On Subsistence and Intelligence

This paper presents a unified theory of self-subsisting agents and self-improving intelligence. To illustrate, consider a computational artificial life system wherein individual agents self-maintain and self-replicate in a world governed by time-reversible and computationally universal dynamics. By construction, random noise is constantly injected into the world. This requires the agents to continually remove random states to maintain their identities against error and to make space for newly created agents. Under time-reversible dynamics, random states can be removed not by free, irreversible erasure, but only by exchange with existing regular states. This further requires a constant source of regular states, without which all existing ones would soon dissipate and the closed system would reach statistical equilibrium. Thus, by design, the agents are supplied with a constant stream of new data, such as real-world texts. From this source, regular states can be continually extracted by losslessly compressing the incoming data. Crucially, the agents are not optimized by external processes. Instead, they possess the complete codebases encoding themselves, which span features, learning algorithms, and all meta-levels above. On this basis, they produce candidate self-improvements at any of these levels, with all knowledge ultimately measured by subsistence. This establishes an unrestricted improvement potential. Coupled with it, the challenge of compressing complex real-world data creates a selective pressure that drives open-ended improvement.This artificial life system substantiates a philosophical picture in which agents indefinitely self-subsist and unboundedly self-improve as part of a natural world characterized by a finite set of metaphysical elements. The paper explicates this picture and explores an optimistic prospect for our future.