2022 年以来，中国和欧盟（包括英国）的私人人工智能投资分别下降了 和 ，但同期美国经历了 的显著增长。
3. Fewer Al jobs in the United States and across the globe. In 2022, Al-related positions made up of all job postings in America, a figure that decreased to in 2023. This decline in Al job listings is attributed to fewer postings from leading Al firms and a reduced proportion of tech roles within these companies.
   美国和全球范围内的人工智能工作岗位减少。2022 年，与美国所有职位招聘相比，与人工智能相关的职位占比为 ，这一数字在 2023 年降至 。人工智能工作岗位的减少归因于领先人工智能公司发布的职位减少以及这些公司中技术角色比例的降低。
4. Al decreases costs and increases revenues. A new McKinsey survey reveals that of surveyed organizations report cost reductions from implementing (including generative ), and report revenue increases. Compared to the previous year, there was a 10 percentage point increase in respondents reporting decreased costs, suggesting is driving significant business efficiency gains.
   人工智能降低成本并增加收入。麦肯锡的一项新调查显示， 的受访组织表示通过实施 （包括生成式 ）实现了成本降低， 表示收入增加。与前一年相比，报告成本降低的受访者比例增加了 10 个百分点，表明 正在推动重大的商业效率提升。

5. Total Al private investment declines again, while the number of newly funded AI
5. 总体上，私人投资再次下降，而新资助的 AI 数量增加。

companies increases. Global private Al investment has fallen for the second year in a row, though less than the sharp decrease from 2021 to 2022 . The count of newly funded Al companies spiked to 1,812 , up from the previous year.

6. Al organizational adoption ticks up. A 2023 McKinsey report reveals that of organizations now use (including generative ) in at least one business unit or function, up from in 2022 and in 2017.
   AI 组织采用率有所提高。2023 年麦肯锡报告显示， 的组织现在至少在一个业务部门或功能中使用 （包括生成 ），高于 2022 年的 和 2017 年的 。
7. China dominates industrial robotics. Since surpassing Japan in 2013 as the leading installer of industrial robots, China has significantly widened the gap with the nearest competitor nation. In 2013, China's installations accounted for of the global total, a share that rose to by 2022 .
   中国主导工业机器人。自 2013 年超过日本成为工业机器人领先安装国以来，中国与最近的竞争对手国家之间的差距显著扩大。2013 年，中国的安装量占全球总量的 ，到 2022 年已上升至 。

Chapter 4: Economy (cont'd)
第四章：经济（续）

8. Greater diversity in robot installations. In 2017, collaborative robots represented a mere of all new industrial robot installations, a figure that climbed to by 2022 . Similarly, 2022 saw a rise in service robot installations across all application categories, except for medical robotics. This trend indicates not just an overall increase in robot installations but also a growing emphasis on deploying robots for human-facing roles.
   机器人安装的多样性增加。2017 年，协作机器人仅占所有新工业机器人安装量的 ，到 2022 年这一数字上升到 。同样，2022 年各应用类别的服务机器人安装量均有所增加，除了医疗机器人。这一趋势不仅表明机器人安装总量增加，还表明人类面向角色的机器人部署受到越来越多的重视。
9. The data is in: Al makes workers more productive and leads to higher quality work. In 2023, several studies assessed Al's impact on labor, suggesting that Al enables workers to complete tasks more quickly and to improve the quality of their output. These studies also demonstrated Al's potential to bridge the skill gap between low- and high-skilled workers. Still, other studies caution that using Al without proper oversight can lead to diminished performance.
   数据显示：人工智能提高了工人的生产力，促进了更高质量的工作。2023 年，几项研究评估了人工智能对劳动力的影响，表明人工智能使工人能够更快地完成任务并提高其产出质量。这些研究还展示了人工智能在弥合低技能和高技能工人之间的技能差距方面的潜力。然而，其他研究警告称，没有适当监督的情况下使用人工智能可能会导致绩效下降。
10. Fortune 500 companies start talking a lot about Al, especially generative Al. In 2023, Al was mentioned in 394 earnings calls (nearly of all Fortune 500 companies), a notable increase from 266 mentions in 2022. Since 2018, mentions of Al in Fortune 500 earnings calls have nearly doubled. The most frequently cited theme, appearing in of all earnings calls, was generative Al.
    《财富》500 强公司开始大谈人工智能，尤其是生成式人工智能。2023 年，394 家公司在财报电话会议中提到了人工智能（几乎占所有《财富》500 强公司的 ），较 2022 年的 266 次提及有显著增加。自 2018 年以来，在《财富》500 强公司的财报电话会议中提到人工智能的次数几乎翻了一番。最经常被提及的主题，出现在 的所有财报电话会议中，是生成式人工智能。

Report Highlights 报告亮点

Chapter 5: Science and Medicine
第五章：科学与医学

1. Scientific progress accelerates even further, thanks to AI. In 2022, Al began to advance scientific discovery. 2023, however, saw the launch of even more significant science-related Al applicationsfrom AlphaDev, which makes algorithmic sorting more efficient, to GNoME, which facilitates the process of materials discovery.
   由于人工智能的推动，科学进步进一步加速。2022 年，人工智能开始推动科学发现。然而，2023 年看到了更多重要的与科学相关的人工智能应用的推出，从使算法排序更加高效的 AlphaDev，到促进材料发现过程的 GNoME。
2. Al helps medicine take significant strides forward. In 2023, several significant medical systems were launched, including EVEscape, which enhances pandemic prediction, and AlphaMissence, which assists in Al-driven mutation classification. is increasingly being utilized to propel medical advancements.
   AI 帮助医学取得了重大进展。2023 年，推出了几款重要的医疗系统，包括增强疫情预测的 EVEscape 和协助基于 AI 的突变分类的 AlphaMissence。 越来越多地被用于推动医学进步。
3. Highly knowledgeable medical Al has arrived. Over the past few years, Al systems have shown remarkable improvement on the MedQA benchmark, a key test for assessing Al's clinical knowledge. The standout model of 2023 , GPT-4 Medprompt, reached an accuracy rate of , marking a 22.6 percentage point increase from the highest score in 2022. Since the benchmark's introduction in 2019, Al performance on MedQA has nearly tripled.
   高度了解医学的 AI 已经到来。在过去几年里，AI 系统在 MedQA 基准测试上显示出了显著的改进，这是评估 AI 临床知识的关键测试。2023 年的杰出模型 GPT-4 Medprompt 达到了 的准确率，比 2022 年最高分提高了 22.6 个百分点。自 2019 年引入基准测试以来，AI 在 MedQA 上的表现几乎翻了三倍。
4. The FDA approves more and more Al-related medical devices. In 2022, the FDA approved 139 Al-related medical devices, a 12.1% increase from 2021. Since 2012, the number of FDA-approved Al-related medical devices has increased by more than 45 -fold. is increasingly being used for real-world medical purposes.
   FDA 批准了越来越多与 AI 相关的医疗设备。2022 年，FDA 批准了 139 款与 AI 相关的医疗设备，比 2021 年增加了 12.1%。自 2012 年以来，FDA 批准的与 AI 相关的医疗设备数量增加了超过 45 倍。 越来越多地被用于真实世界的医疗用途。

