PDF(1875 KB)
PDF(1875 KB)
PDF(1875 KB)
迈向“全球网关”：接入视角下的大语言模型基础设施化研究
To be the Global Gateways: An Empirical Study on the Infrastructuralization of Large Language Models from the Access Perspective
本研究通过实证路径探究大语言模型在全球范围内的接入情况，以此考察其基础设施化程度及其迈向通用人工智能的困境。从传播与媒介研究的连接观入手，借基础设施核心概念“网关”，研究使用计算机网络实验法，引入丢包率、时延与抖动等指标，考察大语言模型的全球接入状况（可达、速度与稳定）及其基础设施化与数字不平等的问题。基于全球62城网络节点对中外6款大语言模型所进行的近20万次网络发包，研究发现相较全球北方城市节点，全球南方城市节点在接入层面有较为普遍的劣势，在访问欧美产大模型时尤其明显。在可达性与速度层面，部分大模型已基本完成对搜索引擎、数据库等前代信息基础设施的超越，且全球南方城市节点接入大模型的速度已显著快于数据库；但在稳定性层面，大模型尚未展示出超越前代信息基础设施的显著优越性。尽管大语言模型的问世在隐喻意义上可被视作全球信息交往与人机互联的“初级阶段”，但亟需面对和解决连接中的地缘矛盾。在接入、使用、稳定、普及与价值等多层面发力，或是大语言模型激发其潜质、通往“全球网关”的可能路径。
This paper empirically explores the access of large language models (LLMs) on a global scale, to examine their infrastructuralization and dilemmas towards general AI. Drawing upon the basic idea of connectivity from communication and media studies, and adopting “gateway” as the core concept of infrastructure, this study employs computer network experiments as the primary method and utilizes indicators including packet loss, latency, and jitter to examine the global access conditions including accessibility, speed, stability of LLMs as new digital infrastructure along with the potential inequality issues behind them. Through conducting nearly 200,000 network probes across 62 global network nodes, this study finds that compared to city nodes in the Global North, city nodes in the Global South generally exhibit a disadvantage in terms of accessibility, especially when accessing LLMs produced in Western countries. Furthermore, some LLMs have significantly surpassed “old-type” information infrastructures like search engines and databases in terms of accessibility and speed, with nodes in the Global South accessing LLMs notably faster than databases. However, LLMs have not yet demonstrated a significant superiority in stability over previous-generation information infrastructures. Although the emergence of LLMs metaphorically signifies a primary stage of global information exchange and human-machine interactions, it is still imperative to address and resolve geopolitical conflicts within connectivity. Addressing the factors like access, usability, stability, universalisation and value will be crucial for LLMs to realize their full potential and evolve into global gateways.
大语言模型 / 可接入性 / 网关 / 基础设施化 / 数字接入沟
Large language model / accessibility / gateway / infrastructuralization / digital access divide
| [1] |
百川大模型(2024年4月10日). 百川奔腾入沧海,直挂云帆向远方. https://mp.weixin.qq.com/s/DEQdyBaQvlXzJRGDpNJZCQ
|
| [2] |
卞冬磊(2021). 遗忘与重建:作为“传播”的“交通”. 《新闻大学》,(1),36-47+118-119.
|
| [3] |
曹冲(2021). “一带一路”倡议下中国与中亚五国基础设施的贸易效应研究. 《大连理工大学学报(社会科学版)》,(3),36-45.
|
| [4] |
陈昌凤, 黄阳坤(2023). ChatGPT的知识功能与人类的知识危机. 《现代出版》,(6),10-18.
|
| [5] |
陈昌凤, 袁雨晴(2024). 智能新闻业:生成式人工智能成为基础设施. 《内蒙古社会科学》,(1),40-48.
|
| [6] |
杜莉华, 吴世文(2023). 退隐互联网:主动网络失联的演进、实践与论争. 《国际新闻界》,(10),6-27.
|
| [7] |
胡翼青, 胡欣阅(2023). 作为语言基础设施的ChatGPT. 《新闻记者》,(6),21-27.
