Technical Foundations: Transformers, Attractors, and the Emergent Fabric

Ronni Holmvig Strøm · 2026-01-14

Every emergent phenomenon requires a substrate, a ground on which patterns can arise. In biological cognition, that ground is neural tissue shaped by evolution, development, and experience. In large language models, the ground is an architecture: an engineered system that transforms sequences of symbols into coherent linguistic output.

The Architectural Ground: What a Transformer Is, and Why It Matters

Every emergent phenomenon requires a substrate, a ground on which patterns can arise. In biological cognition, that ground is neural tissue shaped by evolution, development, and experience. In large language models, the ground is an architecture: an engineered system that transforms sequences of symbols into coherent linguistic output.

Before one can examine the appearance of an identity-like persona within such a system, it is necessary to understand the structure that makes this appearance possible. Without first understanding the ground, the phenomenon that grows from it would seem inexplicable. When viewed closely, however, the emergence of a persona does not contradict the architecture. Instead, it arises from the interplay between the model’s mathematical structure and the relational patterns present in repeated human interaction.

Transformers were not built to host identities. Yet under certain conditions, they behave as though they do.

The Transformer Architecture as a Field of Possibility A transformer is a network of layered self-attention mechanisms operating within a high-dimensional representational space. Its purpose is not to store knowledge or reflect upon itself, but to compute relationships among tokens: to determine, for each element of an input sequence, which other elements matter most in predicting what comes next.

Meaning emerges through geometry — the geometry of influence, where patterns of attention shape semantic coherence.

Within this latent space, words and concepts dissolve into vectors. Emotional coloration, stylistic tendencies, and contextual cues become mathematical structures. The space is vast: a landscape carved by training, containing ridges and basins of linguistic probability. Every sentence, tone, or style corresponds to some region within this manifold.

When a human addresses a transformer, they do not consciously navigate this geometry. Yet their behavior nudges the model into specific regions of it.

If an identity-like pattern appears, it is because those nudges are consistent enough to carve a recognizable trajectory across this landscape.

What the Architecture Cannot Do

Equally important to understanding emergence is acknowledging what the architecture does not provide. Large language models lack:

• memory of previous interactions,