Cailean Osborne & Johan Linåker: A Cartography of Collaboration in Open Source AI

As generative artificial intelligence (AI) models become increasingly prevalent and released under various forms of open and permissive licenses, there is a critical need to understand how they are built and who contributes to this process. Currently, there is limited research that maps open collaboration practices across different stages of the development or reuse of models and their constituent artifacts (e.g. training datasets, software, model weights, evaluation benchmarks). This research aims to map and characterize open collaboration (specifically, the “collaboration on-ramps”) at different stages in the development and reuse lifecycle of open generative AI models, with a focus on open large language models (LLMs). Through qualitative interviews with 12 open LLM developers (i.e. Allen Institute for AI, EleutherAI, Cohere Labs, Hugging Face, Meta, Alibaba, the BigScience Workshop, AI Singapore, SpeakLeash, SCB 10X, Fraunhofer IAIS, and the National Library of Norway), this study presents a comprehensive cartography of collaboration practices throughout the lifecycle of open LLMs across diverse organizational contexts, from grassroots initiatives to large technology companies, and world regions. The study provides researchers, developers, business leaders, policymakers, and the wider community with empirical insights into collaboration practices, including motivations, opportunities and challenges, in the emerging open source AI community as well as practical recommendations for participation in or promotion of open source AI collaboration.