AI2.0 时代的数字学术及其范式变革

摘要/Abstract

摘要：

数字学术是数字技术和方法介入学术研究过程的研究范式。以ChatGPT 为代表的生成式人工智能的突破，标志着通用人工智能的到来，有学者称之为AI2.0 时代，也使得科学研究从数据密集型向计算密集型过渡，从而进入以人工智能介入研究过程的第五范式，越来越多的研究过程将有人工智能参与，甚至主导。与数据驱动型研究有所不同的是，大模型不仅可以更加高效甚至“自动”进行数据获取、提取、管理和分析等工作，而且有可能实现海量科学数据的“理解”，从而发现新的规律和趋势。大模型不仅可以帮助科学家更好地理解数据，还可以为科学研究提供更加全面和深入的视角，同时，它还能应用于多种相关场景，帮助科学家发现数据或成果中的错误和缺陷，从而提高科学研究的可靠性和可重复性。当然，随着人工智能技术的广泛应用，也带来了一些挑战。对于图书馆而言，面向科研的服务将更加学术化和技术化，除了基础设施、科研交流环境建设和数据支撑服务之外，还将以AI 思维和数字化方法与工具更加深入地介入科研过程，并在保护数据的隐私和安全、避免数据泄露和滥用，以及数据伦理方面提供必要的支持。

Abstract:

Digital scholarship is a research paradigm in which digital technologies and methodsintervene in the academic research process. The breakthrough of generative AI represented by ChatGPTmarks the arrival of generalized AI, which some scholars call the “AI2.0 era”, and also makes scientificresearch transition from data-intensive to computationally-intensive, thus entering the fifth paradigm.More and more of the research process will be involved in, or even dominated by AI. Different from datadrivenresearch, big models not only can conduct data acquisition, extraction, management and analysisin a more efficient, or even “automatic” way, but also have the potential to realize the “understanding” ofmassive scientific data, so as to discover new laws and trends. The big model can not only help scientistsbetter understand the data, but also provide a more comprehensive and in-depth perspective for scientificresearch. At the same time, it can also be applied to a variety of related scenarios, to help scientistsfind errors and defects in the data or results, thereby improving the reliability and repeatability ofscientific research. Of course, with the wide application of AI technology, it also brings some challenges.For libraries, research-oriented services will become more academic and technical. In addition toinfrastructure, construction of research communication environment and data support services, they willintervene more deeply in the research process with AI thinking and digital methods and tools, provide thenecessary support in protecting the privacy and security of data, and avoid data leakage, misuse or ethicalissues.

金家琴刘炜（上海图书馆）. AI2.0 时代的数字学术及其范式变革[J]. 图书馆杂志, 2023, 42(391): 10-21.

Jin Jiaqin, Liu Wei (Shanghai Library). The Transformation of Digital Scholarship Paradigm in AI2.0 Era[J]. Libraly Journal, 2023, 42(391): 10-21.

参考文献

[ 1 ] Hey T， Tansley S， Tolle K，等 . 第四范式：数据密集型科学发现[M]. 北京：科学出版社， 2012.

[ 2 ] 邓仲华，李志芳 . 科学研究范式的演化——大数据时代的科学研究第四范式[J]. 情报资料工作，2013（4）：19–23.

[ 3 ] J Shaw. Why “Big Data” Is a Big Deal：InformationScience Promises to Change the World[J]. HarvardMagazine，2014，2（3/4）： 30–35；74–75.

[ 4 ] Jinghai Li， Wenlai Huang.Paradigm shift inscience with tackling global challenges[J]. NationalScience Review， 2019， 6（6）： 1091–1093.

[ 5 ] C Leng，Z Tang，Y-G Zhou，et al. Fifth Paradigmin Science： A Case Study of an Intelligence-DrivenMaterial Design[J/OL]. Engineering， https://doi.org/10.1016/j.eng.2022.06.027.

[ 6 ] Wang H， Fu T， Du Y， et al. Scientific discoveryin the age of artificial intelligence[J]. Nature，2023， 620（7972）： 47–60.

[ 7 ] Wang H， Fu T， Du Y， et al. Artificial intelligencein science： An emerging general method ofinvention[J]. Research Policy， 2022， 620（7972）：47–60.

[ 8 ] Christopher Bishop.AI4Science to empower thefifth paradigm of scientific discovery[EB/OL].[2023-07-10]. https://www.microsoft.com/en-us/research/blog/ai4science-to-empower-the-fifthparadigm-of-scientific-discovery/.

[ 9 ] 张伟 . AI 赋能科研“第五范式”引变革[N]. 中_国高新技术产业导报，2022-08-15（13） .

[10] 王飞跃，缪青海 . 人工智能驱动的科学研究新范式：从AI4S 到智能科学[J]. 中_国科学院院刊，2023，38（4）：536–540.

[11] 王飞跃，缪青海，张军平，等 . 探讨AI for Science的影响与意义：现状与展望[J]. 智能科学与技术学报，2023，5（1）：1–6.

