Narrow-Band Senses: An Information-Theoretic Framework for Multimodal AI Perception

Multimodal AI is usually discussed as a frontier-lab arms race: who can fuse vision, audio and text into the largest model. The paper proposes a different frame. Treat every non-language signal — DNA sequences, music notation, manufacturing code, medical telemetry, animal vocalisation — as a language with its own statistical structure, and the question "can AI perceive this?" becomes a translation question with a measurable answer.

Three empirical pillars carry the argument. First, an entropy-based criterion: a structure score computed from raw data alone, in minutes and without any training, predicts eventual learnability across roughly thirty diverse domains with a rank correlation near −0.92. The more structured the stream, the more learnable the bridge. Second, forward-bridge universality: across sixteen specialised domains, models translate the specialised stream into natural language once sufficient paired data exists. Third, reverse bridges: the same machinery generates specialised data from natural-language description in multiple symbolic domains.

A four-factor capacity model assembles the pillars into a practical instrument: given a domain’s data structure, volume, pairing and symbol granularity, it predicts what an AI system can achieve there before anyone trains anything. The decade’s implication runs against the arms-race frame — which expert-augmenting tools become possible depends more on the structure and availability of domain data than on which laboratory trains the largest model. Code and data are public.