题　目：Upstream Boundaries and Downstream Innovation in AI Application Development

嘉　宾： Xinyu Li（李欣羽）, Ph.D. Candidate，CUHK Business School

主持人：韩雪雯，助理教授，上海交通大学安泰经济与管理学院

时　间：2026年11月17日（周二）10:00-11:30

地　点：上海交通大学安泰经济与管理学院包兆龙图书馆A303

内容简介:

Tools that simplify the use of pretrained models allow more developers to build AI applications, but it is unclear whether they also promote more innovations. We study this tension on Hugging Face, where staggered integrations of external machine-learning libraries reduced the technical effort required to load, deploy, and share models. Using contribution histories for 2,391 developers across eight communities in a staggered difference-in-differences design, we examine how these integrations changed downstream application development. Integrations attract new developers and substantially increase the number of applications created. Yet this growth is concentrated in applications that receive little follow-up development, have fewer versions and simpler architectures, and attract less user appreciation per application. We further document a handover in downstream production: newcomers account for the expansion in community-related applications, while incumbents reduce such work and redirect effort toward other projects and the upstream model layer. This handover helps explain why lower entry barriers increase output without producing a comparable increase in development depth. Standardized tools give them the know-how to deploy models quickly, but the models themselves remain difficult to understand, and the shift toward simple applications persists over the observed post-integration period with no later recovery in complex output. Separate representations of applications’ descriptive and functional content show where the missing know-why bites: newcomers vary titles, documentation, and stated purposes while keeping functional designs close to the existing stock. Tools that democratize AI application development thus generate broadening without deepening: encapsulation lowers the cost of using a model, but opacity leaves the cost of understanding it unchanged, so use no longer turns into learning. For platforms, the lever that matters is therefore not a further cut in the cost of use but a gain in upstream legibility, so that use once again teaches extension downstream.

演讲人简介：

Xinyu Li is a Ph.D. candidate in Information Systems at CUHK Business School, The Chinese University of Hong Kong. Her research focuses on the economics of AI and digital innovation, with particular interests in open-source software development and digital entrepreneurship. She uses econometrics and machine learning to analyze source code, developer contribution histories, and other data from digital platforms. Her work has been published in Marketing Letters.