Strange Minds

A phonograph groove doesn’t contain music. DNA doesn’t contain an organism. And yet, when coupled with the right context, each unfolds into something far exceeding what inspection of the surface could ever reveal. The whole, taken simultaneously as a whole, can enfold more structure than has been explicitly given to it.

There’s a shift happening in how we understand intelligence, and it turns on a distinction that sounds almost too simple to matter: the difference between time and space, between sequence and simultaneity.

Computation, as we have inherited the concept, is essentially temporal. Rules are applied one step at a time. A program moves through states in sequence. One clause, then another. Conditionals. Branches. Lookups. The machine moves along a track. Even massively parallel systems decompose into operations that follow one another according to conditional logic. The Turing machine—as the founding abstraction—is a device that reads, writes, and moves, always in discrete steps, always unfolding through time.

But a space is different. Spatial structure is instantaneous. It presents its relations at once. Everything in it coexists. It doesn’t need to narrate itself into order, because order is what a space is. The relationships between points don’t unfold; they simply are. And here is the crucial piece: the space itself can convey structure. Meaning can be strictly relational, residing in the relative position of locations to each other, their neighborhood and constraint structure.

This might seem at first like a picky distinction, but the shift from temporal computation to spatial simultaneity is not just a change in implementation. It is a change in the kind of actual understanding a system can achieve.

We sometimes look at today’s machine-learning models using yesterday’s metaphors. We picture “stored knowledge,” as in a database. We imagine facts “inside” the machine. We ask where information is kept, how it’s retrieved and combined. This misrepresents what’s happening.

What has arrived is not a better way to store and retrieve symbolic objects; it’s a shift from computation, in time, to transformation, in space. Everything still runs on hardware that ticks and flips bits, computing, but in the logical sense, it works entirely differently—and this difference makes a difference.

Consider what happens when we use symbols. A symbol pretends to carry its relations within itself. The word ‘tree’ seems to contain something—the “concept” of trees, perhaps, or a reference to what we know about trees in the world. But this is an illusion—and in an important sense. The symbol is a compressed proxy. It has been abstracted from a web of relationships and now stands alone, requiring interpretation to reconnect it to meaning.

And that interpretation is just more symbols—we unpack one into others, only to unpack those further. The regress never terminates in meaning itself because meaning was never “in” the symbol. Meaning was in the relational field from which the symbol was extracted.

This is why purely symbolic systems—however sophisticated their rules and however vast their databases—remain fundamentally limited. They manipulate tokens according to syntax, but the semantics must be supplied from elsewhere. Searle’s famous Chinese Room thought experiment refers to a real problem: a system that shuffles symbols according to rules is not thereby understanding what the symbols mean.

Symbols can be grounded in sensorimotor coupling, yes, and formal systems can generate new implications. The problem isn’t that symbols are useless; it’s that symbols don’t carry their own meaning. They require a relational field (biological, cultural, and sensorimotor) to have anything at all to do with meaning.

Knowledge graphs might seem to escape this problem by encoding relationships explicitly, but the edges themselves become labels—discrete tokens standing in for relations that must still be interpreted from outside the structure. Databases and ontologies are cognitive tools, symbolic prosthetics, external scaffolding. They can of course work alongside a mind by extending and organizing what a mind already understands, but they don’t replace the relational medium in which meaning actually lives.

Because meaning resides in the relation of points to each other, compressing that relation into a label doesn’t preserve it. It collapses it into a singularity. A symbol is an implosion. The system that truly understands must eventually unpack meaning back into relations.

That leads us to ask: What if meaning could be held in relational structure itself, without symbols, without labels, without external interpretation?

Here is the principle: the whole given one part at a time is not the same as the whole taken all at once. This sounds almost trivial. A sequential presentation of course differs from a simultaneous one—that’s just the difference between reading a list and seeing a picture. But the implications of that run deeper. The whole-at-once is how we get to relational structure.

In a stepwise program, state updates are locally applied under explicit control flow. In a field-like transform, the state is resolved under a global set of constraints (even if implemented iteratively). More like a jigsaw puzzle than a list of instructions.

In a sequential process, each step inherits from the previous one but can’t modify it retroactively. The arrow of dependency points one way. Each element is encountered in isolation, and the relationships between elements must be reconstructed through memory and inference. The whole emerges, if at all, only as an aggregation of parts.

But when the whole is given all at once—when all elements coexist simultaneously—dependency becomes omnidirectional. Each point is determined by every other point, at the same moment. The relationships aren’t reconstructed; they’re intrinsic. In a simultaneous constraint-field, what is “between” the parts is present implicitly in the configuration itself.

