Centralization, the concentration of control or trust in a small number of organizations, is the dominant organizing principle of modern AI. While it has powered rapid progress, centralization also means that control over how AI systems are trained, deployed, and coordinated rests with few parties. This thesis argues that decentralized AI is not only possible but practical, and presents eight contributions across three frontiers of modern AI: training on distributed data, deploying trustworthy autonomous agents, and enabling multi-agent collaboration across organizational boundaries.
We begin by establishing foundational infrastructure with DecentralizePy, a modular framework for decentralized learning research that scales to over 1,000 agents and serves as the experimental backbone for subsequent contributions.
Efficient Decentralized Training. We address coordination, communication, and asynchrony in decentralized learning. Epidemic Learning replaces static communication topologies with randomized ones, achieving up to 1.7× faster convergence while handling data heterogeneity. JWINS introduces wavelet-based sparsification that reduces communication by up to 64% without significant accuracy loss. DivShare introduces model fragmentation to asynchronous settings, enabling nodes with slow links to contribute partial updates rather than blocking progress, achieving 3.9× faster time-to-accuracy in the presence of stragglers.
Trustworthy Decentralized Training. We address privacy and fairness as essential requirements for adoption. Shatter provides noiseless privacy protection through model chunking and virtual nodes, reducing privacy attack success rates by up to 52% without adding noise or sacrificing model utility. Facade addresses fairness under feature heterogeneity through an emergent clustering mechanism with a multi-head architecture, achieving 60% minority-cluster accuracy while reducing communication by up to 41.3%.
Decentralizing Agentic AI. We extend decentralized design principles to the emerging landscape of AI agents. Prudentia introduces IFC-aware agent planning that turns security policy compliance into an explicit objective alongside task completion, reducing human-in-the-loop interventions by up to 1.9× while defending against prompt injection attacks. The Web of Agents proposes minimal standards for open multi-agent collaboration, identifying four building blocks, agent-to-agent messaging, interaction interoperability, state management, and discovery, that enable interoperability across framework boundaries without imposing a single protocol.
Together, these contributions demonstrate that the limitations imposed by centralization are not inherent, and that modern AI systems can be decentralized by design without sacrificing efficiency, security, or scalability.
