A Quantitative Definition of Intelligence

Abstract: We propose an operational, quantitative definition of intelligence for arbitrary physical systems. The intelligence density of a system is the ratio of the logarithm of its independent outputs to its total description length. A system memorizes if its description length grows with its output count; it knows if its description length remains fixed while its output count diverges. The criterion for knowing is generalization: a system knows its domain if a single finite mechanism can produce correct outputs across an unbounded range of inputs, rather than storing each answer individually. We argue that meaning over a domain is a selection and ordering of functions that produces correct outputs, and that a system whose intelligence density diverges necessarily captures this structure. The definition (1) places intelligence on a substrate-independent continuum from logic gates to brains, (2) blocks Putnam's pancomputationalist triviality argument via an independence condition on outputs, and (3) resolves Searle's Chinese Room Argument by showing that any finite rulebook handling an infinite domain must generalize.

• The paper introduces intelligence density, quantifying generalization through the ratio of logarithmic output count to description length.
• It distinguishes between memorization and true knowing by analyzing scaling behaviors in systems like neural networks and lookup tables.
• The framework resolves classical debates and offers substrate-independent criteria for evaluating a system's computational generalization.

Quantifying Intelligence: Formalizing Generalization via Intelligence Density

Introduction

Kang-Sin Choi's "A Quantitative Definition of Intelligence" (2604.10873) presents a formal, operational, and substrate-independent metric for intelligence in arbitrary physical systems. The work distinguishes between memorization and genuine generalization, and provides a foundation for evaluating intelligence not as a binary, behavioral property but as a continuum grounded in the scaling relationship between a system’s descriptive complexity and its capacity to produce independent outputs. The core contribution is the notion of intelligence density—the ratio of the logarithm of a system’s independent outputs to its description length—and the rigorous, asymptotic characterization of "knowing" as opposed to mere memorization.

Formal Definition and Theoretical Framework

The intelligence density $\mathcal{I}(S)$ for a physical system $S$ is defined as

$\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$

where $C(S)$ is the system's total description length (in bits), and $N(S)$ is the number of independent outputs producible in response to distinct inputs. Independence is characterized via Kolmogorov complexity: two outputs are independent if there is no algorithmically shorter description of one from the other.

The operational criterion for "knowing" a domain $D$ is that, as domain size parameter $n \to \infty$, $C(S)$ remains finite while $\log_2 N(S)$ diverges; i.e., $\mathcal{I}(S, n) \to \infty$. This distinguishes systems that generalize (those capable of generating correct outputs for an unbounded range of inputs using a fixed mechanism) from those that only memorize (where the system's description length grows with the coverage of new outputs).

This criterion is crucially substrate-independent and abstracts away from implementation details: it applies equally to neural tissue, silicon, or documentary rulebooks.

Philosophical Implications and Argumentative Strength

The framework directly addresses and resolves several classical debates:

• Searle's Chinese Room: Any finite rulebook or symbolic system capable of passing a Turing Test across infinite domains (e.g., Chinese arithmetic) must be algorithmic, not a lookup table. The intelligence resides in the generalization capacity of the rulebook, not the symbol manipulator, thus formalizing the location and nature of "knowing" in the system and resolving the Systems Reply ambiguity.
• Block's Blockhead: Lookup tables for linguistically rich domains are physically unrealizable due to exponential growth in required storage. The vanishing of $S$0 as domain size grows quantitatively refutes claims that such design can constitute intelligence.
• Putnam's Pancomputationalism: The requirement that outputs be independent (grounded in Kolmogorov complexity) prevents the trivial assignment of computational structure to arbitrary physical systems—a limitation often raised as a challenge to computationalist views.
• Syntax vs. Semantics: The paper asserts and formalizes that meaning over a domain reduces to the correct selection and ordering of functions producing outputs. Syntax (rules, programs) is sufficient for semantics if and only if it enables generalization over the domain, precisely when $S$1. This locates the semantic content in the system’s functional organization, not in observer-relative interpretation or embodied context.
• Grounding and Correctness: The separation of generative capacity ($S$2) from correctness ensures that intelligence is evaluated independently of external evaluators or reward structures. Correctness is a property of system-domain correspondence, not a requisite component of intelligence density.

Comparative Analysis with Prior Frameworks

The proposed metric delineates itself from prior approaches:

• Legg-Hutter Universal Intelligence [legg2007]: Choi's metric obviates the need for explicit reward functions or environment enumerations. While Legg-Hutter sums over all computable environments using reward-weighted performance, $S$3 is intrinsic, computable for many simple physical systems, and ties generalization capacity to the physical structure of the system itself. The paper sketches a correspondence between the two in evolutionary equilibria but preserves distinctions in methodological prerequisites and operability.
• Kolmogorov Complexity: Intelligence density is dual to Kolmogorov complexity. While $S$4 measures the shortest program to produce $S$5, $S$6 measures the generative power (the breadth of independent outputs) for a fixed system. The duality characterizes the relationship between compression and generalization.
• Chollet's Skill-Acquisition Efficiency [chollet2019]: The focus of $S$7 is the final state of a system, not its learning path. Lookup tables, regardless of training data, remain at low $S$8 due to their lack of generalization, thus sidestepping issues raised regarding the acquisition of intelligence.
• Integrated Information Theory (IIT): Unlike $S$9, which attempts to quantify consciousness, $\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$0 targets operational intelligence and is robust to concerns regarding the spurious assignment of high values to structurally trivial systems.

Application and Implications for Real Systems

Through explicit analyses (e.g., multiplication algorithms, neural networks, LLMs, biological brains, logic gates), the paper demonstrates the scaling properties of $\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$1:

• Static circuits or lookup tables have constant or vanishing $\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$2 with domain expansion, showing absence of generalization.
• True generalization is only realized when a fixed, finite mechanism supports an unbounded, independent output space.
• Large pre-trained models (LLMs with finite parameters, human brains) qualify as knowing their domain insofar as their outputs diverge across expanding domains, constrained by their parameterization, supporting the claim that these architectures "know" in the formal sense.

The framework exposes a rigorous partition: no computation ($\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$3), memorization ($\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$4), computation without knowing ($\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$5), and knowing ($\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$6). This taxonomy allows unambiguous and substrate-agnostic assessment—yielding both theoretical clarity and practical criteria for system evaluation.

Empirical and Theoretical Outlook

While the computation of exact $\mathcal{I}(S) = \frac{\log_2 N(S)}{C(S)}$7 for complex systems (e.g., LLMs, the human brain) remains open due to Kolmogorov complexity’s incomputability, the asymptotic criterion is readily applicable. Compression-based approximations enable practical estimation, and future work may refine effective independence quantification and extend the framework to learning and meta-generalization.

A key vector for future theoretical development is the formalization of understanding as distinct from knowing—potentially operationalized in terms of a system’s capacity to construct new generalizing arrangements (meta-learning, program synthesis) rather than merely to execute fixed ones.

Conclusion

Choi’s operational definition of intelligence via intelligence density provides a rigorous, information-theoretic foundation for analyzing and comparing the generalization capacity of physical systems. It clarifies longstanding philosophical debates regarding computationalism, generalization, and meaning, and supplies a robust framework for the scientific measurement and comparison of artificial and biological intelligence. The metric's core innovation is its focus on the divergence of generative capacity per description length, furnishing a precise criterion that is neither anthropocentric nor contingent on environment, evaluator, or purpose. This formulation sets the stage for principled study of both natural and machine intelligence, and foregrounds the centrality of generalization in the mechanistic conception of intelligence.