|
| [8] |
姬德强, 闫伯维(2023). 人工智能的地缘政治:传播政治经济学的视角. 《南昌大学学报(人文社会科学版)》,(6),84-92.
|
| [9] |
匡文波, 姜泽玮(2024). 论智能传播研究的基本理论问题. 《中国人民大学学报》,(3),115-126.
|
| [10] |
雷少华(2019). 超越地缘政治——产业政策与大国竞争. 《世界经济与政治》,(5),131-154+160.
|
| [11] |
瞭望东方周刊(2020年4月26日). 新基建,是什么. 新华网. http://www.xinhuanet.com/politics/2020-04/26/c_1125908061.htm
|
| [12] |
彭兰(2013). “连接”的演进——互联网进化的基本逻辑. 《国际新闻界》,(12),6-19.
|
| [13] |
任孟山, 李呈野(2023). 作为国际传播议题的人工智能: 知识生产与全球权力. 《中国出版》,(17),5-12.
|
| [14] |
史安斌, 朱泓宇(2022). 发展传播学的叙事更新与逻辑转化:“传播基础设施”的概念与取向之辩. 《南昌大学学报(人文社会科学版)》,(5),77-86.
|
| [15] |
束开荣(2021). 互联网基础设施:技术实践与话语建构的双重向度——以媒介物质性为视角的个案研究. 《新闻记者》,(2),39-50.
|
| [16] |
唐海娜, 李俊(2004). 网络性能监测技术综述. 《计算机应用研究》,(8),10-13.
|
| [17] |
涂良川, 乔良(2023). “关键技术”创新逻辑的政治叙事——基于《芯片战争》的政治哲学考察. 《山东社会科学》,(11),103-112.
|
| [18] |
王浩宇, 王永杰(2023). 基础设施工具理性的缺陷及其价值理性的回归. 《中国人民大学学报》,(1),33-46.
|
| [19] |
王维佳, 张涵抒(2022). 超越地缘?全球卫星网络的创造与失败. 《全球传媒学刊》,(6),3-20.
|
| [20] |
沃尔夫冈·恩斯特, 高山, 黄家圣(2023). 中国特有的技术进路存在吗?. 《全球传媒学刊》,(2),196-210.
|
| [21] |
夏立平, 田博(2020). 论国际新智缘政治的范式与影响. 《同济大学学报(社会科学版)》,(6),53-63.
|
| [22] |
新浪科技(2024年5月9日). 阿里云CTO周靖人:通义千问API日调用量破亿,企业用户突破9万. https://tech.hexun.com/2024-05-09/212786657.html
|
| [23] |
喻发胜, 张振宇, 黄海燕(2017). 从传播到“传联”:一个新概念提出的学理依据、现实背景与理论内涵. 《新闻大学》,(2),63-72+149.
|
| [24] |
袁连海, 陆利刚(2018). 《计算机网络实验教程》. 清华大学出版社.
|
| [25] |
证券时报网(2024年4月16日). 文心一言用户数突破2亿,API日均调用量突破2亿. https://stcn.com/article/detail/1177150.html
|
| [26] |
Complexity is all around us in this increasingly digital world. Global digital infrastructure, social media, Internet of Things, robotic process automation, digital business platforms, algorithmic decision making, and other digitally enabled networks and ecosystems fuel complexity by fostering hyper-connections and mutual dependencies among human actors, technical artifacts, processes, organizations, and institutions. Complexity affects human agencies and experiences in all dimensions. Individuals and organizations turn to digitally enabled solutions to cope with the wicked problems arising out of digitalization. In the digital world, complexity and digital solutions present new opportunities and challenges for information systems (IS) research. The purpose of this special issue is to foster the development of new IS theories on the causes, dynamics, and consequences of complexity in increasing digital sociotechnical systems. In this essay, we discuss the key theories and methods of complexity science, and illustrate emerging new IS research challenges and opportunities in complex sociotechnical systems. We also provide an overview of the five articles included in the special issue. These articles illustrate how IS researchers build on theories and methods from complexity science to study wicked problems in the emerging digital world. They also illustrate how IS researchers leverage the uniqueness of the IS context to generate new insights to contribute back to complexity science.