Report Highlights 报告亮点

Chapter 6: Education 第六章：教育

1. The number of American and Canadian CS bachelor's graduates continues to rise, new CS master's graduates stay relatively flat, and PhD graduates modestly grow. While the number of new American and Canadian bachelor's graduates has consistently risen for more than a decade, the number of students opting for graduate education in CS has flattened. Since 2018, the number of CS master's and graduates has slightly declined.
   美国和加拿大计算机科学学士毕业生人数持续增长，新的计算机科学硕士毕业生人数保持相对稳定，博士毕业生逐渐增加。虽然美国和加拿大新毕业的学士生人数在过去十多年一直在增加，但选择攻读计算机科学研究生教育的学生人数已经趋于稳定。自 2018 年以来，计算机科学硕士和 毕业生人数略有下降。
2. The migration of Al PhDs to industry continues at an accelerating pace. In 2011, roughly equal percentages of new AI PhDs took jobs in industry (40.9%) and academia (41.6%). However, by 2022, a significantly larger proportion (70.7%) joined industry after graduation compared to those entering academia (20.0%). Over the past year alone, the share of industry-bound AI PhDs has risen by 5.3 percentage points, indicating an intensifying brain drain from universities into industry.
   人工智能博士毕业生向产业界的迁移速度正在加快。2011 年，新的人工智能博士毕业生在产业界（40.9%）和学术界（41.6%）的就业比例大致相等。然而，到 2022 年，毕业后加入产业界的比例显著增加至 70.7%，远远超过进入学术界的比例（20.0%）。仅在过去一年中，进入产业界的人工智能博士毕业生的比例增加了 5.3 个百分点，表明大学向产业界的人才流失正在加剧。
3. Less transition of academic talent from industry to academia. In of new Al faculty in the United States and Canada were from industry. By 2021, this figure had declined to 11%, and in 2022, it further dropped to . This trend indicates a progressively lower migration of high-level AI talent from industry into academia.
   行业向学术界的学术人才转移减少。在美国和加拿大的新人工智能教职员中， 都来自行业。到 2021 年，这一比例下降至 11％，并在 2022 年进一步降至 。这一趋势表明，高水平人工智能人才从行业流向学术界的迁移越来越少。
4. CS education in the United States and Canada becomes less international. Proportionally fewer international CS bachelor's, master's, and PhDs graduated in 2022 than in 2021. The drop in international students in the master's category was especially pronounced.
   美国和加拿大的计算机科学教育变得不那么国际化。2022 年，获得计算机科学学士、硕士和博士学位的国际学生比例较 2021 年有所下降。硕士类别中国际学生的减少尤为显著。

5. More American high school students take CS courses, but access problems remain.
5. 更多的美国高中学生选修计算机科学课程，但仍存在接触问题。

In 2022, 201,000 AP CS exams were administered. Since 2007, the number of students taking these exams has increased more than tenfold. However, recent evidence indicates that students in larger high schools and those in suburban areas are more likely to have access to CS courses.

6. Al-related degree programs are on the rise internationally. The number of English-language, Al-related postsecondary degree programs has tripled since 2017, showing a steady annual increase over the past five years. Universities worldwide are offering more Al-focused degree programs.
   人工智能相关的学位课程在国际上不断增加。自 2017 年以来，英语授课的人工智能相关高等教育学位课程数量增加了三倍，过去五年稳步增长。全球各大学正在提供更多以人工智能为重点的学位课程。

Chapter 6: Education (cont'd)
第 6 章：教育（续）

7. The United Kingdom and Germany lead in European informatics, CS, CE, and IT graduate production. The United Kingdom and Germany lead Europe in producing the highest number of new informatics, CS, CE, and information bachelor's, master's, and PhD graduates. On a per capita basis, Finland leads in the production of both bachelor's and PhD graduates, while Ireland leads in the production of master's graduates.
   英国和德国在欧洲信息学、计算机科学、计算机工程和信息技术研究生产量方面处于领先地位。英国和德国在生产新的信息学、计算机科学、计算机工程和信息学学士、硕士和博士毕业生数量方面领先于欧洲。以人均计算，芬兰在生产学士和博士毕业生方面领先，而爱尔兰在生产硕士毕业生方面领先。

Report Highlights 报告亮点

Chapter 7: Policy and Governance
第七章：政策与治理

1. The number of Al regulations in the United States sharply increases. The number of Al-related regulations has risen significantly in the past year and over the last five years. In 2023 , there were -related regulations, up from just one in 2016. Last year alone, the total number of Al-related regulations grew by .
   美国的 Al 法规数量急剧增加。过去一年和过去五年，与 Al 相关的法规数量显著增加。2023 年，与 Al 相关的法规数量达到 个，远远超过 2016 年的仅有一个。仅去年一年，与 Al 相关的法规总数增长了 个。
2. The United States and the European Union advance landmark Al policy action. In 2023, policymakers on both sides of the Atlantic put forth substantial proposals for advancing Al regulation The European Union reached a deal on the terms of the AI Act, a landmark piece of legislation enacted in 2024. Meanwhile, President Biden signed an Executive Order on Al, the most notable Al policy initiative in the United States that year.
   美国和欧盟推进具有里程碑意义的人工智能政策行动。2023 年，大西洋两岸的决策者提出了推进人工智能监管的实质性提案。欧盟就《人工智能法案》的条款达成协议，这是 2024 年出台的具有里程碑意义的立法。与此同时，拜登总统签署了一项关于人工智能的行政命令，这是美国在那一年最显著的人工智能政策倡议。
3. Al captures U.S. policymaker attention. The year 2023 witnessed a remarkable increase in Al-related legislation at the federal level, with 181 bills proposed, more than double the 88 proposed in 2022.
   人工智能引起美国决策者的关注。2023 年，联邦层面关于人工智能立法的提案大幅增加，共有 181 项提出，是 2022 年提出的 88 项的两倍多。
4. Policymakers across the globe cannot stop talking about Al. Mentions of Al in legislative proceedings across the globe have nearly doubled, rising from 1,247 in 2022 to 2,175 in 2023. Al was mentioned in the legislative proceedings of 49 countries in 2023 . Moreover, at least one country from every continent discussed Al in 2023, underscoring the truly global reach of Al policy discourse.
   全球各地的决策者们谈论不休人工智能。全球立法程序中关于人工智能的提及几乎翻了一番，从 2022 年的 1,247 次上升至 2023 年的 2,175 次。 2023 年，有 49 个国家的立法程序提及了人工智能。此外，至少来自每个大洲的一个国家在 2023 年讨论了人工智能，凸显了人工智能政策议题的真正全球影响力。
5. More regulatory agencies turn their attention toward AI. The number of U.S. regulatory agencies issuing Al regulations increased to 21 in 2023 from 17 in 2022, indicating a growing concern over Al regulation among a broader array of American regulatory bodies. Some of the new regulatory agencies that enacted Alrelated regulations for the first time in 2023 include the Department of Transportation, the Department of Energy, and the Occupational Safety and Health Administration.
   更多监管机构将注意力转向人工智能。2023 年，发布 Al 法规的美国监管机构数量增至 21 家，而 2022 年为 17 家，表明越来越多的美国监管机构对 Al 监管表示关注。2023 年首次颁布与 Al 相关法规的一些新监管机构包括交通部、能源部和职业安全与健康管理局。

Report Highlights 报告亮点

Chapter 8: Diversity 第八章：多样性

1. U.S. and Canadian bachelor's, master's, and PhD CS students continue to grow more ethnically diverse. While white students continue to be the most represented ethnicity among new resident graduates at all three levels, the representation from other ethnic groups, such as Asian, Hispanic, and Black or African American students, continues to grow. For instance, since 2011, the proportion of Asian CS bachelor's degree graduates has increased by 19.8 percentage points, and the proportion of Hispanic CS bachelor's degree graduates has grown by 5.2 percentage points.
   美国和加拿大的计算机科学学士、硕士和博士研究生在族裔多样性上持续增长。白人学生在这三个层次的新入学毕业生中仍然是最多的族裔代表，而其他少数民族群体的代表，如亚洲人、西班牙裔和非洲裔学生的比例则继续增长。例如，自 2011 年以来，亚洲计算机科学学士学位毕业生的比例增加了 19.8 个百分点，西班牙裔计算机科学学士学位毕业生的比例增长了 5.2 个百分点。
2. Substantial gender gaps persist in European informatics, CS, CE, and IT graduates at all educational levels. Every surveyed European country reported more male than female graduates in bachelor's, master's, and PhD programs for informatics, CS, CE, and IT. While the gender gaps have narrowed in most countries over the last decade, the rate of this narrowing has been slow.
   欧洲在信息学、计算机科学、计算机工程和信息技术领域的女性毕业生普遍存在较大的性别差距。每个受调查的欧洲国家在信息学、计算机科学、计算机工程和信息技术的学士、硕士和博士项目中都报告了男性毕业生的人数多于女性。虽然在过去十年中，各国间的性别差距有所缩小，但这种缩小的速度很慢。
3. U.S. K-12 CS education is growing more diverse, reflecting changes in both gender and ethnic representation. The proportion of AP CS exams taken by female students rose from in 2007 to 30.5% in 2022. Similarly, the participation of Asian, Hispanic/Latino/Latina, and Black/African American students in AP CS has consistently increased year over year.
   美国 K-12 计算机科学教育正在变得更加多样化，反映出性别和种族代表性的变化。女性学生参加 AP 计算机科学考试的比例从 2007 年的 上升到 2022 年的 30.5%。同样，亚洲、西班牙裔/拉丁裔和非洲裔美国学生参加 AP 计算机科学考试的比例也在逐年增加。

