by Marlon Barrios Solano
Across poetry, choreography, artificial intelligence, and molecular biology, a shared problem emerges: how does form arise from sequence? How does a linear progression—of words, movements, symbols, or amino acids—give rise to coherent structure, meaning, or material organization? This essay proposes that contemporary large language models (LLMs) offer a crucial conceptual bridge for understanding this problem, linking the poetics of William Carlos Williams, the choreographic thinking of William Forsythe, and the predictive morphology of AlphaFold.
William Carlos Williams famously asserted that “unless there is a new mind there cannot be a new line,” insisting that formal innovation in poetry depends on a transformation of cognition itself. Forsythe, working in the domain of dance, similarly treats the body as a thinking system capable of generating novel form through constraint-based improvisation. LLMs extend these insights computationally: they demonstrate how linear sequences generate form by traversing high-dimensional latent spaces, where each step is both constrained and generative.
By integrating poetics, AI architectures, choreographic systems, and biological modeling, this essay argues that creativity and cognition—human and artificial alike—can be understood as vectorial improvisations: sequential traversals through structured spaces of possibility in which form emerges through movement rather than predefinition.
For William Carlos Williams, the poetic line is not a decorative unit but a cognitive event. In Paterson, he writes: “unless there is a new mind there cannot be a new line, the old will go on repeating itself with recurring deadliness.” The line, in Williams’s poetics, is the visible trace of a mental trajectory. If perception and attention remain unchanged, form will stagnate; novelty requires a reorientation of thought.
Williams rejected inherited metrical structures not for aesthetic rebellion alone, but because they encoded outdated modes of perception. His insistence on “no ideas but in things” signals a procedural understanding of form: meaning arises through engagement with material reality, unfolding step by step. The poetic line functions as a reset of attention, a moment where cognition reconfigures itself in response to what has just occurred.
Crucially, Williams treats form as emergent rather than imposed. The poem discovers its structure in the act of writing, much as thought discovers itself in motion. This idea—that thinking is not the execution of a pre-existing plan but an improvisational unfolding—prefigures both contemporary cognitive science and the architecture of generative AI systems.
Large language models provide a concrete instantiation of this principle. At the surface level, LLMs generate text linearly, one token after another. Internally, however, each token is represented as a vector in a high-dimensional latent space, where semantic, syntactic, and contextual relationships are encoded geometrically.
Language generation in LLMs is therefore not retrieval but traversal. Each generated token updates the model’s internal state, shifting its position in latent space and reshaping the probability landscape of what may follow. What appears externally as a sentence is internally a trajectory through a space of possibilities.
This architecture makes explicit what Williams intuited: a “new line” requires a new cognitive state. Without a shift in vectorial orientation, the model will continue along familiar, high-probability paths, producing repetition or cliché. Techniques such as temperature adjustment, prompt framing, or chain-of-thought prompting do not add content; they redirect movement, pushing the model into less-traveled regions of latent space.
LLMs thus enact a form of synthetic improvisation. Like a poet responding to the momentum of language already written, the model responds to its own prior outputs. Each step is both consequence and condition. Meaning is not stored in depth but assembled on the fly, through sequential negotiation with context.
This model aligns closely with the “mind is flat” thesis proposed by cognitive scientist Nick Chater, which argues that human cognition does not draw from deep, pre-formed internal structures but constructs beliefs, explanations, and decisions in real time. According to this view, thinking is an improvisational process shaped by context, memory, and narrative coherence rather than by hidden mental depths.
LLMs, despite lacking consciousness, mirror this structure with unsettling clarity. They generate convincing reasoning, explanations, and reflections without accessing any inner truth state. When asked to explain their reasoning, they produce narratives that resemble metacognition, yet these explanations are themselves improvised outputs, not windows into an underlying cognitive core.
This parallel suggests that much of what humans experience as reasoning may likewise be surface-level narrative construction. In both biological and artificial systems, cognition emerges from sequential pattern completion constrained by prior trajectories, not from consultation with a deep representational store.
William Forsythe’s choreographic practice extends these ideas into embodied space. Forsythe treats choreography not as a fixed sequence of steps but as a system of propositions, constraints, and modalities that enable dancers to generate movement in real time. His “Improvisation Technologies” articulate movement as a form of thinking—an exploration of spatial, temporal, and anatomical possibilities.
Forsythe’s modalities (folding, shearing, dropping, rotating) function much like vectors: directional prompts that guide motion without determining outcome. A dancer operating within these constraints navigates a high-dimensional space of bodily possibility, discovering form through action rather than executing a pre-scripted plan.
Forsythe has described these choreographic instructions as “little language machines,” highlighting their algorithmic character. Like prompts in an LLM, they shape the space of possible continuations while preserving improvisational freedom. Each movement alters the dancer’s state, creating new affordances and constraints for what can follow.
Here, choreography becomes a morphological process, where form emerges through continuous negotiation between structure and variation. The dancer’s body, like an LLM’s latent space, is not a repository of fixed forms but a dynamic field of potential transitions.
The transition from LLMs and choreography to AlphaFold is therefore not a conceptual leap, but a scaling shift—from linguistic and bodily form to molecular form. AlphaFold addresses a fundamental biological problem: how a linear sequence of amino acids folds into a stable three-dimensional protein structure.
Like LLMs, AlphaFold operates by learning a latent space—this time of molecular conformations—where sequence-to-structure relationships are encoded as vectorial constraints. Protein folding is not solved by explicit physical simulation alone, but by predicting the most probable geometric configuration given a sequence, based on learned patterns across vast datasets.
AlphaFold thus performs morphological reasoning: it navigates a high-dimensional possibility space to arrive at coherent form. The folded protein is the material trace of this traversal, analogous to a generated sentence, a danced phrase, or a poetic line. In all cases, form is not imposed from outside but emerges from the system’s internal negotiation with constraints.
What distinguishes AlphaFold is that its output is not symbolic or aesthetic but material. The geometry it predicts has physical consequences, shaping biochemical function. Yet the underlying logic remains consistent: linear input, vectorial traversal, emergent form.
Across poetry, AI, dance, and biology, we encounter the same structural principle: vectorial improvisation within a constrained space of possibilities. Each system operates sequentially, yet generates coherence by continuously updating its internal state in response to its own actions.
In each case, novelty arises not from randomness alone, but from directional deviation—a change in orientation within a structured field. The system must risk leaving the most probable path to discover new form.
Reintegrating William Carlos Williams’s poetic insight with contemporary computational and choreographic systems reveals a shared epistemology of form: thinking is movement, and form is the residue of traversal. Whether enacted by a poet, a dancer, a neural network, or a protein, creativity emerges through the sequential exploration of possibility spaces shaped by constraint.
LLMs make this logic explicit, translating linguistic intuition into vectorial computation. Forsythe renders it embodied, exposing the algorithmic intelligence of movement. AlphaFold extends it into matter itself, demonstrating that morphology is not designed but predicted—discovered through navigation rather than imposed through blueprint.
A new line is a new mind.
A new vector is a new orientation.
A new fold is a new form.
Across scales and substrates, cognition reveals itself not as depth, but as trajectory.
End of essay