CALLIOPE · BCOM

Kolmogorov Theory (KT) proposes that minds are best understood as algorithms that compress, evaluate, and act. This primer is the elevator pitch: in roughly two pages, we sketch the central insight (compression as cognition), define the algorithmic agent in terms of three minimal modules,---,a Modeling Engine (ME), an Objective Function (OF), and a Planning Engine (PE),---,explain why every successful regulator must contain a model of what it regulates, and outline what this framework buys us in neuropsychiatry, AI safety, and the science of subjective experience. The aim is to give a working scientist enough of the architecture to read deeper into the corpus without first wading through the formal machinery.

Minds are compression engines — and this two-page primer explains exactly what that means and why it matters.

The central bet is simple: understanding something and compressing it are the same operation. When your brain recognizes a face, it isn't storing a pixel grid — it's finding a short program that generates the pixel grid. Kolmogorov complexity (equation 1 in the paper) formalizes this: the complexity of a string is the length of the shortest program that produces it. A mind that "gets" something has found a short generator. A mind that doesn't is stuck quoting the data verbatim. This isn't a metaphor — it's a definition, and it applies equally to bacteria, brains, and language models.

From that single idea, the paper derives a three-module architecture for any agent. A Modeling Engine (ME) compresses incoming sensory data into a compact internal model of the world. An Objective Function (OF) maps that model state to a single scalar — positive for good, negative for bad — collapsing all competing drives into one number so that action selection is even possible. A Planning Engine (PE) simulates candidate actions, picks the one that maximizes expected valence, and closes the loop back through the world. The thermostat example is deliberately provocative: a bang-bang thermostat has all three modules in their minimal form. That's not a joke — it's the proof that the agent category is non-empty and well-defined, with richness as the only variable that scales from thermostat to human.

The paper also gives a distribution-free restatement of the Good Regulator Theorem (originally Conant & Ashby, 1970): any system that reliably makes the world more compressible must share algorithmic information with that world. Modeling isn't a design choice — it's a mathematical consequence of sustained competence. Every bit of missing shared structure costs a factor of two in posterior support for the regulator's success.

The framework then reaches into experience. The paper's Central Hypothesis is that structured experience — what it's like to be an agent — tracks how compressive, broad, and accurate the ME's models are. Emotion maps cleanly onto the three modules: the ME supplies what an emotion is about, the OF supplies how it feels, and the PE supplies its urgency. This is testable: measure model compressiveness and breadth, then check whether reported experience varies as predicted. Psychedelics, the paper notes, transiently flatten the model landscape (measurable as elevated Lempel-Ziv complexity in EEG), which is exactly what you'd want when the existing model is the pathology — as in depression, framed here as a ME stuck in a low-valence attractor.

The paper is explicitly a primer — short, non-technical, pointing toward the formal machinery in WP0062 and P13. It earns its brevity: the three-box loop is genuinely the whole spine, and everything else follows.