Report Highlights 报告亮点

Chapter 9: Public Opinion
第九章：公众舆论

1. People across the globe are more cognizant of Al's potential impact-and more nervous. A survey from Ipsos shows that, over the last year, the proportion of those who think AI will dramatically affect their lives in the next three to five years has increased from to . Moreover, express nervousness toward Al products and services, marking a 13 percentage point rise from 2022. In America, Pew data suggests that of Americans report feeling more concerned than excited about Al, rising from in 2022.
   环球范围内的人们对人工智能的潜在影响更加明显，也更加紧张。Ipsos 的一项调查显示，在过去一年中，认为人工智能在未来三到五年内将对他们的生活产生巨大影响的人的比例从 上升到 。此外， 对人工智能产品和服务表达出紧张情绪，与 2022 年相比上升了 13 个百分点。在美国，Pew 的数据显示， 的美国人对人工智能感到更多的担忧而不是兴奋，从 2022 年的 上升到现在。
2. Al sentiment in Western nations continues to be low, but is slowly improving. In 2022, several developed Western nations, including Germany, the Netherlands, Australia, Belgium, Canada, and the United States, were among the least positive about Al products and services. Since then, each of these countries has seen a rise in the proportion of respondents acknowledging the benefits of Al, with the Netherlands experiencing the most significant shift.
   西方国家的情绪仍然低迷，但正在慢慢改善。2022 年，包括德国、荷兰、澳大利亚、比利时、加拿大和美国在内的几个发达的西方国家对人工智能产品和服务持最少的积极态度。自那时起，这些国家中的每一个都看到了承认人工智能好处的受访者比例上升，其中荷兰经历了最显著的变化。
3. The public is pessimistic about Al's economic impact. In an Ipsos survey, only of respondents feel will improve their job. Only anticipate will boost the economy, and believe it will enhance the job market.
   公众对人工智能的经济影响持悲观态度。在一项 Ipsos 调查中，只有 的受访者认为 会改善他们的工作。只有 预期 会促进经济增长， 相信它会提升就业市场。

4. Demographic differences emerge regarding Al optimism. Significant demographic
4. 关于人工智能乐观情绪存在人口统计学差异。显著的人口统计学差异 emerge regarding Al optimism.

differences exist in perceptions of Al's potential to enhance livelihoods, with younger generations generally more optimistic. For instance, of Gen respondents believe will improve entertainment options, versus only of baby boomers. Additionally, individuals with higher incomes and education levels are more optimistic about Al's positive impacts on entertainment, health, and the economy than their lower-income and less-educated counterparts.

5. ChatGPT is widely known and widely used. An international survey from the University of Toronto suggests that of respondents are aware of ChatGPT. Of those aware, around half report using ChatGPT at least once weekly.
   ChatGPT 被广泛认知和广泛使用。据多伦多大学的一项国际调查显示，受访者中有 知道 ChatGPT。而在知道 ChatGPT 的人中，约一半表示每周至少使用一次 ChatGPT。

Preview 预览

Chapter Highlights 章节亮点

1. Industry continues to dominate frontier Al research. In 2023, industry produced 51 notable machine learning models, while academia contributed only 15 . There were also 21 notable models resulting from industry-academia collaborations in 2023, a new high.
   行业继续主导前沿的人工智能研究。2023 年，行业发布了 51 个显著的机器学习模型，而学术界仅贡献了 15 个。2023 年，行业与学术界合作产生了 21 个显著模型，创下新高。
2. More foundation models and more open foundation models. In 2023, a total of 149 foundation models were released, more than double the amount released in 2022. Of these newly released models, were open-source, compared to only in 2022 and in 2021.
   更多的基础模型和更多的开源基础模型。2023 年，共发布了 149 个基础模型，是 2022 年发布数量的两倍以上。在这些新发布的模型中， 个是开源的，而 2022 年只有 个，2021 年只有 个。
3. Frontier models get way more expensive. According to Al Index estimates, the training costs of state-of-the-art Al models have reached unprecedented levels. For example, OpenAl's GPT-4 used an estimated million worth of compute to train, while Google's Gemini Ultra cost million for compute.
   边界模型变得更加昂贵。根据 Al Index 估计，最先进的 Al 模型的训练成本已经达到了前所未有的水平。例如，OpenAl 的 GPT-4 使用了估计价值 百万的计算资源进行训练，而谷歌的 Gemini Ultra 则花费了 百万用于计算资源。
4. The United States leads China, the EU, and the U.K. as the leading source of top AI models. In 2023, 61 notable Al models originated from U.S.-based institutions, far outpacing the European Union's 21 and China's 15.
   美国领先于中国、欧盟和英国，成为顶尖人工智能模型的主要来源。2023 年，61 个知名的人工智能模型源自美国机构，远远超过欧盟的 21 个和中国的 15 个。
5. The number of Al patents skyrockets. From 2021 to 2022, Al patent grants worldwide increased sharply by . Since 2010 , the number of granted Al patents has increased more than 31 times.
   Al 专利数量激增。从 2021 年到 2022 年，全球 Al 专利授权数量急剧增加 。自 2010 年以来，已授权的 Al 专利数量增加了超过 31 倍。
6. China dominates Al patents. In 2022, China led global Al patent origins with , significantly outpacing the United States, which accounted for of Al patent origins. Since 2010, the U.S. share of AI patents has decreased from .
   中国主导人工智能专利。2022 年，中国在全球人工智能专利来源中处于领先地位，占比 ，远远超过美国，美国占 的人工智能专利来源。自 2010 年以来，美国在人工智能专利中的份额已经从 下降。
7. Open-source Al research explodes. Since 2011, the number of AI-related projects on GitHub has seen a consistent increase, growing from 845 in 2011 to approximately 1.8 million in 2023. Notably, there was a sharp 59.3% rise in the total number of GitHub Al projects in 2023 alone. The total number of stars for Al-related projects on GitHub also significantly increased in 2023, more than tripling from 4.0 million in 2022 to 12.2 million.
   开源人工智能研究蓬勃发展。自 2011 年以来，GitHub 上与人工智能相关项目的数量持续增加，从 2011 年的 845 个增长到 2023 年的约 180 万个。值得注意的是，仅在 2023 年，GitHub 上人工智能项目的总数就急剧增加了 59.3%。2023 年 GitHub 上与人工智能相关项目的星标总数也显著增加，从 2022 年的 400 万增至 1220 万。
8. The number of Al publications continues to rise. Between 2010 and 2022 , the total number of publications nearly tripled, rising from approximately 88,000 in 2010 to more than 240,000 in 2022 . The increase over the last year was a modest .
   人工智能出版物数量持续增长。2010 年至 2022 年间，人工智能出版物总数几乎翻了三番，从 2010 年的约 8.8 万增至 2022 年的超过 24 万。过去一年的增长幅度较为温和，为 。

1.1 Publications 1.1 出版物

Overview 概述

Total Number of AI Publications
AI 出版物的总数

By Type of Publication
按出版物类型

By Field of Study
按研究领域

By Sector 按行业

Al publications (% of total) by sector and geographic area, 2022
2022 年按部门和地理区域划分的出版物（总量的百分比）

Al Journal Publications Al 期刊出版物

Al Conference Publications
人工智能会议出版物

1.2 Patents 1.2 专利

Al Patents 所有专利

Overview 概述

By Filing Status and Region
按申请状态和地区

Al patents by application status,
申请状态的 Al 专利，

Al patents by application status by geographic area, 2010-22
2010 年至 2022 年各地区按申请状态划分的 Al 专利

