How Open Source Machine Learning Software Shapes AI

Abstract

If we want a future where AI serves a plurality of interests, then we should pay attention to the factors that drive its success. While others have studied the importance of data, hardware, and models in directing the trajectory of AI, I argue that open source software is a neglected factor shaping AI as a discipline. I start with the observation that almost all AI research and applications are built on machine learning open source software (MLOSS). This thesis presents four contributions. First, it quantifies the outsized impact of MLOSS by using Github contributions data. By contrasting the costs of MLOSS and its economic benefits, I find that the average dollar of MLOSS investment corresponds to at least $100 of global economic value created, corresponding to $30B of economic value created this year. Second, I leverage interviews with AI researchers and developers to develop a causal model of the effect of open sourcing on economic value. I argue that open sourcing creates value through three primary mechanisms: standardization of MLOSS tools, increased experimentation in AI research, and creation of commuities. Third, I analyze the various incentives behind MLOSS by examining three key factors: business strategy, sociotechnical factors, and ideological motivations. In the last section, I explore how MLOSS may help us understand the future of AI and make a number of probabilistic predictions. I intend this thesis to be useful for technologists and academics who want to analyze and critique AI, and policymakers who want to better understand and regulate AI systems.

AI thinking and the enterprise of science

“…For AI thinking to be effective, AI methodologies must be accessible to non-experts. Recent development of machine learning toolkits has made complex AI methods much more accessible and contributed significantly to the current AI boom (12). But using these toolkits still requires significant expertise in programming and large computational resources, as well as specific paradigms for data and analysis, excluding many scientists and much of the Global South (13).…”

Reproducibility Starts at the Source: R, Python, and Julia Packages for Retrieving USGS Hydrologic Data

“…In this regard, one might consider the development and availability of scientific code packages to be a revolution in scientific philosophy (metascience). Recent advances in machine learning, data science, and many other domains have been accelerated through the availability of open-source packages [13,14].…”

The ethics of AI business practices: a review of 47 AI ethics guidelines

“…These concerns include: (1) economic issues, (2) justice, (3) human freedoms, (4) broader societal issues, and (5) unknown issues. Economic issues are especially significant in AI value chains, and issues of particular note can be found in the use of automation and biometrics in hiring, contracting, dismissal, and surveilling workers (Bales & Stone, 2020;Hickok & Maslej, 2023); labor exploitation, distributions of wealth, capital, and other financial resources, particularly in transnational and inter-class contexts (Dyer-Witheford, Kjøsen, & Steinhoff, 2019;; open-sourcing of and access to AI-related data, code, and other software resources (Langenkamp & Yue, 2022;Masiello & Slater, 2023); distributions of data, computational, technological, and financial resources obtained through the development and/or use of AI systems, as well as distributions of political and economic power that emerge from those resource distributions (Dyer-Witheford, Kjøsen, & Steinhoff, 2019;Pasquale, 2020). Examples of specific cases involving some of these economic issues include: OpenAI's outsourcing of data labeling to workers employed by Sama AI in Kenya, many of whom were psychologically harmed and undercompensated during their employment (Perrigo, 2023); consolidation of models and datasets in an increasingly small group of companies (Ahmed, Wahed, & Thompson, 2023); Amazon's development, use, and subsequent disuse of a hiring automation tool that discriminated against women (Dastin, 2018); striking Screen Actors Guild and Writers Guild of America workers demanding their employers refrain from using their likenesses or unionprotected creative materials in training datasets and from introducing generative AI applications into production processes (Broderick, 2023;Webster, 2023).…”