This essay examines a structural resemblance between two generative processes: the composition of formal verse by a skilled poet and the inference process of autoregressive large language models. It does not argue for mechanistic or neurobiological equivalence. Its aim is interpretive: to identify a recurring pattern that clarifies both human composition and machine generation. The comparison is developed across four dimensions. First, corpus formation: the poet's experiential, multimodal corpus is contrasted with the LLM's textual training corpus. The essay argues that the critical distinction is not corpus size but grounding: current language models recombine distributional facts about tokens, while the poet recombines sensory primitives acquired through embodiment, assembled into scenes that may be lived, observed, or imagined. Second, constrained generation: prosodic form is treated as an analogue to sampling constraints, showing how limitation can intensify rather than diminish output quality. Third, emergent semantics: in both systems, meaning can arise without being pre-specified by the operator. Fourth, the verifier function: the poet's evaluative judgment is compared to alignment, oversight, and selection mechanisms in deployed AI systems. The essay concludes that formal poetic practice offers a historically deep model of disciplined interaction with generative systems: one that depends neither on servile acceptance nor on total dictation, but on guided generation and rigorous selection. On that view, the poet-language relationship illuminates not only literary composition but also broader questions of AI governance, operator training, alignment, and the epistemology of working with generative systems. This record includes the English paper and an author-prepared Russian companion version of the same work. The Russian text is a companion rendering of the same core argument, adapted for Russian prose rather than presented as a line-by-line translation.
Edward Meyman (Wed,) studied this question.