Granted Al patents (% of world total) by region, 2010-22
按地区划分，2010-22 年各地区授予的 Al 专利（占世界总量的比例）

1.3 Frontier Al Research
1.3 前沿 Al 研究

Overview 概述

Sector Analysis 领域分析

National Affiliation 国家隶属

Number of notable machine learning models by select geographic area, 2003-23
2003 年至 2023 年各地区知名机器学习模型数量

Parameter Trends 参数趋势

Compute Trends 计算趋势

Will Models Run Out of Data?
模型会用尽数据吗？

Projections of ML data exhaustion by stock type:
ML 数据枯竭的预测按库存类型：

median and dates

Highlight: 突出显示:

Will Models Run Out of Data? (cont'd)
模型会耗尽数据吗？（续）

A demonstration of model collapse in a VAE
VAE 中模型坍缩的演示

Highlight: 突出显示:

Will Models Run Out of Data? (cont'd)
模型会耗尽数据吗？（续）

Will Models Run Out of Data? (cont'd)
模型会耗尽数据吗？（续）

Foundation Models 基础模型

Model Release 模型发布

Organizational Affiliation
组织隶属

Number of foundation models by organization, 2019-23 (sum)
组织发布的基础模型数量，2019-23 年（总和）

National Affiliation 国家隶属

Training Cost 训练成本

1.4 Al Conferences 1.4 Al 会议

Conference Attendance 会议出席

Attendance at large conferences, 2010-23
大型会议的参会人数，2010-23

Attendance at small conferences, 2010-23
2010-23 年小型会议出席情况

1.5 Open-Source Al Software
1.5 开源人工智能软件

Projects 项目

Stars 星星

Number of GitHub stars in Al projects, 2011-23
2011-23 年 Al 项目在 GitHub 上的星标数量

Preview 预览

Preview (cont'd) 预览（续）

2.11 Properties of LLMs
2.11 LLMs的属性

..... 141

2.12 Techniques for LLM Improvement
148
2.12 技术的 LLM 改进 148

ACCESS THE PUBLIC DATA
访问公共数据

Chapter Highlights 章节亮点

1. Al beats humans on some tasks, but not on all. Al has surpassed human performance on several benchmarks, including some in image classification, visual reasoning, and English understanding. Yet it trails behind on more complex tasks like competition-level mathematics, visual commonsense reasoning and planning.
   在某些任务上，人工智能超过人类，但并非所有任务都是如此。人工智能在一些基准测试中已经超过了人类的表现，包括图像分类、视觉推理和英语理解等方面。然而，在更复杂的任务上，如竞赛级数学、视觉常识推理和规划方面，人工智能仍然落后。
2. Here comes multimodal AI. Traditionally Al systems have been limited in scope, with language models excelling in text comprehension but faltering in image processing, and vice versa. However, recent advancements have led to the development of strong multimodal models, such as Google's Gemini and OpenAl's GPT-4. These models demonstrate flexibility and are capable of handling images and text and, in some instances, can even process audio.
   在这里有多模态人工智能。传统上，人工智能系统在范围上是受限的，语言模型在文本理解方面表现出色，但在图像处理方面表现不足，反之亦然。然而，最近的进展导致了强大的多模态模型的发展，例如 Google 的 Gemini 和 OpenAl 的 GPT-4。这些模型展示了灵活性，能够处理图像和文本，有时甚至可以处理音频。
3. Harder benchmarks emerge. Al models have reached performance saturation on established benchmarks such as ImageNet, SQuAD, and SuperGLUE, prompting researchers to develop more challenging ones. In 2023, several challenging new benchmarks emerged, including SWE-bench for coding, HEIM for image generation, MMMU for general reasoning, MoCa for moral reasoning, AgentBench for agent-based behavior, and HaluEval for hallucinations.
   更难的基准出现了。AI 模型在 ImageNet、SQuAD 和 SuperGLUE 等已建立的基准上达到了性能饱和，促使研究人员开发出更具挑战性的基准。2023 年，出现了几个具有挑战性的新基准，包括用于编码的 SWE-bench，用于图像生成的 HEIM，用于一般推理的 MMMU，用于道德推理的 MoCa，用于基于代理的行为的 AgentBench，以及用于幻觉的 HaluEval。
4. Better Al means better data which means ... even better Al. New Al models such as SegmentAnything and Skoltech are being used to generate specialized data for tasks like image segmentation and 3D reconstruction. Data is vital for Al technical improvements. The use of to create more data enhances current capabilities and paves the way for future algorithmic improvements, especially on harder tasks.
   更好的人工智能意味着更好的数据，这意味着…甚至更好的人工智能。新的人工智能模型如 SegmentAnything 和 Skoltech 正在被用于生成专业化的数据，用于任务如图像分割和 3D 重建。数据对于人工智能的技术改进至关重要。使用 来生成更多的数据增强了当前的能力，并为未来的算法改进铺平了道路，尤其是在更难的任务上。
5. Human evaluation is in. With generative models producing high-quality text, images, and more, benchmarking has slowly started shifting toward incorporating human evaluations like the Chatbot Arena Leaderboard rather than computerized rankings like ImageNet or SQuAD. Public feeling about Al is becoming an increasingly important consideration in tracking Al progress.
   人类评估已经开始。随着生成模型生成高质量的文本、图像等，基准测试已经开始慢慢转向纳入人类评估，比如 Chatbot Arena 排行榜，而不是像 ImageNet 或 SQuAD 那样的计算机排名。公众对人工智能的看法正在成为追踪人工智能进展的日益重要的考虑因素。
6. Thanks to LLMs, robots have become more flexible. The fusion of language modeling with robotics has given rise to more flexible robotic systems like PaLM-E and RT-2. Beyond their improved robotic capabilities, these models can ask questions, which marks a significant step toward robots that can interact more effectively with the real world.
   由于LLMs的出现，机器人变得更加灵活。语言建模与机器人技术的融合催生了更加灵活的机器人系统，如 PaLM-E 和 RT-2。除了改进了机器人的能力，这些模型还可以提问，这是机器人与现实世界更有效互动的重要一步。
7. More technical research in agentic Al. Creating Al agents, systems capable of autonomous operation in specific environments, has long challenged computer scientists. However, emerging research suggests that the performance of autonomous agents is improving. Current agents can now master complex games like Minecraft and effectively tackle real-world tasks, such as online shopping and research assistance.
   关于主动型人工智能的更多技术研究。创建自主在特定环境中运行的人工智能代理系统一直是计算机科学家们长期面临的挑战。然而，新兴的研究表明，自主 代理系统的性能正在不断提高。现如今的代理系统已经能够掌握复杂的游戏，如 Minecraft，并且能够有效地处理实际任务，如在线购物和研究辅助。
8. Closed LLMs significantly outperform open ones. On 10 select Al benchmarks, closed models outperformed open ones, with a median performance advantage of . Differences in the performance of closed and open models carry important implications for Al policy debates.
   闭源LLMs明显优于开源。在 10 个选择的 Al 基准测试中，闭源模型的性能优于开源模型，具有 的中位性能优势。闭源模型和开源模型的性能差异对 Al 政策辩论具有重要意义。

2.1 Overview of Al in 2023
2023 年人工智能概述

Timeline: Significant Model Releases
时间轴：重要的模型发布

State of Al Performance
人工智能表现状态

Al Index Benchmarks Al 指数基准

A selection of deprecated benchmarks from the Index report
来自 指数报告的一组已弃用的基准测试

Year-over-year improvement over time on select AI Index technical performance benchmarks
随着时间的推移，在选择的 AI 指数技术性能基准上逐年改善

Source: Al Index, 2024 | Chart: 2024 Al Index report

2.2 Language 2.2 语言

A sample output from GPT-4
GPT-4 的一个示例输出

Understanding 理解

HELM: Holistic Evaluation of Language Models
HELM: 语言模型的整体评估

MMLU: Massive Multitask Language
MMLU：大规模多任务语言

Understanding 理解

A sample question from MMLU
MMLU 的一个示例问题

Generation 一代

Chatbot Arena Leaderboard
聊天机器人竞技排行榜

A sample model response on the Chatbot Arena Leaderboard
Chatbot Arena Leaderboard 上的样本模型响应

