The Biological Foundations of Compression
All enduring systems are physical systems, and energy is the most honest feedback there is.
Every living process unfolds under a simple condition: to persist, a system must predict; to predict, it must compress.
In biology, this necessity is enforced, not chosen. Each cell, network, or organism exists far from equilibrium, expending energy to maintain order against entropy. That energy cost acts as a continuous audit: models of the world that waste energy or mispredict reality collapse. Only representations efficient enough to sustain function survive. Compression and truth become coupled through consequence.
This coupling is what I describe as epistemic entropy: a measure of how uncertain a system remains about the structures that sustain its own persistence. Reducing epistemic entropy is not an abstract cognitive act; it is metabolic. Every adaptive response, from enzyme regulation to cortical prediction, narrows uncertainty about which actions will keep the system within viable bounds. Biology therefore embodies a form of continual inference: energy expenditure translates directly into information about what is real.
The comparison to reality is not philosophical; it is physical. A neuron misfiring or a cell signaling incorrectly pays an energetic price. Each interaction with the environment becomes a test of correspondence between internal model and external cause. Over time, the systems that survive are those whose compressions best mirror the world's causal grain. The result is evolution itself: a history of successful approximations to truth.
Artificial systems operate under analogous constraints, though the feedback loops differ in directness and timescale. Computation consumes energy; training costs scale with model complexity; inference requires power. MDL principles apply: shorter programs that capture regularities are favored not merely for elegance but because they reduce the total description length—and thus the resources required to represent and deploy them. The distinction is not that artificial systems escape energetic constraint, but that the coupling between prediction error and immediate existential consequence is mediated differently. A neural network's misclassification doesn't directly threaten its own persistence the way a failing kidney does.
Yet the principle remains: compression coupled to feedback converges on truth. Whether the feedback arrives as metabolic crisis, reproductive failure, or optimization pressure, systems that persist are those whose internal models economically capture the structure of their environment. Biology demonstrates this coupling at its most direct and unforgiving. It reminds us that knowledge is never free: every coherent representation, from a protein fold to a trained model to a thought, is paid for in energy. Compression is not merely efficient. It is the signature of any system forced to predict under constraint. And constraint, energetic and informational, is what binds representation to reality.
