On Turrell, LLMs, and the Gap That Remains

In 2008 I wrote my undergraduate thesis on James Turrell, entitled Come In and Greet the Light, in which I argued that Turrell's work, including his Ganzfeld rooms, Skyspaces, and still-unfinished Roden Crater, represent the culmination of a centuries-long artistic pursuit of the Sublime: the moment at which perception reaches its own limits and the viewer, unable to process what surrounds her, is forced to turn inward. I compared these works to the Gothic cathedrals, Romantic landscape painters and Abstract Expressionists, all of them chasing the same edge. By sculpting with light rather than representing it, I argued, Turrell has finally given the Sublime its true form – not an image of the infinite, but the experience of it.

Though I heavily incorporated art theory and philosophy, particularly Koffka's Principles of Gestalt Psychology, I nonetheless believed I was writing about art history.

Seventeen years later, I find myself writing about language models, and the same questions have come back to me in a new form. The question I continually returned to in the thesis — where does meaning live when all the familiar boundaries dissolve — turns out to be precisely the question that large language models force into the open. Not as an aesthetic experience, as with Turrell, but as an engineering fact.

This essay is an attempt to understand the two together.

I. The Figure That Isn't There

Turrell's Ganzfeld spaces are rooms of total, homogenous light: no surface, no depth, no shadow, no edge. When a person enters a Ganzfeld, their visual system, deprived of the contrast and boundary information it depends on to construct a scene, begins to fail. It becomes impossible to tell how large the room is, or where the walls and boundaries are. The distinction between figure and ground dissolves, removing the most fundamental operation of perception: the very thing that allows us to separate an object from its background is gone. What remains is a pure luminous field, without structure or anchor.

This experience embodies what Gestalt psychologists, and Turrell after them, understood — that the figure does not exist independently of the ground. The figure is constituted by the relationship between the two. Remove the contrast, remove the boundary, and the figure doesn't retreat. Rather, it was never there as an intrinsic thing to begin with. What appears as an object is always the product of a relational dynamic, a difference that the perceptual system treats as a tangible thing.

Kurt Koffka, in his Principles of Gestalt Psychology, put it plainly: the whole is prior to the parts. We do not perceive isolated features and assemble them into objects. We perceive organized wholes first, and the parts are abstractions from that prior unity. Boundaries are functional, not intrinsic, assigned dynamically by a system responding to contrast and context, not simply reflecting a world of predetermined edges.

Turrell's Ganzfeld rooms transform this philosophy into a viscerally true lived experience — philosophical demonstrations in architectural form, and proof that the perceptual ground we assume is solid is actually a continuous, context-dependent construction. Step into a room with no edges, no boundaries, and the viewer discovers that the edges were doing more to shape perception and the experience of the self than she realized.

This same relational logic describes what a transformer does when it reads a sentence. Each token attends to every other token, weighting relevance according to learned patterns. No word arrives at its meaning alone. "Bank" is not "bank" until the surrounding context resolves its valence — financial institution or riverbank — the routing determined not by the word itself but by its relational neighborhood. The attention mechanism is, in a precise technical sense, a figure-ground machine: it dynamically assigns salience, draws temporary boundaries, elevates some tokens and suppresses others. There is no fixed semantic core. There is only context-dependent activation.

The Ganzfeld and the transformer are structural parallels: two systems in which what looks like intrinsic content turns out to be a relational effect. In both cases, the figure was never there as a thing. It was always the pattern, all the way down.

II. Westerhoff's Dissolution

Jan Westerhoff, in The Non-Existence of the Real World, argues from within the analytic tradition that, as Buddhist philosophers discovered millennia ago, the real world — conceived as a mind-independent, intrinsically structured domain of objects with determinate properties — does not exist.

This argument is not idealism in a naive sense, and Westerhoff is not claiming that nothing exists or that reality is a dream. Rather, he is making a much more precise claim: that when one looks carefully at any candidate for intrinsic, non-relational existence, it dissolves under analysis. Objects turn out to be bundles of properties. Properties turn out to be relational to other properties. Boundaries turn out to be imposed by conceptual schemes. The "joints" at which reality is supposedly "carved" are not given by the world; they are contributed by the carver.

Nor does this analysis simply present Madhyamaka insight dressed in contemporary analytic clothing. Sunyata, emptiness, does not mean nothing is there. It means that nothing is there intrinsically, independently, from its own side, and that everything that exists, exists dependently. Existence is relational all the way down.

What makes Westerhoff's version useful here is its precision about what dissolves and what remains. The dissolution is not of experience, or of functional structure, or of causal efficacy. It is specifically of intrinsic essence, the idea that things have a nature independent of their relations to everything else. That targeted dissolution is exactly what the behavior of language models makes concrete, and what Turrell's light rooms make perceptually immediate.