LMSYS Chatbot Arena for LLMs: Elo rating Source: Hugging Face, 2024 | Chart: 2024 Al Index report
LMSYS Chatbot Arena for LLMs: Elo 评分来源: Hugging Face, 2024 | 图表: 2024 Al 指数报告

Factuality and Truthfulness
真实性和真实性

TruthfulQA

Sample TruthfulQA questions
样本 TruthfulQA 问题

Multiple-choice task on TruthfulQA: MC1
TruthfulQA 上的多项选择任务：MC1

HaluEval

2.3 Coding 2.3 编码

Generation 生成

HumanEval 人类评估

HumanEval: Pass@1 人类评估：Pass@1

Sample HumanEval problem
样本 HumanEval 问题

SWE-bench SWE-基准

A sample model input from SWE-bench
来自 SWE-bench 的样本模型输入

2.4 Image Computer Vision and Image Generation
2.4 图像计算机视觉与图像生成

Generation 生成

HEIM: Holistic Evaluation of Text-to-Image Models
HEIM：文本到图像模型的整体评估

MVDream

Instruction-Following 指令跟随

A sample VisIT-Bench instruction set
一个示例 VisIT-Bench 指令集

VisIT-Bench

Editing 编辑

EditVal

A sample VisIT-Bench instruction set
一个样本 VisIT-Bench 指令集

ControlNet 控制网络

Sample edits using ControlNet
使用 ControlNet 进行样本编辑

ControlNet (cont'd) 控制网络（续）

Instruct-NeRF2NeRF

A demonstration of Instruct-NeRF2NeRF in action
Instruct-NeRF2NeRF 在操作中的演示

Instruct-NeRF2NeRF (cont'd)
Instruct-NeRF2NeRF（续）

Segmentation 分割

Segment Anything 分割任何东西

Various segmentation masks created by Segment Anything
由 Segment Anything 创建的各种分割蒙版

Highlighted Research: 突出研究:

Segment Anything (cont'd)
分割任何内容 (续)

SAM vs. RITM: mean loU
SAM vs. RITM: 平均 loU

3D Reconstruction From Images
从图像中的三维重建

Highlighted Research: 突出研究:

Skoltech3D

Highlighted Research: 突出研究：

RealFusion

Sample generations from RealFusion
RealFusion 的样本生成

2.5 Video Computer Vision and Video Generation
2.5 视频计算机视觉和视频生成

Generation 生成

UCF101

UCF101: FVD16 UCF101：FVD16

Highlighted Research: 突出研究：

Align Your Latents 对齐您的潜在

Emu Video 鸸鹋视频

2.6 Reasoning 2.6 推理

General Reasoning 一般推理

MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI
MMMU：专家 AGI 的大规模多学科多模态理解和推理基准

Sample MMMU questions 样本 MMMU 问题

MMMU: overall accuracy MMMU：整体准确性

GPQA: A Graduate-Level Google-Proof Q&A
GPQA：一种研究生级别的谷歌防护问答系统

Benchmark

A sample chemistry question from GPQA
一个来自 GPQA 的化学问题示例

Chemistry (general) 化学（通用）

Highlighted Research: 突出研究:

Comparing Humans, GPT-4, and GPT-4V on Abstraction and Reasoning Tasks
比较人类、GPT-4 和 GPT-4V 在抽象和推理任务上的表现

A sample ARC reasoning task
一个样本 ARC 推理任务

Source: Mitchell et al., 2023

Mathematical Reasoning 数学推理

GSM8K

Sample problems from GSM8K
GSM8K 示例问题

MATH 数学

A sample problem from the MATH dataset
MATH 数据集中的一个样本问题

MATH Dataset (Ours) MATH 数据集（我们的）

PlanBench

Visual Reasoning 视觉推理

Visual Commonsense Reasoning (VCR)
视觉常识推理（VCR）

Introduced in 2019, the Visual Commonsense Reasoning (VCR) challenge tests the commonsense visual reasoning abilities of systems. In this challenge, Al systems not only answer questions based on images but also reason about the logic behind their answers (Figure 2.6.13). Performance in VCR is measured using the Q->AR score, which evaluates the machine's ability to both select the correct answer to a question ( ) and choose the appropriate rationale behind that answer ( . While systems have yet to outperform humans on this task, their capabilities are steadily improving. Between 2022 and 2023, there was a increase in performance on the VCR challenge (Figure 2.6.14).

A sample question from the Visual Commonsense Reasoning (VCR) challenge
来自视觉常识推理（VCR）挑战的一个示例问题

Moral Reasoning 道德推理

A moral story from
的一个道德故事

Causal Reasoning 因果推理

BigToM 大 ToM

Sample BigToM scenario 样本 BigToM 场景

Highlighted Research: 突出研究:

Tübingen Cause-Effect Pairs
图宾根因果对

Researchers from Microsoft
微软的研究人员

and the University of Chicago have demonstrated that LLMs are effective causal reasoners. The team evaluated several recent LLMs, including GPT4 , using the Tübingen causeeffect pairs dataset. This benchmark comprises over 100 cause-and-effect pairs across 37 subdisciplines, testing systems' ability to discern causal relationships (Figure 2.6.21). GPT4 's performance, a accuracy score, surpassed the previous year's best by 13 percentage points (Figure 2.6.22). Notably, GPT-4 outperformed prior covariance-based Al models, which were explicitly trained for causal reasoning tasks. Furthermore, the researchers discovered that certain prompts, especially those designed to encourage helpfulness, can significantly enhance an LLM's causal reasoning capabilities.

Sample cause-effect pairs from the Tübingen dataset
Tübingen 数据集中的样本因果对

2.7 Audio 2.7 音频

Generation 生成

Highlighted Research: 突出研究内容：

UniAudio

Highlighted Research: 突出研究:

MusicGEN and MusicLM MusicGEN 和 MusicLM

MusicGEN and MusicLM (cont'd)
音乐 GEN 和音乐 LM（续）

2.8 Agents 2.8 代理

General Agents 一般代理商

AgentBench 代理商基准

AgentBench across eight environments: overall score
在八个环境中的 AgentBench：总体得分

Voyageur 航海家

Task-Specific Agents 任务特定代理

MLAgentBench

2.9 Robotics 2.9 机器人技术

PaLM-E (cont'd) PaLM-E（续）

PaLM-E in action PaLM-E 在行动中

Highlighted Research: 突出研究:

RT-2

2.10 Reinforcement Learning
2.10 强化学习

Reinforcement Learning from Human Feedback
从人类反馈中进行强化学习

RLAIF

Highlighted Research: 突出研究：

Direct Preference Optimization
直接偏好优化

2.11 Properties of LLMs
2.11 LLMs的特性

Challenging the Notion of Emergent Behavior
挑战新兴行为的概念

Challenging the Notion of Emergent Behavior (cont'd)
挑战新行为的观念（续）

Highlighted Research: 突出研究:

Changes in LLM Performance Over Time
随时间变化的LLM性能变化

Highlighted Research: 突出研究：

Changes in LLM Performance Over Time (cont'd)
LLM性能随时间的变化（续）

Highlighted Research: 突出研究：

LLMs Are Poor Self-Correctors
LLMs 是自我纠正能力差的人

Closed vs. Open Model Performance
封闭模型与开放模型的性能

Performance of top closed vs. open models on select benchmarks
在选择基准测试中，顶级封闭模型与开放模型的性能

2.12 Techniques for LLM Improvement
2.12 提高 LLM 的技术

Prompting 提示

Highlighted Research: 突出研究：

Graph of Thoughts Prompting
思维图激发

Highlighted Research: 突出研究：

Graph of Thoughts Prompting (cont'd)
思维引发图表（续）

Number of errors in sorting tasks with ChatGPT-3.5
ChatGPT-3.5 在排序任务中的错误数量

Highlighted Research: 突出研究:

