Year: 2026
Type: preprint
Abstract: Thermodynamic accounts of intelligence have advanced rapidly, but they do not yet resolve a prior definitional problem: what makes internal structure intelligent rather than merely ordered? This paper argues that any thermodynamic intelligence metric should include a usefulness filter. The numerator must track internal structure that is relevant to viability, prediction, and control, not merely raw information or benchmark score. Intelligence is therefore defined as the resource-bounded capacity of an adaptive system to acquire, compress, and deploy useful internal structure so as to improve prediction, control, and robustness across a range of environments. Capability and efficiency are separated. Capability is adaptive reach, A, defined over a reference distribution of environments or task families. Efficiency is acquisition efficiency, I_eff = Delta C_u / W_diss, where C_u denotes useful internal structure and W_diss physically dissipated work. Concretely, C_u is the predictive and control-relevant information that a system's internal state carries about viability-relevant environmental structure, making it a semantic rather than purely syntactic information term. I_eff is a learning or adaptation metric, not a complete scalar measure of deployed intelligence: trained systems at steady state may have Delta C_u = 0 while still exhibiting high adaptive reach. A toy numerical example with exact Landauer-referenced dissipation costs, explicit boundary conditions, temperature-normalised reporting conventions, and falsifiable comparisons against size-based and energy-only baselines are provided.
Source: arXiv (Cornell University)
Author: Djordjevic, Boris
Language: English
Topic: Complex Systems and Dynamics
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221801052