|
| [27] |
|
| [28] |
|
| [29] |
|
| [30] |
|
| [31] |
|
| [32] |
|
| [33] |
Finance Center for South-South Cooperation. (2015). Global South Countries (Group of 77 and China). http://www.fc-ssc.org/en/partnership_program/south_south_countries
|
| [34] |
|
| [35] |
We propose a design theory that tackles dynamic complexity in the design for Information Infrastructures (IIs) defined as a shared, open, heterogeneous and evolving socio-technical system of Information Technology (IT) capabilities. Examples of IIs include the Internet, or industry-wide Electronic Data Interchange (EDI) networks. IIs are recursively composed of other infrastructures, platforms, applications and IT capabilities and controlled by emergent, distributed and episodic forms of control. II's evolutionary dynamics are nonlinear, path dependent and influenced by network effects and unbounded user and designer learning. The proposed theory tackles tensions between two design problems related to the II design: (1) the bootstrap problem: IIs need to meet directly early users’ needs in order to be initiated; and (2) the adaptability problem: local designs need to recognize II's unbounded scale and functional uncertainty. We draw upon Complex Adaptive Systems theory to derive II design rules that address the bootstrap problem by generating early growth through simplicity and usefulness, and the adaptability problem by promoting modular and generative designs. We illustrate these principles by analyzing the history of Internet exegesis.
|
| [36] |
|
| [37] |
|
| [38] |
|
| [39] |
Studies of infrastructure have demonstrated broad differences between Northern and Southern cities, and deconstructed urban theory derived from experiences of the networked urban regions of the Global North. This includes critiques of the universalisation of the historically–culturally produced normative ideal of universal, uniform infrastructure. In this commentary, we first introduce the notion of ‘heterogeneous infrastructure configurations’ (HICs) which resonates with existing scholarship on Southern urbanism. Second, we argue that thinking through HICs helps us to move beyond technological and performative accounts of actually existing infrastructures to provide an analytical lens through which to compare different configurations. Our approach enables a clearer analysis of infrastructural artefacts not as individual objects but as parts of geographically spread socio-technological configurations: configurations which might involve many different kinds of technologies, relations, capacities and operations, entailing different risks and power relationships. We use examples from ongoing research on sanitation and waste in Kampala, Uganda – a city in which service delivery is characterised by multiplicity, overlap, disruption and inequality – to demonstrate the kinds of research questions that emerge when thinking through the notion of HICs.
|
| [40] |
|
| [41] |
|
| [42] |
OpenAI. (2023, November 8). Major Outage across ChatGPT and API. https://status.openai.com/incidents/00fpy0yxrx1q
|
| [43] |
|
| [44] |
|
| [45] |
Two theoretical approaches have recently emerged to characterize new digital objects of study in the media landscape: infrastructure studies and platform studies. Despite their separate origins and different features, we demonstrate in this article how the cross-articulation of these two perspectives improves our understanding of current digital media. We use case studies of the Open Web, Facebook, and Google to demonstrate that infrastructure studies provides a valuable approach to the evolution of shared, widely accessible systems and services of the type often provided or regulated by governments in the public interest. On the other hand, platform studies captures how communication and expression are both enabled and constrained by new digital systems and new media. In these environments, platform-based services acquire characteristics of infrastructure, while both new and existing infrastructures are built or reorganized on the logic of platforms. We conclude by underlining the potential of this combined framework for future case studies.
|
| [46] |
|
| [47] |
Universal access (UA) to the Internet and the associated information infrastructure has become an important economic and societal goal. However, UA initiatives tend to focus on issues such as physical access and geographical ubiquity, and they measure adoption through penetration rates. In this paper, we apply an interpretive case study approach to analyze the Philadelphia wireless initiative to provide insights into the nature of UA and extend this concept to also consider universal use (UU). UU is important because simply providing access does not guarantee use. UU is presented as a conceptual goal that starts with the challenge of physical access, but which necessarily also leads to considerations of use. The results show that the human and technological elements underlying individual access and use are deeply embedded within various institutional elements and collectives that enable but also constrain meaningful use. We integrate our findings into a multilevel framework that shows how access and use are influenced by both micro and macro factors. This framework provides new insights into the study of the information infrastructure, digital divide, and public policy.
