公共人工智能（Public AI）白皮书

Legal notice Commissioned by© Bertelsmann Stiftung, GüterslohMay 2025 Rights Thetextof this publication is licensed under theCreative Commons Attribution 4.0 InternationalLicense. You can find the complete license textat: https://creativecommons.org/licenses/by/4.0/legalcode.en Publisher Bertelsmann StiftungCarl-Bertelsmann-Straße 25633311 GüterslohPhone +49 5241 81-0www.bertelsmann-stiftung.de Theinfographicsare licensed under the CreativeCommons Attribution-NonCommercial-NoDerivatives4.0 International License. You can find the completelicense text at: https://creativecommons.org/licenses/by-nc-nd/4.0/legalcode.en Supported byOpen Future Authors Dr. Felix Sieker,Bertelsmann StiftungDr. Alek Tarkowski,Open FutureLea Gimpel,Digital Public Goods AllianceDr. Cailean Osborne,Oxford Internet Institute The visualizations are not meant to be exhaustive.Alllogosare excluded, as they are protected bycopyright, not covered by the above-mentioned CClicense, and may not be used. ResponsibleDr. Felix Sieker,Bertelsmann Stiftung Recommended citation styleSieker/Tarkowski/Gimpel/Osborne (2025). Public AI –White Paper. Bertelsmann Stiftung. Gütersloh. EditingBarbara Serfozo, Berlin DOI 10.11586/2025040 InfographicsJakub Koźniewski Layout and TypesettingNicole Meyerholz, Bielefeld Public AI – White Paper Dr. Felix Sieker,Dr. Alek Tarkowski,Lea Gimpel,Dr. Cailean Osborne Reviewer list Albert Cañigueral,Barcelona Supercomputing CenterAmin Oueslati,The Future SocietyBen Burtenshaw,Hugging FaceBrandon Jackson,Public AI NetworkHuw Roberts,Oxford Internet Institute, University ofOxfordIsabel Hou,Taiwan AI AcademyJakob Mökander,Digital Ethics Center, Yale UniversityJennifer Ding,Boundary Object StudioLaura Galindo,AI policy expertLuca Cominassi,AI policy expertMartin Hullin,Bertelsmann StiftungMartin Pompéry,SINE FoundationMarta Ziosi,Oxford Martin AI Governance Initiative,University of OxfordPaul Keller,Open FuturePaul Sharratt,Sovereign Tech AgencyRavi Iyer,USC MarshallYacine Jernite,Hugging FaceZoe Hawkins,Tech Policy Design Institute Table of contents Preface6 Executive summary8 Glossary11 1 |Introduction13 2 |Technical primer: What are AI technologies and how do they work?17Defining artificial intelligence17The deep learning paradigm21“Attention is all you need”: Transformers and the rise of generative AI22The generative AI development process24Pretraining phase24Post-training phase26Deployment26AI scaling laws: The contested future of AI26What are AI scaling laws?27The evolution of AI scaling laws27Scaling and AI’s environmental footprint28What is the future of AI scaling laws?29 3 |The generative AI stack32Overview of the AI stack32Advantages of the AI stack concept33Concentrations of power in the AI stack35Characteristics of key layers of the stack37Compute37Data39Models42 4 |The public AI framework45The concept of public digital infrastructure45Public, private and civic actors in public digital infrastructure47Proposals for public AI47Public AI Network47Mozilla Foundation48Vanderbilt Policy Accelerator49Defining public AI infrastructure49Gradient of publicness of AI systems52Goals and governance principles of public AI policies55Governance of public AI56 5 |AI strategy and three pathways to public AI58 Elements of a public AI strategy58The public AI ecosystem and its orchestrating institution60Three pathways toward public AI infrastructure: compute, data and model61Compute pathway to public AI61Compute: bottlenecks61Compute: opportunities63 Data pathway to public AI65 Data: bottlenecks65Data: opportunities66 Models: bottlenecks68Models: opportunities69 Additional measures71 Coda: mission-driven public AI policy73 Preface Grounded in a realistic analysis of the constraintsacross the AI stack – compute, data and models – thepaper translates the concept of Public AI into a con-crete policy framework with actionable steps. Cen-tral to this framework is the conviction that public AIstrategies must ensure the continued availability ofat least one fully open-source model with capabili-ties approaching those of proprietary state-of-the-art systems. Achieving this goal requires three keyactions: coordinated investing in the open-sourceecosystem, providing public compute infrastructure,and building a robust talent base and institution-al capacity. Artificial Intelligence stands at a pivotal crossroads.While its potential to transform society is immense,the power to shape its trajectory is becoming in-creasingly concentrated. Today, a small number ofdominant technology firms hold sway not only overthe most advanced AI models but also the founda-tional infrastructure – compute capacity, data re-sources and cloud platforms – that makes thesesystems possible. This consolidation of influencerepresents more than a market imbalance; it poses adirect threat to the principles of openness, transpar-ency and democratic accountability. When only a handful of actors define how AI systemsare built and used,