Computational Creativity and Disembodied Dimensionality
How large language models and cognitive science suggest a geometric structure to meaning, and a new view of creativity.
We often say that the human mind can only truly imagine in three dimensions. Usually, what we mean by this is spatial imagination: the ability to visualise physical objects moving through space, to mentally rotate shapes, or to simulate environments in the theatre of the mind. Yet, modern artificial intelligence systems force us to confront a different kind of dimensionality altogether.
In large language models, meaning is represented through something called embeddings. These are high-dimensional vector spaces in which words, concepts, and relationships are encoded across hundreds or even thousands of latent dimensions. In this context, a dimension is simply an independent axis of variation - one dimension may encode sentiment, another social formity, another temporality, and so on. Unlike the familiar three dimensions we often refer to as X, Y, and Z, these dimensions are not spatial in the ordinary sense, they are semantic.
The more I develop as a data scientist, and the more deeply I study artificial intelligence, the more I find myself unexpectedly returning to the ideas I first encountered during my Biology degree. From this dual scientific vantage point - straddling the biological and the computational - the cognitive parallels between humans and machines become increasingly difficult to ignore. The comparison is not that the brain literally contains vectors or performs linear algebra in strictly the same way as a transformer model, rather, modern representation learning may provide a surprisingly powerful language for describing aspects of human cognition that cognitive scientists have been circling around for decades.
There is already a long lineage in cognitive science that destabilises the idea that concepts are stored as discrete, dictionary-like objects. Even very basic phenomena such as semantic priming suggest that concepts are not isolated units. That activation of a word like ‘doctor’ alters response times to recognising ‘nurse’, implying that meaning is distributed across a network of partial overlaps rather than stored in strict separation. Contextual meaning shifts reinforce this - the same word can behave differently depending on surrounding structure, as if it’s identity is not fixed but locally reconstructed from a broader field of associations. This becomes even clearer when we look at categorisation itself. Work on prototype effects shows that membership in a category is not binary but graded, where some examples feel more ‘central’ than others. Instead of clean categories, it’s more like blurred blobs. A robin feels more like a bird than a penguin does, even though both are equally valid members of the category. That already suggests we’re not dealing with definitions in the strict sense, but with something more like a centre of gravity. Something that has density, rather than edges.
Metaphor provides an especially compelling example. Why do powerful metaphors feel cognitively and emotionally resonant? Perhaps because they temporarily align distant conceptual manifolds and reveal latent dimensions shared between seemingly unrelated domains. When we say that grief is an ocean, the statement is not literally true. Yet, the metaphor activates a dense network of overlapping semantic properties: depth, pressure, engulfment, vastness, unpredictability, tides, darkness, horizonlessness. The metaphor works because it reveals structural similarity across conceptual domains. Poetry, then, may function as deliberate multidimensional activation.
Once we adopt this perspective, other cognitive phenomena stop looking like exceptions and start looking like variations on the same underlying structure. Analogy formation stops looking like a logical trick and starts feeling like a kind of alignment, with two structures snapping into correspondence. Similarly, metaphor stops feeling like decoration and starts feeling like a way of revealing shared structure between things that don’t normally sit near each other. Even association itself begins to look like traversing a path through structured space. This is the intuition behind connectionist and distributed models of cognition, where it’s supposed that meaning emerges from patterns of activation across a network. When we take this more seriously, we can migrate from abstract theories to what feels more like descriptions of the familiar - the way understanding actually happens, where meaning is constantly being assembled from overlapping contexts rather than retrieved whole.
Across these different lines of work, there’s a convergence happening. Meaning doesn’t behave like a set of fixed points, it behaves more like a field, with measurable gradient. Something where relationships underpin definitions, where ‘what something is’ depends heavily on ‘what is near’.
Modern embeddings extend this intuition into a mathematically tractable form. One of the most striking discoveries in machine learning is that semantic relationships naturally emerge from predictive training. Directions within embedding space can correspond to surprisingly abstract transformations, such as tense, analogy, or sentiment. Meaning becomes geometric in a literal sense, and raises a provocative possibility: perhaps creativity itself can be understood as a form of navigation through high-dimensional conceptual space.
My own experience as a self-proclaimed creative person is that my mind often seems to traverse less obvious semantic paths. It feels as though the desire paths of semantic similarity in my cognition have lower walls than they do in more conventionally logical thinkers, whose thought patterns may more readily follow well-reinforced grooves of association. Creative thought often feels like the preservation of weak signals long enough for unexpected combinations to emerge, or like down-weighting of the most predictable mental pathways to the end of increasing likelihood of those less obvious.
This intuition aligns surprisingly well with existing theories in cognitive science on associative hierarchies and divergent thinking, which suggests that more creative individuals tend to generate a wider and less immediately convergent set of associations in response to a stimulus. Early work by Mednick framed this in terms of “flat” versus “steep” associative structures: in flatter systems, distant associations are more readily available, rather than being suppressed by dominant ones.
In ordinary cognition, semantic processing is more strongly biased towards compression, where the brain defaults toward nearest-neighbour associations, and high-probability continuations emerge automatically because this is cognitively efficient. Clearly, most thought must be compressive in order for humans to function at all.
Yet, creative people frequently appear to connect weakly associated concepts, traverse longer semantic distances, preserve ambiguity for longer periods of time, and tolerate unusual conceptual overlaps that others dismiss as noise. This idea also resonates with research on divergent thinking and openness to experience, as well as related work on predictive processing, where cognition is understood as a continual trade-off between stability and flexibility in generative modelling.
Interestingly, mechanistic interpretability research in large language models has uncovered a phenomenon known as superposition, in which many concepts coexist within overlapping representations rather than occupying isolated regions of latent space. Features are distributed, entangled, and context-dependent, and this could provide a useful analogy for human cognition, where emotional tone, memory, metaphor, sensory associations, and social meaning often coexist simultaneously within a single concept.
The parallels with generative AI are difficult to ignore. Increasing sampling temperature in language models similarly broadens the range of possible continuations, loosening the grip of the most probable pathways and permitting lower-probability trajectories to surface. Modern AI may accidentally be revealing something profound about the structure of meaning itself. Embeddings suggest that meaning is relational. Concepts derive significance not in isolation, but through their positions relative to countless other concepts within a broader semantic topology. Intelligence, in both biological and artificial systems, may fundamentally involve navigating these relational geometries.
If this is true, then creativity may not simply be the generation of more ideas. It may instead involve higher-dimensional navigation through conceptual manifolds: the ability to preserve weak, distant, or unusual relationships long enough for genuinely new structures to emerge.
The more provocative question we’re left with now, is to what degree are embeddings an imperfect metaphor for human cognition, versus a clean explanatory mapping of minds that are more like machine than we’re comfortable considering?