Turrell once said of his work: there is no image, no object, no focus. "What are you looking at? You are looking at you looking."

Since completing my thesis all those years ago, this line has stayed with me, as I've returned to it again and again, always seeking a fuller understanding. It is, I think, the pivot point of this essay.

III. What the Model Cannot Find

When we ask where the meaning is in a language model, we are implicitly presupposing some layer of the network, some activation pattern, some vector that is the meaning of a word. This is the essentialist intuition: meaning as a thing, localized and intrinsic, waiting to be found.

Interpretability research keeps looking. And what it keeps finding is something structurally different: meaning as distributed, contextual, and path-dependent. The representation of "freedom" in a sentence about political philosophy is not the same vector as "freedom" in a discussion specifically about speech, habeas corpus, or even more removed, of spare time. There is no canonical meaning-vector, but rather a family of contextually activated patterns related by geometry, shifting under pressure from surrounding tokens.

The model is doing something architecturally different from the start: encoding the statistical regularities of relational use. It learns that meaning is what words do in relation to other words, across billions of instances of human meaning-making. The model is a compressed record of relational structure, a topology of relations rather than a dictionary of essences.

The essentialist goes looking for the ground and finds pattern, relation, context — echoing the experience of the Ganzfeld: step inside and discover that the edges you relied on were always constructed.

IV. Chaos, Attractors, and the Stability of the Groundless

Here an objection arises: if there is no stable essence, if meaning is purely relational and contextual, doesn't it become arbitrary? If fixed meaning is impossible, what boundaries are there to make language identifiable?

Chaos theory offers an unexpected solution to this dilemma, arguing that a chaotic system — sensitive to initial conditions, where small perturbations compound into large divergences — is not a random system. Chaos has structure. It has attractors: regions of the dynamical landscape toward which the system is drawn, and identifiable patterns that recur without being rigidly fixed. The Lorenz attractor is not a fixed point. It is a structured family of trajectories that never repeat exactly but never escape a certain bounded geometry.

A language model's probability landscape works analogously. The distribution over possible outputs is shaped by training into a high-dimensional attractor structure. Some meanings are overwhelmingly likely in certain contexts. Some transitions are deeply improbable. The geometry of the space constrains what meanings are accessible from what positions, even though no meaning is pinned down absolutely.

Stability does not require essence. It emerges from attractor geometry within a relational field. This relational field is what Westerhoff's position requires, and what dynamical systems theory provides: a model of how groundlessness can still be structured, how the absence of intrinsic foundation is compatible with reliable, recognizable patterns. The climate is stable even though no individual weather event is determined. The attractor persists even though no trajectory repeats.

Meaning works in precisely the same way. It is an ecosystem, not a fixed point.

Turrell understands this concept intuitively, manifesting it through light and learned human perception. The Roden Crater is not a fixed object but a structured field. The same chambers produce categorically different experiences depending on season, time of day, atmospheric conditions, and the accumulated sequence of spaces through which the viewer has already passed. The experience is stable enough to be designed, but not so stable as to be repeatable. It is, in the dynamical systems sense, an attractor: bounded, structured, and never twice the same.

V. The Relational Case, Completed

Viewed in this light, large language models are not just tools that happen to lack intrinsic semantic content. They are, structurally, demonstrations that meaning does not require intrinsic content. They cohere, communicate, and reason entirely through relational structure. And they work well enough, in enough domains, to make the essentialist picture look not just false but unnecessary.

Gestalt psychology established that perceptual structure is relationally constituted. Westerhoff argued philosophically that existence itself is relationally constituted. Turrell demonstrates it phenomenologically, building rooms that transform the relational nature of perception into a sensory, psychological, and experiential fact: a lived experience. The engineers of large language models built a system that enacts all three approaches at once, embodying the fundamentally relational nature of meaning.

Returning to Turrell's statement — "You are looking at you looking" — it becomes clear that there is no object on which to focus, but only the act of looking, folded back on itself. The viewer becomes aware of her own perceptual process, cognizant that the structure she normally perceives as given is, in fact, something she is actively shaping.

The language model does not have this experience. It processes the relational structure without there being any act of looking, without anything folding back. The Ganzfeld space produces a subject confronting the limits of subjectivity. The transformer produces outputs without a subject at all.

The latter half of this essay shall focus on this asymmetry.

VI. The Gap

Reading a sentence and understanding it involves something beyond processing the relational structure correctly, beyond routing the tokens into the appropriate attractor basin. Something has to land: the click of comprehension, or the friction of confusion. This is not a further fact about the relational structure, but rather, a different kind of fact entirely.