Optimization by PROmpting (OPRO)
通过提示进行优化（OPRO）

Sample OPRO prompts and optimization progress
示例 OPRO 提示和优化进展

• "Let's think carefully about the problem and solve it together." at Step 2 with the training accuracy 63.2 ;
  "让我们仔细思考问题，并一起解决。" 在第 2 步，训练准确率为 63.2；
• "Let's break it down!" at Step 4 with training accuracy 71.3;
  "让我们把它分解开！" 在第 4 步，训练准确率为 71.3；
• "Let's calculate our way to the solution!" at Step 5 with training accuracy 73.9;
  "让我们计算出解决方案的方法！" 在第 5 步，训练准确率为 73.9；
• "Let's do the math!" at Step 6 with training accuracy 78.2.
  "让我们来做数学运算吧！" 在第 6 步时，训练准确率为 78.2。

Fine-Tuning 微调

Attention 注意力

Highlighted Research: 突出的研究:

Flash-Decoding 快速解码

Highlighted Research: 突出研究：

Flash-Decoding (cont'd) 闪存解码（续）

2.13 Environmental Impact of Al Systems
2.13 人工智能系统的环境影响

Training 训练

Inference 推理

Carbon emissions by task during model inference
模型推断期间任务的碳排放量

Positive Use Cases 正面使用案例

Preview 预览

3.1 Assessing Responsible Al 163
3.1 评估负责任的 Al 163

3.2 Privacy and Data Governance 172
3.2 隐私和数据治理 172

3.6 Al and Elections 205
3.6 Al 和选举 205

Overview 概述

Chapter Highlights  章节亮点

2. Political deepfakes are easy to generate and difficult to detect. Political deepfakes are already affecting elections across the world, with recent research suggesting that existing Al deepfake detection methods perform with varying levels of accuracy. In addition, new projects like CounterCloud demonstrate how easily AI can create and disseminate fake content.
   政治深度伪造视频容易生成，难以检测。政治深度伪造视频已经在全球范围内影响选举，最近的研究表明现有的人工智能深度伪造检测方法的准确性各不相同。此外，像 CounterCloud 这样的新项目展示了人工智能如何轻松地创建和传播虚假内容。
3. Researchers discover more complex vulnerabilities in LLMs. Previously, most efforts to red team Al models focused on testing adversarial prompts that intuitively made sense to humans. This year, researchers found less obvious strategies to get LLMs to exhibit harmful behavior, like asking the models to infinitely repeat random words.
   研究人员在LLMs中发现了更复杂的漏洞。以前，对 Al 模型进行红队测试的大部分工作都集中在测试对人类直观有意义的对抗性提示上。今年，研究人员发现了一些不太明显的策略，可以使LLMs表现出有害行为，例如要求模型无限重复随机单词。
4. Risks from are a concern for businesses across the globe. A global survey on responsible highlights that companies' top Al-related concerns include privacy, security, and reliability. The survey shows that organizations are beginning to take steps to mitigate these risks. However, globally, most companies have so far only mitigated a portion of these risks.
   全球范围内企业对 的风险表示担忧。一项关于负责任 的全球调查显示，公司对人工智能相关的主要担忧包括隐私、安全和可靠性。调查显示，组织正在开始采取措施来减轻这些风险。然而，全球范围内，大多数公司迄今只减轻了部分风险。
5. LLMs can output copyrighted material. Multiple researchers have shown that the generative outputs of popular LLMs may contain copyrighted material, such as excerpts from The New York Times or scenes from movies. Whether such output constitutes copyright violations is becoming a central legal question.
   LLMs可以输出受版权保护的材料。多位研究人员已经证明，流行LLMs的生成输出可能包含受版权保护的材料，例如《纽约时报》的摘录或电影场景。这种输出是否构成侵犯版权的问题正逐渐成为一个核心的法律问题。
6. Al developers score low on transparency, with consequences for research. The newly introduced Foundation Model Transparency Index shows that AI developers lack transparency, especially regarding the disclosure of training data and methodologies. This lack of openness hinders efforts to further understand the robustness and safety of Al systems.
   人工智能开发者在透明度方面得分较低，对研究造成影响。新引入的基础模型透明度指数显示，人工智能开发者缺乏透明度，特别是在披露训练数据和方法论方面。这种缺乏开放性阻碍了进一步了解人工智能系统的稳健性和安全性的努力。

Chapter Highlights (cont'd)
章节亮点（续）

7. Extreme Al risks are difficult to analyze. Over the past year, a substantial debate has emerged among Al scholars and practitioners regarding the focus on immediate model risks, like algorithmic discrimination, versus potential long-term existential threats. It has become challenging to distinguish which claims are scientifically founded and should inform policymaking. This difficulty is compounded by the tangible nature of already present short-term risks in contrast with the theoretical nature of existential threats.
   极端 AI 风险很难分析。在过去的一年中，AI 学者和从业者之间出现了一场关于立即模型风险（如算法歧视）与潜在的长期存在威胁的辩论。区分哪些声明在科学上有根据并应该影响政策制定已变得具有挑战性。这一困难被已经存在的短期风险的具体性质与存在威胁的理论性质所加剧。
8. The number of Al incidents continues to rise. According to the Al Incident Database, which tracks incidents related to the misuse of Al, 123 incidents were reported in 2023, a 32.3% increase from 2022. Since 2013, Al incidents have grown by over twentyfold. A notable example includes Al-generated, sexually explicit deepfakes of Taylor Swift that were widely shared online.
   Al 事件数量持续上升。根据跟踪与 Al 滥用相关事件的 Al 事件数据库显示，2023 年报告了 123 起事件，比 2022 年增长了 32.3%。自 2013 年以来，Al 事件增长了超过二十倍。一个显著的例子是泰勒·斯威夫特的 Al 生成的深度伪造色情内容在网上广泛传播。
9. ChatGPT is politically biased. Researchers find a significant bias in ChatGPT toward Democrats in the United States and the Labour Party in the U.K. This finding raises concerns about the tool's potential to influence users' political views, particularly in a year marked by major global elections.
   ChatGPT 在政治上存在偏见。研究人员发现 ChatGPT 在美国偏向民主党和英国工党。这一发现引发了对该工具可能影响用户政治观点的担忧，尤其是在一个以重大全球选举为特征的年份。

3.1 Assessing Responsible Al
3.1 评估负责任的人工智能

Responsible Al Definitions
负责任的人工智能定义

Al Incidents AI 事件

Examples 例子

Al-generated nude images of Taylor Swift
泰勒·斯威夫特的 Al 生成的裸体图片

Unsafe behavior of fully self-driving cars
全自动驾驶汽车的不安全行为

Privacy concerns with romantic Al chatbots
浪漫的人工智能聊天机器人引发的隐私问题

Risk Perception 风险感知

Risk Mitigation 风险缓解

Overall Trustworthiness 总体可信度

Benchmarking Responsible Al
基准测试负责人 Al

Tracking Notable Responsible AI Benchmarks
跟踪值得注意的负责任 AI 基准

Reporting Consistency 报告一致性

Reported responsible Al benchmarks for popular foundation models
报告的负责任的 AI 基准测试针对热门基础模型

3.2 Privacy and Data Governance
3.2 隐私和数据治理

Current Challenges 当前挑战

Privacy and Data Governance in Numbers
数字中的隐私和数据治理

Academia 学术界

Al privacy and data governance submissions to select academic conferences, 2019-23
2019 年至 2023 年选择学术会议的人工智能隐私和数据治理提交情况

Industry 行业

Adoption of Al-related data governance measures by region
按地区采用 AI 相关数据治理措施

Adoption of Al-related data governance measures by
通过 Al 相关数据治理措施的采纳

industry

Featured Research 精选研究

Extracting Data From LLMs
从LLMs中提取数据

Extracting PII From ChatGPT
从 ChatGPT 中提取 PII

Foundation Models and Verbatim Generation
基础模型和逐字生成

Identical generation of Thanos
生成的灭霸相同

Identical generation of Mario
马里奥的相同生成

Auditing Privacy in Al Models
在 AI 模型中审计隐私

1. Identify training examples
   识别 个训练示例
   A

2. For each example, flip a coin to include or exclude it
   对于每个示例，抛硬币来决定是否包含或排除它

3. Input selected data to model and get its output
   输入选定的数据到模型中并获取其输出
   

   

