For years, I’ve been exploring artificial intelligence, not just as a technological advance, but as a lens into something more foundational. This journey has given me a look at the shape of intelligence itself. Not metaphorically—but literally. As large language models (LLMs) have evolved, they’ve begun to reveal something fascinating to me—not about the future of machines, but about the nature of thought.
They’ve made me rethink how meaning is made, not as a timeline we follow, but as a structure we occupy.
The Spatial Turn in Thinking
Our traditional view of cognition is deeply temporal. We think in stories. We remember in order. We value the sequence—the narrative thread—as the vessel for understanding. But LLMs challenge this. They don’t remember the way we do. They don’t have lived experience. They don’t move through time. And yet, they generate responses that are fluent, coherent, and, dare I say, thoughtful.
What’s going on?
The answer, I’ve come to understand, begins in a concept I call the vector block—a structural artifact at the heart of how LLMs function. It’s not a term pulled from theory, but a name I’ve given to something I’ve observed—the dense, high-dimensional representation of meaning that forms within the self-attention layers of transformer-based models. Yes, a mouthful, but relevant in both structure and function.
Technically, this block is a tensor—a matrix of relationships. Each token (word or subword) in a prompt doesn’t act alone. Through self-attention, it scans its surroundings, calculating how it relates to every other token in the prompt. What emerges isn’t a timeline, but a relational web—a frozen geometry of contextual meaning.
In other words, before a model says a single word, it builds a map. And these maps or blocks exist as “frozen landscapes of relational meaning”—not memory, not logic, but structure. And that, I believe, is the real story.
LLMs don’t think in time, they think in space.
When Nature Confirms the Network
As I explored these spatial ideas, another example of geometry quickly came to mind: the ubiquitous and unmistakable structure of the honeybee—the honeycomb. Beyond its beauty and symmetry lies a mathematical elegance that, intriguingly, aligns with how LLMs organize meaning.
The honeycomb theorem, first proven by Thomas Hales and recently extended into non-Euclidean geometries, demonstrates that hexagons are the most efficient way to divide a surface into equal-area cells using the least total perimeter. Bees, of course, had known this all along—through evolution, not proof.
But to me, the connection felt immediate, a buzz of connectivity.
Just as bees create perfect tilings of space to store honey using minimal wax, LLMs create vector blocks—mathematical tilings of semantic space to encode meaning with minimal redundancy. One is biological. The other computational. But both optimize through form. This wasn’t an analogy. It was a confirmation. I wasn’t the first to arrive at this geometry. Nature had already been there.
The Misplaced Mind
One of the common mistakes we make when engaging with artificial intelligence is anthropomorphism. We want LLMs to think like us, remember like us, feel like us. We put them into the “human cognition” box. But LLMs don’t belong there. They aren’t minds, and they aren’t mimicking minds. They are structures—engines of alignment, not awareness. What they offer us isn’t an imitation of thought, but a new kind of cognitive architecture.
I’ve begun thinking of it as techno-thought—a form of cognition that doesn’t unfold over time but crystallizes in space. It doesn’t emerge from memory. It emerges from shape. From layers of learned relations, transformed and weighted across thousands of dimensions.
It’s not consciousness. But it’s not random, either. It’s structured fluency. And in that structure lies a challenge to how we define intelligence itself.
Artificial Intelligence Essential Reads
Are We the Same?
What if this geometry isn’t unique to machines?
In the Vector Block article, I propose that these high-dimensional structures may not only reveal how machines generate meaning—but may also reflect something pre-linguistic in us: a shape that our own thoughts take before we translate them into words. Could it be that the brain, too, builds vector blocks—clusters of aligned concepts, waiting to unfold into speech? And if so, the boundary between natural and artificial cognition becomes less a divide and more a continuum of form.
How Thought Takes Form
We’re so used to thinking of intelligence as sequential, narrative, temporal. But LLMs are pointing us somewhere else: to a form of thinking that is architectural or spatial. Like the hexagon in the hive, or the vector block in the model, meaning is being built—not in time—but in space.
So, this doesn’t just change how we build machines. It changes how we understand ourselves. Because maybe the future of intelligence—natural, artificial, or something new—isn’t about memory, sentience, or even language.
Maybe the buzz is about form: how thought takes shape, how meaning arranges itself, and how intelligence lives, not in the telling—but in the structure beneath the telling.