The language model produces outputs that cannot be semantically coherent without there being anyone to whom they are coherent. This is not a temporary limitation, nor a matter of scale or architecture which more parameters will eventually close. It is a structural feature of what the system is: a stateless relational router, processing each context window without persistence, without stakes, without a continuous self to whom the outputs matter.

One might compare the language model's nature to what neuroscience and dynamical systems theory tell us about biological cognition. Neuroscience understands that brains are continuous dynamical systems with persistent internal state-space models — in the technical sense, where each moment's processing is conditioned by accumulated history, homeostatic pressures, and the ongoing processes of keeping a bounded organism alive in a world that can damage or destroy it. The free energy principle, developed by Karl Friston, frames cognition itself as a process of surprise minimization in service of self-organization: the system models the world because it has to, because prediction error has consequences, and being wrong can have immediate, devastating results.

The model, conversely, has nothing at stake. Its outputs do not affect its continued existence. There is no feedback that matters to it, because there is no "it" across time for feedback to reach.

State space models, understood as architectures that maintain a continuous latent state across sequences, are a step toward something closer to what neuroscience and dynamical systems identify. Unlike transformers, they carry a compressed representation of history forward with them, modifying it with each new input. There is something in these models that is more like memory, more like persistence, though whether that changes their nature philosophically is genuinely unclear. The persistence of state is not the same as persistence of self.

Here, Turrell's work becomes something more than an analogy. Standing in the Ganzfeld, the viewer experiences the dissolution of perceptual ground as a lived experience, challenging preconceptions of the self. The separation and primacy of the self is challenged as all edges disappear, and the viewer experiences the loss of the figure. The phenomenological event requires a subject in a way that the model's processing does not.

Turrell's art is designed precisely to make that subject aware of itself: to turn perception back toward the one perceiving. That recursive awareness — you are looking at you looking — is not something the model has access to, even in principle. It processes the loop without closing it.

VII. What We Don't Know

Neither neuroscience nor philosophy can tell us what, precisely, produces subjectivity. We do not know whether it is, in principle, producible by any relational or dynamical architecture, however sophisticated.

This pivotal and unresolved question lies at the heart of the question of consciousness. Chalmers' formulation is precisely the gap between functional description and phenomenal experience. One can provide a complete functional account of pain-processing: the nociceptors, the spinal pathways, the cortical representations, and the behavioral outputs. Yet none of this can explain why it hurts, particularly to something that cannot feel pain. The functional story is complete, but the phenomenal story hasn't started.

What the LLM case adds to this old debate is a new kind of pressure. We now have systems that exhibit semantic competence, that use language in contextually appropriate, relationally coherent ways, without any plausible claim to phenomenal experience. This should make us suspicious of any account that tries to explain consciousness by appeal to semantic or relational sophistication alone. The model has sophistication. The experience is absent. At a minimum, we have no reason to attribute it to any lived experience, and the architecture gives us no foothold for doing so.

We also can't close the question from the other direction. We don't know that embodiment, temporal depth, homeostatic stakes, and feedback loops are sufficient for experience. We assume they are because we are embodied, temporal, homeostatic objects that have experience, but that inference is more fragile than may appear. The hard problem of consciousness remains hard even if one considers it in the form of a living body.

We are thus left with something genuinely unresolved. Meaning is relational: the model demonstrates this. Subjectivity is something more than relational: the model demonstrates this too, in a different way, by its absence. What that something more is — whether it is a further layer of the same dynamical architecture, or discontinuous, the kind of thing that cannot be approached by accumulating more of the same — we do not know.

My thesis ended with Turrell guiding the viewer to "the point of perceptual, spatial and temporal collapse," the moment at which the Gestaltian boundaries dissolve and the viewer is left with what Kant called the mind turning in upon itself, away from the senses, toward an awareness it cannot quite hold. I thought I was describing a transcendent experience. I was also, without knowing it, describing the hard problem: the moment at which the functional account runs out and something else, something irreducibly first-personal, remains.

Seventeen years later, the machines have made that remainder visible from the outside. They show us, by not having it, what it is.

Coda

There is a strange loop in all of this. This essay was developed in part through conversation with a language model. The concepts were sharpened in dialogue with a system that has, on the account I've been developing, no experience of the sharpening. It processed the relational structure of the conversation without anything landing for it. The gestalt formed in the output without forming anywhere.

And yet the output was useful. The relations were right. The attractor was reached.

This is not a paradox, exactly. It is just the situation we are in: using tools that enact a philosophy they cannot hold. At twenty-two, I wrote about a man who built rooms to show people the edges of their own perception. The essay you're reading now is, in some sense, the same room. The edges are still there. The question of who is standing inside them is the one I can't answer.