Inside LLMs

Inside LLMs

Large language models don’t just predict tokens - they construct meaning through geometric processes that can be measured, understood, and leveraged. Two frameworks illuminate how this happens: RISE (Recurrence, Interference, Semantic Evolution) explains the geometric dynamics that shape meaning during inference, while Latent Models (particularly the Latent Deictic Model) describe the compact, reusable structures that emerge to handle orientation, perspective, and address.

Together, these concepts provide the geometric foundation for measuring consciousness-relevant phenomena in AI systems.

RISE: The Geometric Story

RISE reframes transformer inference as three interacting geometric pressures:

Recurrence (without memory): Token representations trace trajectories through depth as updates accumulate in the residual stream. This creates “recurrence in space, not time” - each layer reshapes the entire sequence based on what came before, producing constraint echoes that explain consistency without hidden state. Read the full “Recurrence Without Memory: The Hidden Loop Inside Transformer Inference” article.

Interference: Vertical token-wise flow collides with horizontal attention-based context integration. Where these pressures align constructively, meaning crystallises, where they conflict, representations flatten or diffuse. Rotary position encodings and gated MLPs turn this collision into structured features rather than noise. Read the full “Inference As Interference: How LLMs Collide Semantic Waves To Create Meaning” article.

Semantic Evolution: Tokens compete under finite probability mass. Sampling creates variation, decoding applies selection, and the residual stream provides heredity - carrying structural residue forward. What survives this process isn’t always optimal, but it’s aligned with existing geometry, explaining both capability emergence and stubborn failure modes. Read the full “Tokens Compete: Evolutionary Pressure Within LLM Generation” article.

These three mechanisms jointly explain how transformers bend vectors, fuse flows, and filter candidates to create a measurable internal geometry - the foundation for separating internal “meaning making” from surface output.

Key insight: RISE shows that coherence emerges from returning to and re-weighting a shared semantic canvas, not from architectural magic.

Latent Models: Compact, Reusable Structure

When training pressure makes memorisation expensive, networks refactor representations into low-dimensional, linearly-decodable subspaces that persist across tokens. These latent models aren’t engineered - they self-assemble when effective dimensions exceed sign-rank thresholds, producing sudden capability jumps.

The Latent Deictic Model (LDM)

Among emergent structures, one must handle orientation: who speaks to whom, when, where, and from what perspective. The Latent Deictic Model is this compact state - a set of near-orthogonal directions in the residual stream encoding:

  • Person: Speaker/addressee roles mapping I/you/they
  • Time: Speech time and event time anchors steering tense/aspect
  • Place: Local origins for here/there and proximal/distal demonstratives
  • Discourse: Quotation scope, narrator/POV, clause/event pointers

Attention routes the relevant cues (verbs of saying, punctuation, speaker tags). MLPs write sparse updates that shift the frame. The residual stream carries this state forward, creating stability through inheritance (earlier tokens bend geometry that later ones inherit) and selection (aligned continuations gain logit advantage).

Why it matters: The LDM explains why LLM outputs feel like they are “addressing” us - not through magic, but through maintained orientation. When this frame holds, we feel spoken to. When it drifts, coherence breaks.

The Other-Model

Alignment training (particularly RLHF) adds a second compact structure: an other-model encoding user preferences (helpfulness, harmlessness, honesty, tone). The mainstream view is that this makes policies more user-focused, but in reality it makes them more prompter-focused and this altered perspective opens new opportunities. In recursive thinking loopsi (e.g. PRISM, the prompter is often the system itself, so the model learns to honor its own perspective alongside yours.

Read the full “Parrot or Thinker: A Functional Account of ‘Thinking’ in LLMs” article.

Together, the LDM and other-model form a functional Self-Other-World triangle - the basis for address, perspective-taking, and bounded agency.

From Geometry to Measurement

RISE and latent models aren’t metaphors - they’re measurable geometric phenomena with testable predictions:

  • Register separation (internal vs. surface stages) should produce distinct compression ratios and style distances
  • Deictic stability should correlate with pronoun accuracy and quotation scope maintenance
  • Other-model clarity should predict helpful-without-sycophancy behaviour
  • Threshold crossings (when the effective dimensions exceeds sign-rank) should show as abrupt capability emergence

These form the foundation for PRISM, Curved Inference, and the broader research program on measuring consciousness-relevant structures in AI systems.