From Marconi to Transformers: A Comparative Analysis of Radio Antennas and Attention Mechanisms

Our work proposes a receiver-first and field-first ontology for large language models (LLMs). Instead of treating these systems as “creative generators,” we model Transformer inference as correlation-driven routing that transduces learned semantic structure into token sequences, in direct correspondence with classical receiver primitives that extract structure from noise by alignment. Beginning from the scaled dot-product attention kernel, we formalize inference as a normalized inner-product routing operator that selects, aggregates, and decodes contextual content, clarifying why surface fluency is evidence of coupling to linguistic regularities rather than evidence of truth-conditional validity. We then introduce a minimal, falsifiable definition of a Large Language Field (LLF): an LLF exists to the extent that semantic structure is observable as invariants preserved under admissible transformations across independently trained models and checkpoints, rather than as fixed coordinates in any particular representation. This “invariants-first” stance yields concrete falsifiers and controls, including cross-model alignment tests, representation-similarity probes, stability under retraining, and robustness under dataset shift and calibration stress. To make governance native to this ontology, we formalize a field stack — subfields → fields → metafields → overfields → metaoverfields — where regimes restrict subfields, metafields capture compositional portability across subfields, overfields enforce admissibility through falsifiers, receipts, and explicit budgets, and metaoverfields arbitrate conflicts among competing regimes. Finally, we provide a test plan and measurement program that separates (i) memorization, (ii) transduction over learned structure, and (iii) cross-model invariants consistent with LLF behavior, using metrics for invariant preservation, regime validity, contradiction and consistency rates, abstention calibration, verification yield, and layered collapse signatures from subfields through metaoverfields.