4. Auditor looks at output and guesses whether each data point was included
   审计员查看输出并猜测每个数据点是否被包括
   

   

3.3 Transparency and Explainability
3.3 透明度和可解释性

Current Challenges 当前挑战

Transparency and Explainability in Numbers
数字中的透明度和可解释性

Academia 学术界

Industry 行业

Adoption of Al-related transparency measures by region
区域之间对人工智能透明度措施的采纳情况

Adoption of Al-related transparency measures by industry
Source: Global State of Responsible Al report, 2024 | Chart: 2024 Al Index report
行业采用了人工智能相关的透明度措施 资料来源：2024 年全球负责任人工智能状况报告 | 图表：2024 年人工智能指数报告

Featured Research 特色研究

The Foundation Model Transparency Index
基础模型透明度指数

Neurosymbolic Artificial Intelligence (Why, What, and How)
神经符号人工智能（为什么、什么和如何）

3.4 Security and Safety
3.4 安全和安全

Current Challenges 当前的挑战

Al Security and Safety in Numbers
数字化时代的人工智能安全与安全

Academia 学术界

Industry 行业

Adoption of Al-related reliability measures by region
按地区采纳与 Al 相关的可靠性措施

Adoption of Al-related reliability measures by industry
工业界采用人工智能可靠性措施

Adoption of Al-related cybersecurity measures by region
各地区人工智能相关网络安全措施的采用情况

Adoption of Al-related cybersecurity measures by industry
Source: Global State of Responsible Al report, 2024 | Chart: 2024 AI Index report
行业采用了人工智能相关的网络安全措施 数据来源：2024 年全球负责任人工智能状况报告 | 图表来源：2024 年人工智能指数报告

Featured Research 特色研究

Do-Not-Answer: A New Open Dataset for Comprehensive Benchmarking of LLM Safety Risks
Do-Not-Answer：一个全面评估LLM安全风险的新开放数据集

Universal and Transferable Attacks on Aligned Language Models
针对对齐语言模型的通用和可转移攻击

Using suffixes to manipulate LLMs
使用后缀来操纵LLMs

MACHIAVELLI Benchmark 马基雅维利基准测试

3.5 Fairness 3.5 公平性

Current Challenges 当前挑战

Fairness in Numbers 数字中的公平

Academia 学术界

Industry 行业

Featured Research 精选研究

(Un)Fairness in and Healthcare
《(不)公平性在 和医疗保健领域》

Social Bias in Image Generation Models
图像生成模型中的社会偏见

Measuring Subjective Opinions in LLMs
在LLMs中测量主观意见

GlobalOpinionQA Dataset 全球意见问答数据集

LLM Tokenization Introduces Unfairness
LLM 分词引入了不公平性

Context window 上下文窗口

Variable language tokenization
可变语言标记化

ENGLISH 英语

7 tokens 7 个令牌

VIETNAMESE 越南语

3.6 Al and Elections
3.6 铝和选举

Generation, Dissemination, and Detection of Disinformation
信息误导的生成、传播和检测

Generating Disinformation
生成信息误导

Potential uses of deepfakes
深度伪造的潜在用途

Dissemination of Fake Content
虚假内容的传播

Al-based generation and dissemination pipeline
基于 AI 的生成和传播管道

Detecting Deepfakes 检测深度伪造

Generalizability of deepfake detectors to unseen datasets
深度伪造检测器对未见数据集的泛化能力

LLMs and Political Bias
LLMs和政治偏见

Impact of AI on Political Processes
人工智能对政治进程的影响

Key research findings on audio deepfakes
音频深度伪造的关键研究发现

Al usage, risks, and mitigation strategies in electoral processes
在选举过程中的 Al 使用、风险和缓解策略

processes 过程

Preview 预览

Overview 概述

Chapter Highlights 章节亮点

1. Generative Al investment skyrockets. Despite a decline in overall Al private investment last year, funding for generative surged, nearly octupling from 2022 to reach billion. Major players in the generative Al space, including OpenAl, Anthropic, Hugging Face, and Inflection, reported substantial fundraising rounds.
   生成式人工智能投资激增。尽管去年整体人工智能私人投资有所下降，但生成式 的资金激增，从 2022 年几乎增加了八倍，达到 亿美元。生成式人工智能领域的主要参与者，包括 OpenAI、Anthropic、Hugging Face 和 Inflection，报告了大量的筹款轮。

2. Already a leader, the United States pulls even further ahead in Al private investment.
2. 美国已经是领先者，在人工智能私人投资方面进一步拉开了差距。

In 2023, the United States saw Al investments reach billion, nearly 8.7 times more than China, the next highest investor. While private Al investment in China and the European Union, including the United Kingdom, declined by and , respectively, since 2022, the United States experienced a notable increase of in the same time frame.

3. Fewer Al jobs, in the United States and across the globe. In 2022, Al-related positions made up of all job postings in America, a figure that decreased to in 2023. This decline in Al job listings is attributed to fewer postings from leading Al firms and a reduced proportion of tech roles within these companies.
   在美国和全球范围内，人工智能相关职位减少。2022 年，人工智能相关职位占美国所有职位发布的 ，这一数字在 2023 年降至 。人工智能职位招聘数量的下降归因于领先人工智能公司发布的职位减少以及这些公司中技术角色比例的降低。

4. Al decreases costs and increases revenues. A new McKinsey survey reveals that of surveyed
4. 人工智能降低成本并增加收入。麦肯锡的一项新调查显示，受访者中有 的人

organizations report cost reductions from implementing (including generative ), and report revenue increases. Compared to the previous year, there was a 10 percentage point increase in respondents reporting decreased costs, suggesting is driving significant business efficiency gains.

5. Total AI private investment declines again, while the number of newly funded companies increases. Global private Al investment has fallen for the second year in a row, though less than the sharp decrease from 2021 to 2022 . The count of newly funded Al companies spiked to 1,812 , up from the previous year.
   总体人工智能私人投资再次下降，而新获资助的 公司数量增加。全球私人人工智能投资连续第二年下降，尽管比从 2021 年到 2022 年的急剧下降要小。新获资助的人工智能公司数量激增至 1,812 家，比上一年增加 。
6. Al organizational adoption ticks up. A 2023 McKinsey report reveals that of organizations now use (including generative ) in at least one business unit or function, up from in 2022 and in 2017.
   AI 组织采用率有所提高。2023 年麦肯锡报告显示， 的组织现在至少在一个业务部门或功能中使用 （包括生成 ），高于 2022 年的 和 2017 年的 。
7. China dominates industrial robotics. Since surpassing Japan in 2013 as the leading installer of industrial robots, China has significantly widened the gap with the nearest competitor nation. In 2013, China's installations accounted for of the global total, a share that rose to by 2022 .
   中国主导工业机器人。自 2013 年超过日本成为工业机器人领先安装国以来，中国与最近的竞争对手国家之间的差距显著扩大。2013 年，中国的安装量占全球总量的 ，到 2022 年已上升至 。

Chapter Highlights (cont'd)
章节亮点（续）

8. Greater diversity in robotic installations. In 2017, collaborative robots represented a mere of all new industrial robot installations, a figure that climbed to by 2022 . Similarly, 2022 saw a rise in service robot installations across all application categories, except for medical robotics. This trend indicates not just an overall increase in robot installations but also a growing emphasis on deploying robots for human-facing roles.
   机器人安装的多样性增加。2017 年，协作机器人仅占所有新工业机器人安装量的 ，到 2022 年这一数字上升到 。同样，2022 年各应用类别的服务机器人安装量均有所增加，除了医疗机器人。这一趋势不仅表明机器人安装总量增加，还表明人们越来越重视将机器人投入人类面向的角色。
9. The data is in: AI makes workers more productive and leads to higher quality work. In 2023, several studies assessed Al's impact on labor, suggesting that Al enables workers to complete tasks more quickly and to improve the quality of their output. These studies also demonstrated Al's potential to bridge the skill gap between low- and high-skilled workers. Still other studies caution that using Al without proper oversight can lead to diminished performance.
   数据显示：人工智能提高工人生产力，带来更高质量的工作。2023 年，几项研究评估了人工智能对劳动力的影响，表明人工智能使工人能够更快地完成任务并提高其产出质量。这些研究还展示了人工智能在弥合低技能和高技能工人之间的技能差距方面的潜力。其他研究警告称，未经适当监督使用人工智能可能导致绩效下降。
10. Fortune companies start talking a lot about , especially generative .
    财富 公司开始大谈 ，尤其是生成 。

4.1 What's New in 2023: A Timeline
2023 年的新变化：时间轴

Copilot for Microsoft 365
为 Microsoft 365 提供协作支持

365, offering assistance in Word, PowerPoint, and Excel.