|
| [48] |
|
| [49] |
|
| [50] |
|
| [51] |
|
| [52] |
We analyze a large-scale custom software effort, the Worm Community System (WCS), a collaborative system designed for a geographically dispersed community of geneticists. There were complex challenges in creating this infrastructural tool, ranging from simple lack of resources to complex organizational and intellectual communication failures and tradeoffs. Despite high user satisfaction with the system and interface, and extensive user needs assessment, feedback, and analysis, many users experienced difficulties in signing on and use. The study was conducted during a time of unprecedented growth in the Internet and its utilities (1991–1994), and many respondents turned to the World Wide Web for their information exchange. Using Bateson's model of levels of learning, we analyze the levels of infrastructural complexity involved in system access and designer-user communication. We analyze the connection between systems development aimed at supporting specific forms of collaborative knowledge work, local organizational transformation, and large-scale infrastructural change.
|
| [53] |
Since the inauguration of information systems research (ISR) two decades ago, the information systems (IS) field's attention has moved beyond administrative systems and individual tools. Millions of users log onto Facebook, download iPhone applications, and use mobile services to create decentralized work organizations. Understanding these new dynamics will necessitate the field paying attention to digital infrastructures as a category of IT artifacts. A state-of-the-art review of the literature reveals a growing interest in digital infrastructures but also confirms that the field has yet to put infrastructure at the centre of its research endeavor. To assist this shift we propose three new directions for IS research: (1) theories of the nature of digital infrastructure as a separate type of IT artifact, sui generis; (2) digital infrastructures as relational constructs shaping all traditional IS research areas; (3) paradoxes of change and control as salient IS phenomena. We conclude with suggestions for how to study longitudinal, large-scale sociotechnical phenomena while striving to remain attentive to the limitations of the traditional categories that have guided IS research.
|
| [54] |
UNOPS. (n. d.). Infrastructure. Retrieved January 12, 2024, from https://www.unops.org/expertise/infrastructure
|
| [55] |
Van der Vlist, F. N., Helmond, A., Burkhardt, M., & Seitz, T. (2022). API governance: The case of Facebook’s evolution. Social Media+Society, 8(2), 20563051221086228.
|
| [56] |
|
| [57] |
For a long time, a common opinion among policy-makers was that the digital divide problem would be solved when a country’s Internet connection rate reaches saturation. However, scholars of the second-level digital divide have concluded that the divides in Internet skills and type of use continue to expand even after physical access is universal. This study—based on an online survey among a representative sample of the Dutch population—indicates that the first-level digital divide remains a problem in one of the richest and most technologically advanced countries in the world. By extending basic physical access combined with material access, the study finds that a diversity in access to devices and peripherals, device-related opportunities, and the ongoing expenses required to maintain the hardware, software, and subscriptions affect existing inequalities related to Internet skills, uses, and outcomes.
|
| [58] |
|
| [59] |
|
| [60] |
|
1. 丢包的成因复杂，包括防火墙或交换机阻断、网络线缆故障、软件程序出错、网络拥堵等，但丢包导致的直接可见后果即信息基础设施的访问受阻。
2. 这一设定受宽带发展联盟进行的“中国宽带速率状况”调查启发。因用户上网有相对明确的高峰和非高峰时段，一天四次基本可覆盖网络使用的忙时（常取19时至23时）和闲时（常取1时至6时），进而降低了忙闲时对研究结果的混淆和干扰。
3. PLR的分母不完全一致，即因为这些节点在对相关大模型进行拨测实验时，出现了“数据声明”中所言的情况：网络环境问题造成部分城市网络节点Ping拨测失败，出现了数据缺失。
/
| 〈 |
|
〉 |