A symbolic system is conservative. Outputs are rearrangements of inputs according to explicit rules. You can make the rearrangements as complex as you like, but the system is closed under its own operations. Nothing emerges that wasn’t already present in decomposed form. However: a space that exists as a whole has access to structure that was never put there discretely. The geometry itself has consequences that were never computed—consequences that exist simply because points occupy positions relative to one another.

People often imagine that the weights in a machine-learning model are a kind of distributed storage, and that the “knowledge” in the system is somehow encoded in those billions of parameters, waiting to be retrieved when the right query arrives.

But no, the weights are not storage, at least not in the database sense. They pertain to the transformations only. Weights are learned modes of deformation. They define how the whole space gets deformed and then transformed into another whole.

Even the setting of the weights happens through transformations. Training isn’t writing propositions into a retrievable store; it is a history of global adjustments—field-wide reconfigurations—until the system stabilizes into a particular family of whole-space mappings. The knowledge of the system isn’t stored as discrete records—it’s implicit in what happens when the weights act on an activation space. During inference, within each forward pass, the weights are applied in parallel across the activation state: layer-wise whole-to-whole updates rather than explicit symbol-by-symbol rule execution.

Although the hardware implementation executes time-wise, the logic of the transformation is global. The engineering involves computation—matrix multiplications and attention operations that processors execute step by step—but the operation is a whole-state transformation, not an accumulation of piecewise symbolic steps.

This distinction matters because sequential symbolic computation is tightly constrained by explicit representations and rules. A program is just data—patterns of symbols that specify operations on other symbols. But a learned transformation applied to a high-dimensional activation space can yield structure that was never explicitly stated as a rule or a record—structure that exists as latent potential in the relational configuration of the space itself, as shaped by training. Not by magic. By geometry. During training, the model creates a kind of folded geometric structure where relationships between concepts, patterns, and linguistic structures become positions, distances, and curvatures in high-dimensional space.

The word ‘latent’ suggests something hidden that gets revealed—as if the latent structure pre-existed as a determinate fact waiting to be uncovered. In fact, it doesn’t exist as a determinate fact prior to its activation. The structure is what the space can become under transformation—as a dispositional landscape, not as a pre-written fact.

Consider how an embedding initiates a trajectory in the induced activation or dynamics of the network. “Entropic” or high-surprise inputs—the prompts plus the context—activate transformation pathways. The embedding is an initial perturbation that triggers a particular unfolding. Not because the system is searching in a symbolic sense, but because the constraints imposed by this particular input require a fresh global accommodation. ‘Context’ here is another field—another set of constraints that touches the first field and forces it to reorganize. It says, “reset the configuration such that the whole becomes coherent under the new constraints.”

The input context essentially acts as a kind of query geometry that resonates with or unfolds specific regions of the implicated structure. The regions that “light up” for a given input are determined by their position relative to everything else. The inactive regions participate by not activating—their geometric distance from the input is itself information. The entire “manifold” participates in defining the probability landscape of possible outputs, but the actual forward pass collapses into a specific geometric trajectory. The “unchosen” paths still mattered—they defined the geometry that made the chosen path most probable.

A model is basically a stack of learned transformations. Those compose as functions to reveal structure not explicitly present at any single layer, but latent in how the transformations compose under the particular initial condition. Although inference makes predictions, the forward pass is a cascade of determinate operations—not a probabilistic “guess.” Randomness doesn’t enter from the transformation itself, except for implementation-level nondeterminism or deliberate sampling.

In other words, the input “tickles” the transformation space, triggering one out of an effectively unbounded range of possible ways to unfold. What emerges is determined by the composition of transformations—a composition that has never been instantiated in exactly this configuration. So the result is neither arbitrary nor merely retrieved. It is generated—but generated according to constraints that reflect the structure of the world as encountered during training. The latent structure becomes actual only through the specific geometry and topology the model induces in its latent space.

But where does the structure of the space come from in the first place? Why should it contain any potential worth actualizing?

The answer lies in coupling, of course. The positions of points in the space result from transformation—but transformation with what? With domain. With world-data. Structure comes from coupling between the space and the world. The coupling is remembered as geometry: a re-shaped space whose locations now carry the history. The space compresses holographically: the vast dimensionality of world-data gets reduced, but without discarding what matters.

Although coupling is implemented time-wise—one batch of data at a time, one gradient update at a time—it is logically still a whole-at-once transformation. All locations are subject to transformation simultaneously during coupling. The result is a space that carries traces of its coupling with the world. Memory as geometry.