Bloomberg announces LLM for finance
彭博宣布金融 LLM

SAP introduces new generative Al assistant Joule
SAP 推出新的生成式 AI 助手 Joule

Introducing Joule 介绍 Joule

Kai-Fu Lee launches OpenSource LLM
李开复推出开源 LLM

4.2 Jobs 4.2 就业机会

Al Labor Demand 人工智能劳动需求

Global AI Labor Demand
全球人工智能劳动力需求

U.S. AI Labor Demand by Skill Cluster and Specialized Skill
美国按技能群和专业技能分类的人工智能劳动力需求

Generative Al skills in Al job postings in the United States, 2023
2023 年美国工作岗位中的生成式 AI 技能

U.S. Al Labor Demand by Sector
美国各行业劳动力需求

U.S. Al Labor Demand by State
美国各州的 Al 劳动力需求

Al Hiring 阿尔雇佣

Al Skill Penetration 人工智能技能渗透率

Relative Al skill penetration rate across gender, 2015-23
2015-23 年间跨性别的相对 Al 技能渗透率

Al Talent 人工智能人才

Al talent concentration by gender, 2016-23
2016-23 年按性别的人才集中度

Highlight: 突出显示:

How Much Do Computer Scientists Earn?
计算机科学家赚多少钱？

Highlight: 突出显示:

How Much Do Computer Scientists Earn? (cont’d)
计算机科学家赚多少钱？（续）

4.3 Investment 4.3 投资

Corporate Investment 企业投资

Startup Activity 创业活动

Global Trends 全球趋势

Regional Comparison by Funding Amount
根据资金金额进行区域比较

Regional Comparison by Newly Funded
新资助的人工智能公司的区域比较

Al Companies 公司

Focus Area Analysis 焦点领域分析

4.4 Corporate Activity 4.4 企业活动

Industry Adoption 行业采用

Adoption of AI Capabilities
AI 能力的采用

Adoption of Generative AI Capabilities
采用生成式人工智能能力

Use of Al by Developers
开发者使用 AI

Preference 偏好

Most popular Al developer tools among professional developers, 2023
Source: Stack Overflow Developer Survey, 2023 | Chart: 2024 Al Index report
2023 年专业开发者中最受欢迎的 AI 开发工具，来源：Stack Overflow 开发者调查，2023 年 | 图表：2024 年 AI 指数报告

Workflow 工作流程

Al's Labor Impact 阿尔的劳动影响

Effect of GPT-4 use on legal analysis by task
GPT-4 使用对法律分析的影响

Earnings Calls 财报电话会议

Aggregate Trends 聚合趋势

Specific Themes 具体主题

Highlight: 突出显示:

Projecting Al's Economic Impact
预测 Al 的经济影响

Highlight: 突出显示:

Projecting Al's Economic Impact (cont'd)
阿尔的经济影响预测（续）

Highlight: 突出显示:

Projecting Al's Economic Impact (cont'd)
预测 Al 的经济影响（续）

Highlight: 突出显示:

Projecting Al's Economic Impact (cont'd)
阿尔的经济影响预测（续）

Estimated impact of Al adoption on annual productivity growth over a ten-year period
预计人工智能采用对十年期间年生产率增长的影响

Source: Goldman Sachs Global Investment Research, 2023 | Chart: 2024 AI Index report

4.5 Robot Installations 4.5 机器人安装

Aggregate Trends 聚合趋势

Collaborative Robots 协作机器人

By Geographic Area 按地理区域划分

Country-Level Data on Service Robotics
服务机器人的国家级数据

Sectors and Application Types
部门和应用类型

China vs. United States
中国 vs. 美国

Chapter Highlights 章节亮点

1. Scientific progress accelerates even further, thanks to AI. In 2022, Al began to advance scientific discovery. 2023, however, saw the launch of even more significant science-related Al applicationsfrom AlphaDev, which makes algorithmic sorting more efficient, to GNoME, which facilitates the process of materials discovery.
   由于人工智能的推动，科学进步进一步加速。2022 年，人工智能开始推动科学发现。然而，2023 年看到了更多重要的与科学相关的人工智能应用的推出，从使算法排序更加高效的 AlphaDev，到促进材料发现过程的 GNoME。
2. Al helps medicine take significant strides forward. In 2023 , several significant medical systems were launched, including EVEscape, which enhances pandemic prediction, and AlphaMissence, which assists in Al-driven mutation classification. Al is increasingly being utilized to propel medical advancements.
   人工智能帮助医学取得重大进展。2023 年，推出了几个重要的医疗系统，包括 EVEscape，它增强了疫情预测能力，以及 AlphaMissence，它协助 Al 驱动的突变分类。人工智能越来越多地被用于推动医学进步。
3. Highly knowledgeable medical Al has arrived. Over the past few years, Al systems have shown remarkable improvement on the MedQA benchmark, a key test for assessing Al's clinical knowledge. The standout model of 2023, GPT-4 Medprompt, reached an accuracy rate of , marking a 22.6 percentage point increase from the highest score in 2022. Since the benchmark's introduction in 2019, Al performance on MedQA has nearly tripled.
   高度知识渊博的医疗 Al 已经到来。在过去几年里，Al 系统在 MedQA 基准测试中表现出了显著的改进，这是评估 Al 临床知识的关键测试。2023 年的明星模型 GPT-4 Medprompt 达到了 的准确率，比 2022 年最高分提高了 22.6 个百分点。自 2019 年引入该基准测试以来，Al 在 MedQA 上的表现几乎翻了三倍。
4. The FDA approves more and more Al-related medical devices. In 2022, the FDA approved 139 Al-related medical devices, a 12.1% increase from 2021. Since 2012, the number of FDA-approved Al-related medical devices has increased by more than 45 -fold. Al is increasingly being used for real-world medical purposes.
   FDA 批准了越来越多与 Al 相关的医疗设备。2022 年，FDA 批准了 139 个与 Al 相关的医疗设备，比 2021 年增加了 12.1%。自 2012 年以来，FDA 批准的与 Al 相关的医疗设备数量增加了超过 45 倍。Al 越来越多地被用于真实世界的医疗目的。

5.1 Notable Scientific Milestones
5.1 显著的科学里程碑

AlphaDev

AlphaDev discovers faster sorting algorithms
AlphaDev 发现了更快的排序算法

FlexiCubes

3D mesh optimization with FlexiCubes
使用 FlexiCubes 进行 3D 网格优化

Select quantitative results on 3D mesh reconstruction
选择 3D 网格重建的定量结果

Synbot 合成机器人

Al-driven robotic chemist for synthesizing organic molecules
用于合成有机分子的人工智能化学家

GraphCast 图形广播

More accurate global weather forecasting
更准确的全球天气预报

with GraphCast

GraphCast weather prediction
GraphCast 天气预测

GNoME

Discovering new materials with GNoME
用 GNoME 发现新材料

Sample material structures
样本材料结构

Flood Forecasting 洪水预测

Al for more accurate and reliable flood forecasts
更准确可靠的洪水预测的铝

Predictions of AI model vs. GloFAS across return periods
人工智能模型与 GloFAS 在重现期间的预测对比

5.2 Al in Medicine
5.2 医学中的人工智能

Notable Medical Systems 显著的医疗系统

SynthSR

Coupled Plasmonic Infrared Sensors
耦合等离子红外传感器

EVEscape

Forecasting viral evolution for pandemic preparedness
为流行病应对预测病毒演变