報(bào)告簡(jiǎn)讀：僅用于記錄

私人投資在AI領(lǐng)域的變化-投資集中度高且傾向于頭部研究機(jī)構(gòu)---投資
AI領(lǐng)域的研究以中美主導(dǎo)和合作為主---研究模式
語(yǔ)言模型的表達(dá)能力得到增強(qiáng)，但是也增加了偏見---語(yǔ)言模型NLP
AI模型中的精英特色在崛起---投資
AI硬件的價(jià)格越來(lái)越低同時(shí)性能也逐年增強(qiáng)---硬件
數(shù)據(jù)是AI永恒的主題---數(shù)據(jù)
全球立法中關(guān)于AI模型的研究越來(lái)越常見---模型認(rèn)知
機(jī)械臂的價(jià)格變得更便宜---硬件

Eng:

Part 1:

Despite rising geopolitical tensions, the United States and China had the greatest number of cross-country collaborations in AI publications from 2010 to 2021, increasing five times since 2010. The collaboration between the two countries produced 2.7 times more publications than between the United Kingdom and China—the second highest on the list. In 2021, China continued to lead the world in the number of AI journal, conference, and repository publications—63.2% higher than the United States with all three publication types combined. In the meantime, the United States held a dominant lead among major AI powers in the number of AI conference and repository citations. From 2010 to 2021, the collaboration between educational and nonprofit organizations produced the highest number of AI publications, followed by the collaboration between private companies and educational institutions and between educational and government institutions. The number of AI patents filed in 2021 is more than 30 times higher than in 2015, showing a compound annual growth rate of 76.9%.

Part 2:

Data, data, data: Top results across technical benchmarks have increasingly relied on the use of extra training data to set new state-of-the-art results. As of 2021, 9 state-of-the-art AI systems out of the 10 benchmarks in this report are trained with extra data. This trend implicitly favors private sector actors with access to vast datasets. Rising interest in particular computer vision subtasks: In 2021, the research community saw a greater level of interest in more specific computer vision subtasks, such as medical image segmentation and masked-face identification. For example, only 3 research papers tested systems against the Kvasir-SEG medical imaging benchmark prior to 2020. In 2021, 25 research papers did. Such an increase suggests that AI research is moving toward research that can have more direct, real-world applications. AI has not mastered complex language tasks, yet: AI already exceeds human performance levels on basic reading comprehension benchmarks like SuperGLUE and SQuAD by 1%–5%. Although AI systems are still unable to achieve human performance on more complex linguistic tasks such as abductive natural language inference (aNLI), the difference is narrowing. Humans performed 9 percentage points better on aNLI in 2019. As of 2021, that gap has shrunk to 1.Turn toward more general reinforcement learning: For the last decade, AI systems have been able to master narrow reinforcement learning tasks in which they are asked to maximize performance in a specific skill, such as chess. The top chess software engine now exceeds Magnus Carlsen’s top ELO score by 24%. However, in the last two years AI systems have also improved by 129% on more general reinforcement learning tasks (Procgen) in which they must operate in novel environments. This trend speaks to the future development of AI systems that can learn to think more broadly. AI becomes more affordable and higher performing: Since 2018, the cost to train an image classification system has decreased by 63.6%, while training times have improved by 94.4%. The trend of lower training cost but faster training time appears across other MLPerf task categories such as recommendation, object detection and language processing, and favors the more widespread commercial adoption of AI technologies. Robotic arms are becoming cheaper: An AI Index survey shows that the median price of robotic arms has decreased by 46.2% in the past five years—from 22,600 in 2021. Robotics research has become more accessible and affordable.

Part 3:

Language models are more capable than ever, but also more biased: Large language models are setting new records on technical benchmarks, but new data shows that larger models are also more capable of reflecting biases from their training data. A 280 billion parameter model developed in 2021 shows a 29% increase in elicited toxicity over a 117 million parameter model considered the state of the art as of 2018. The systems are growing significantly more capable over time, though as they increase in capabilities, so does the potential severity of their biases. The rise of AI ethics everywhere: Research on fairness and transparency in AI has exploded since 2014, with a fivefold increase in related publications at ethics-related conferences. Algorithmic fairness and bias has shifted from being primarily an academic pursuit to becoming firmly embedded as a mainstream research topic with wide-ranging implications. Researchers with industry affiliations contributed 71% more publications year over year at ethics-focused conferences in recent years. Multimodal models learn multimodal biases: Rapid progress has been made on training multimodal languagevision models which exhibit new levels of capability on joint language-vision tasks. These models have set new records on tasks like image classification and the creation of images from text descriptions, but they also reflect societal stereotypes and biases in their outputs—experiments on CLIP showed that images of Black people were misclassified as nonhuman at over twice the rate of any other race. While there has been significant work to develop metrics for measuring bias within both computer vision and natural language processing, this highlights the need for metrics that provide insight into biases in models with multiple modalities.

Part 4:

? New Zealand, Hong Kong, Ireland, Luxembourg, and Sweden are the countries or regions with the highest growth in AI hiring from 2016 to 2021. ? In 2021, California, Texas, New York, and Virginia were states with the highest number of AI job postings in the United States, with California having over 2.35 times the number of postings as Texas, the second greatest. Washington, D.C., had the greatest rate of AI job postings compared to its overall number of job postings. The private investment in AI in 2021 totaled around 500 million or more; in 2021, there were 15.“Data management, processing, and cloud” received the greatest amount of private AI investment in 2021— 2.6 times the investment in 2020, followed by “medical and healthcare” and “fintech.”In 2021, the United States led the world in both total private investment in AI and the number of newly funded AI companies, three and two times higher, respectively, than China, the next country on the ranking. Efforts to address ethical concerns associated with using AI in industry remain limited, according to a McKinsey survey. While 29% and 41% of respondents recognize “equity and fairness” and “explainability” as risks while adopting AI, only 19% and 27% are taking steps to mitigate those risks. In 2020, 1 in every 5 CS students who graduated with PhD degrees specialized in artificial intelligence and machine learning, the most popular specialty in the past decade. From 2010 to 2020, the majority of AI PhDs in the United States headed to industry while a small fraction took government jobs.

Part 5:

An AI Index analysis of legislative records on AI in 25 countries shows that the number of bills containing “artificial intelligence” that were passed into law grew from just 1 in 2016 to 18 in 2021. Spain, the United Kingdom, and the United States passed the highest number of AI-related bills in 2021, with each adopting three. The federal legislative record in the United States shows a sharp increase in the total number of proposed bills that relate to AI from 2015 to 2021, while the number of bills passed remains low, with only 2% ultimately becoming law. State legislators in the United States passed 1 out of every 50 proposed bills that contain AI provisions in 2021, while the number of such bills proposed grew from 2 in 2012 to 131 in 2021. In the United States, the current congressional session (the 117th) is on track to record the greatest number of AI-related mentions since 2001, with 295 mentions by the end of 2021, half way through the session, compared to 506 in the previous (116th) session.