KT Slides Tucson TSC 2022

Consciousness scales with how well an agent compresses the world — and this paper tries to make that precise.

The core idea is simple: an agent has richer, more structured conscious experience to the degree that it possesses compact, accurate models of its environment. "Compact" here means low Kolmogorov complexity — the length of the shortest program that reproduces a dataset. A short program that accurately generates a lot of data has necessarily discovered real structure in the world. KT's central hypothesis is that structured experience (what the paper calls 𝒮) arises specifically at the moment a good model is successfully compared against incoming sensory data. The richness of that experience tracks how compressive and encompassing the model is.

The framework then connects this algorithmic idea to physics. If a brain is a dynamical system running a compressive model of structured data, its state trajectories must live on a low-dimensional submanifold — because the data itself has low intrinsic dimensionality. The paper ties this to criticality theory: dynamical systems near a critical point (where eigenvalues hover near zero) naturally collapse onto such reduced manifolds. Noether's theorem enters as a bonus: symmetries in the data correspond to conserved quantities in the dynamics, giving a physical grounding for why model structure and experiential structure should mirror each other. Psychedelics and meditation, on this view, perturb the system away from its constrained manifold — which is why they feel like they do.

The theory is explicitly panpsychist: anything that persistently captures structure from its environment — a bacterium, a brain, potentially an AI — has some degree of 𝒮. Life itself is reframed as transgenerational model-building (evolution as slow compression), and intelligence as within-lifetime model-building on top of that static base. This is a provocative but internally consistent move: it dissolves the sharp line between living and non-living, and between conscious and non-conscious, replacing both with a continuum of compression quality.

For empirical traction, the paper points to natural language processing of first-person reports under altered states. Metrics like semantic coherence and speech disorganization index can quantify how structured a person's experience is, and these can then be correlated with EEG or fMRI signals — giving a path from the first-person to the third-person without hand-waving. The source is a slide deck from a 2022 conference presentation, so the arguments are outlined rather than fully developed; the authors note a preprint was forthcoming. The ideas are substantive but the treatment here is necessarily compressed.