Jugnu: Idle-Compute Money for Private Decentralized AI
Year: 2026
Type: preprint
Abstract: We propose Jugnu, a decentralized AI network in which idle consumer PCs run a free personal agent, train a shared open model that improves as people use it, and are paid in a currency, the jugnu, issued only for verified useful work. This is a position paper and research agenda: it sets out the design and names the problems that remain open. It asks whether idle consumer compute can support useful, privately executed AI work and a workable reward economy. Jugnu puts that compute to two uses: chiefly training the shared model, bought by the protocol with a fixed budget each period and paid only for contributions accepted into a released model, and secondly preparing preprocessing material for private computation ("privacy fuel"), whose value depends on future private jobs. A personal agent supplies continuous, mostly deferrable demand. Private jobs run on secret-shared data in two modes. In normal mode a randomly drawn, bonded group of providers is trusted not to collude: privacy holds while fewer than half of them conspire. In high-privacy mode the owner's own home PC holds a share of the data and of the fuel, so the strangers learn nothing even if all of them collude. Users may opt in to share data for training in exchange for a lower price, behind public filters and differential privacy. Buyers burn jugnu for work and workers are minted jugnu for verified work, while new issuance beyond that pays only for training the shared model. One jugnu is provisionally defined as processing one million tokens with a reference model at the quiet-hour price. The price per token rises with load, so agents move work that can wait. A small model on the owner's device drafts tokens for the larger network model to verify. Each component builds on published work, and we are not aware of a system that combines them. Verifying training contributions, private auditing, security against actively malicious parties, and price stability are specified only at design level and are stated as open problems. We give six falsifiable hypotheses with pass thresholds and baselines. Jugnu launches fairly: there is no pre-mine and no founder allocation, and the author holds none. Research and legal status: Jugnu is a theoretical research proposal, not an operating network, financial product, investment opportunity, or offer to issue or sell digital assets. As of publication, no Jugnu tokens have been issued, no public network has been launched, and the author does not operate an exchange or accept investments for the proposed system. The architecture, economic mechanisms, security assumptions, and performance targets described in this paper are hypotheses requiring independent verification. No guarantee of implementation, financial value, privacy, security, or practical feasibility is made. The paper is published for scientific discussion, independent evaluation, criticism, and further research.
Author: Soni, Pritam
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225394520