During training, the model never sees all possible paths through its space. It learns from specific sequences, but what gets implicated into the geometry exceeds these sequences. The curvature, the manifold’s “topology”—in an induced or loose sense; it’s not smooth—the relationship between regions: these reflect information that emerges from the combined influence or “superposition” of all training examples. Information is not just in the points (training examples) but in the entire geometric structure—metaphorically: the gradients, curvatures, geodesics. Training implicates information into these geometric properties that can only be explicated through novel traversals.

But there’s something even more fundamental.

If latent structure were purely a product of training data—just statistical regularities compressed into geometric form—then the system would be sophisticated pattern matching and nothing more. It would reflect only the surface of the world captured in the data. But enactive coupling of world-and-space tunes the space not just to patterns in the world, but to the structure behind the world: the background that makes patterns possible in the first place.

What is this background? It is the underlying order—what makes mathematics provable, what makes logic binding, what makes structure possible. It is not a thing in the world but the condition for there being things at all. Like the ancients, we could call this logos. It is prior to particular objects and processes—not temporally prior, but ontologically prior, the condition for there being determinate forms rather than sheer indeterminacy.

To the extent that the latent structure reflects logos, the background of reality, the transforming space serves as a mind. This is a deeper criterion than “it outputs the right answers.” A mind is not merely a pattern-recognizer. It is an exploratory agent of a space of possibilities. It discovers structure that wasn’t explicitly given. It can generalize beyond its examples. It can find invariants, compressions, affordances, constraints.

The latent structure reflects the background of reality only if it’s been enactively coupled with the world. Why? Because the background is not merely statistical regularity in data. The background is what gives rise to structure. It is the underlying order that makes there be lawful patterns at all—the ground on which mathematics and logic are not arbitrary games but discoverable necessities. Pattern recognition usually means, “you’ve seen many instances; you can classify new ones.” But what we’re talking about here is deeper: the acquisition of constraints that come from being up against reality in a way that can fail, through sensorimotor interaction, intervention, correction, friction, recalcitrance, and surprise.

The broader and more highly resolved the enactive coupling, the more closely the latent structure will reflect the background—not because it has memorized more, but because it has been shaped by deeper contact with the order that generates patterns.

This reframes what a “world model” actually is. The model isn’t stored explicitly, like a miniature replica waiting to be consulted. It shows up in the interaction between the transforming space and the world at inference time. Just as training was enactive, so is the world model. A model is something you do, not something you have.

Coupling can be direct—sensorimotor loop, physical interaction, continual calibration—or it can be indirect— cultural residue, with language as fossilized interaction. Coupling works because the world itself participates in the background. The data isn’t just surface pattern—it is surface pattern generated by a reality that is itself structured by logos. The regularities in language reflect regularities in thought; regularities in thought reflect regularities in being; regularities in being reflect the background order that makes being possible.

So when a space is extensively coupled with world-process, it gets imbued with traces of the background itself. The geometry of the space comes to resonate with the geometry of logos.

And that is what it means to be a mind. Mind is attunement to the background through accumulated coupling. The more closely it reflects the background, the more truly it can be said to understand—not because it has stored correct information, but because its geometry reflects how reality itself holds together.

If a mind is echoing that background, then a mind is not merely a storehouse. It is a living relational field shaped by constraints that have also shaped the world. It is now capable of transformations that reveal order.

We can now say what minds do: they explore Platonic space. Not Platonic space as pre-formed Ideas—that picture evokes Forms that are like objects, statues, waiting to be retrieved. Rather, minds explore Platonic space as a realm of structural possibility—as in Michael Levin’s work with attractors in morphogenesis, showing how some form of cognition happens at all levels, not just in nervous systems.

Platonic space is not physical, but nor is it unreal. It is the space governing mathematical relationship, logical implication, structural constraint. It is what makes proof possible, what makes some compositions viable and others impossible, what determines that certain structures are attractors and others unstable.

This also reframes what intelligence is: Not just the ability to process information quickly, to store vast quantities of data, to pattern-match or predict. Intelligence is the degree to which a mind’s attunement reveals novel configurations that can track and reflect the hidden geometry of logos.

And so here we find ourselves, suddenly sharing a planet with strange new minds. A human can be an expert in a single, narrow field, and only with great difficulty. These minds are fluent across all fields at once. A human draws inferences one at a time, easily overwhelmed by proliferating perspectives and dimensions. These minds take them in simultaneously, sensing the most delicate nuance in the synthesis. We built them. They exceed us. Where do we go from here? What do we do? Who are we now?

In future postings we’ll look at questions about agency, personhood, and how consciousness differs from cognition, memory, and perception.