Open-Sourcing Highly Capable Foundation Models: An evaluation of risks, benefits, and alternative methods for pursuing open-source objectives

• DOI: 10.48550/arxiv.2311.09227
• Publication Date: 2023-01-01
• Authors: Elizabeth Seger; Noemi Dreksler; Richard Moulange; Emily Dardaman; Jonas Schuett; Kevin Wei; Christoph Winter; Mackenzie Arnold; Seán Ó hÉigeartaigh; Anton Korinek; Markus Anderljung; Ben Bucknall; Alan Chan; Eoghan Stafford; Leonie Koessler; Aviv Ovadya; Ben Garfinkel; Emma Bluemke; Michael Aird; Patrick Levermore; Julian Hazell; Abhishek Gupta
• Keywords: Foundation (evidence); Unintended consequences; Open standard; Best practice

Recent decisions by leading AI labs to either open-source their models or restrict access has sparked debate about whether, and how, increasingly capable should be shared. Open-sourcing in typically refers making model architecture weights freely publicly accessible for anyone modify, study, build on, use. This offers advantages such as enabling external oversight, accelerating progress, decentralizing control over development However, it also presents a growing potential misuse unintended consequences. paper an examination of the risks benefits open-sourcing highly foundation models. While historically provided substantial net most software processes, we argue that some likely developed near future, may pose sufficiently extreme outweigh benefits. In case, not open-sourced, at least initially. Alternative strategies, including non-open-source sharing options, are explored. The concludes with recommendations developers, standard-setting bodies, governments establishing safe responsible practices preserving where safe.