[12] 梁娜，曾燕 . 推进数据密集科学发现提升科技创新能力：新模式、新方法、新挑战——《第四范式：数据密集型科学发现》译著出版[J]. 中_国科学院院刊，2013，28（1）：115–121.

[13] Zhao W X，Zhou K，Li J，et al. A survey of largelanguage models[EB/OL]. [2023-07-17]. arXivpreprint arXiv：2303.18223， 2023.

[14] Terry S. 深度学习：智能时代的核心驱动力量[M].北京：中_信出版社，2019.

[15] Brown T， Mann B， Ryder N， et al. Languagemodels are few-shot learners[J]. Advances inneural information processing systems， 2020， 33：1877–1901.

[16] Radford A， Wu J， Child R， et al. Languagemodels are unsupervised multitask learners[J].OpenAI blog， 2019， 1（8）：9.

[17] 王静静，叶鹰，王婉茹 . GPT 类技术应用开启智能信息处理之颠覆性变革[J]. 图书馆杂志，2023（5）：9–13.

[18] Radford A， Narasimhan K， Salimans T， et al.Improving language understanding by generativepre-training[EB/OL]. [2023-07-17]. https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf.

[19] Rober W McGee.Annie Chan：Three Short StoriesWritten with Chat GPT[EB/OL]. [2023-07-17].DOI：10.13140/RG.2.2.21169.66401（2023） .

[20] Zhou J，Ke P，Qiu X，et al.ChatGPT： potential，prospects， and limitations[J]. Front InformTechnol Electron Eng（2023） . DOI：https://doi.org/10.1631/FITEE.2300089.

[21] Wang Y， Yao Q， Kwok J T， et al. Generalizing froma few examples： A survey on few-shot learning[J].ACM computing surveys，2020， 53（3）：1–34.

[22] How Chat GPT Will Revolutionize The Futureof Bakery[EB/OL]. [2023-07-10]. https://www.puratos.com/blog/how-chat-gpt-will-revolutionizethe-future-of-bakery.

[23] Wei J， Tay Y， Bommasani R， et al. Emergentabilities of large language models[J]. Transactionson Machine Learning Research，2022. DOI：https://doi.org/10.48550/arXiv.2206.07682.

[24] Wei J，Wang X，Schuurmans D， et al. Chainof-thought prompting elicits reasoning inlarge language models[J]. Advances in NeuralInformation Processing Systems， 2022， 35：24824–24837.

[25] Silver D，Huang A， Maddison C J，et al. Masteringthe game of Go with deep neural networks and treesearch[J]. nature， 2016， 529（7587）： 484–489.

[26] GPT-4[EB/OL]. [2023-07-10]. https://openai.com/research/gpt-4.

[27] Scott Clark. Unlock the potential of your data with AI：How to break down data silos and extract valuableinsights with AI and NLP technology[EB/OL].[2023-07-10]. https://www.cmswire.com/customerexperience/how-artificial-intelligence-can-breakthrough-data-silos/.

[28] Antoine Tardif. Is GPT-4 a Leap Forward TowardsReaching AGI?[EB/OL]. [2023-07-10]. https://www.unite.ai/is-gpt-4-a-leap-forward-towardsreaching-agi/.

[29] Bubeck S， Chandrasekaran V， Eldan R， et al.Sparks of artificial general intelligence： Earlyexperiments with gpt-4[EB/OL]. [2023-07-10].arXiv preprint arXiv：2303.12712， 2023.

[30] Sergio De Simone. Running Large LanguageModels Natively on Mobile and Laptops[EB/OL].[2023-07-10]. https://www.infoq.com/news/2023/ 05/mlc-llm-mobile-laptops/.

[31] Liz O’Sullivan，John Dickerson. Here are a fewways GPT-3 can go wrong[EB/OL]. [2023-07-10].https://techcrunch.com/2020/08/07/here-are-a-fewways-gpt-3-can-go-wrong/.

[32] 郑一波，曾建勋 . 图书馆发现系统的实质与应对策略[J]. 情报科学，2022，40（8）：134–137；143.

[33] Perplexity[EB/OL]. [2023-07-10]. https://www.perplexity.ai/.

[34] Tang R， Han X， Jiang X， et al. Does syntheticdata generation of llms help clinical textmining?[EB/OL]. [2023-07-10]. arXiv preprintarXiv：2303.04360， 2023.

[35] Jamall Brown. Building a Search Chatbot withGPT-3[EB/OL]. [2023-07-10]. https://medium.com/@jamall.j.brown/enhancing-knowledgeretrieval-with-gpt-3-part-1-c0f0c3b27d01.

[36] 浙商证券 . GPT-4 发布多模态处理+ 复杂推理能力有望打开应用空间[EB/OL]. [2023-07-10]. https://pdf.dfcfw.com/pdf/H3_AP202303161584308890_1.pdf?1678965405000.pdf.

[37] 托马斯·库恩 . 科学革命的结构[M]. 北京：北京大学出版社， 2022.