Signal-Based Modular Intelligence (SBMI) A Conceptual Framework for Non-Linguistic, Signal-First Artificial Intelligence Systems

Current artificial intelligence systems are dominated by monolithic, token-based architectures that process linguistic intermediates end-to-end. These systems face well-documented structural limitations: compute inefficiency, memory bandwidth bottlenecks, opaque internal reasoning, and tight coupling between perception, reasoning, and output. This paper proposes Signal-Based Modular Intelligence (SBMI), a conceptual architecture organized around a different principle. Instead of tokens, SBMI uses a shared, non-linguistic signal representation that encodes entities, spatial relationships, motion, uncertainty, and relational structure in a modality-invariant form. Instead of a single monolithic model, SBMI distributes work across independently trained transformation modules, called discs, that communicate bidirectionally with a central coordination layer responsible for maintaining signal format consistency and semantic alignment. Perceptual input is progressively compressed across the network into decision-ready representations, with the smallest possible representation that still preserves decision-relevant structure. The paper specifies the architectural shape of such a system, defines the role of each component, names the open problems that any working implementation would need to solve (signal format design, alignment, independent training compatibility, optical substrate feasibility), and positions SBMI against transformer, CNN, vector database, and knowledge graph approaches. The contribution is a framing and a vocabulary intended to invite specification and critique, not a working system or empirical result. This is a position paper from an independent researcher, written with explicit AI assistance disclosed in the Author’s Note.