TLDR: A new research paper proposes a definition of intelligence as the ability to generate samples consistent with a given category, measured by how indistinguishable generated samples are from original ones. This “ε-category intelligence” framework is designed to be operational, falsifiable, and generalizable across biological and artificial systems. It suggests that intelligence does not necessarily require consciousness or learning, and introduces concepts of synchronic (breadth at a time) and diachronic (adaptive capacity over time) capabilities for general intelligence. The definition has implications for AI evaluation, safety, and generalization to novel tasks.
The quest to build Artificial General Intelligence (AGI) brings forth a fundamental question: what exactly is intelligence? Despite its widespread use, a formal and universally applicable definition of intelligence has remained elusive. A new research paper, “On the Definition of Intelligence”, proposes a novel framework that aims to unify our understanding of intelligence across biological, artificial, and hybrid systems.
Traditional definitions of intelligence often fall short when applied to the diverse and evolving landscape of AI. For instance, defining intelligence as the ability to achieve goals in various environments, while useful for reinforcement learning, doesn’t fully capture the essence of generative AI systems that create images or text without explicit goals or environments. Similarly, the Turing Test, while historically significant, is most effective when machine intelligence is comparable to or below human levels, and doesn’t account for domains where machines vastly outperform humans. Other definitions, often anthropocentric, tie intelligence to human capabilities, which becomes problematic as AI surpasses human performance in specialized tasks.
A New Perspective: Generative Fidelity to a Category
The paper introduces a minimal and foundational definition: intelligence is the ability, given one or more samples from a specific category, to generate new samples that also belong to that same category. This concept is formalized as “ε-category intelligence.” In essence, a system is considered ε-intelligent with respect to a category if a chosen “distinguisher” (like a human judge or an automated critic) cannot tell the difference between the original samples and the generated samples beyond a small tolerance, ε.
Consider an example: if an AI system generates images in the style of Studio Ghibli, and human judges cannot reliably distinguish these generated images from genuine Ghibli images, then the AI is deemed ε-intelligent in the “Ghibli-style images” category. The key insight here is that the similarity within a category can be highly abstract. For instance, a horse and a car, despite their physical differences, could both belong to the abstract category of “conveyances for human transport.” An intelligent entity, in this context, could generate the concept of a car given the concept of a horse, as both fit the category.
This definition is designed to be:
- Operational: Intelligence can be measured through observable outputs.
- Falsifiable: Systems that consistently produce outputs outside the reference category are clearly identified as less intelligent.
- Generalizable: It applies uniformly to humans, machines, and natural systems, avoiding human-centric biases.
Intelligence Beyond Consciousness and Learning
One of the paper’s intriguing assertions is that intelligence, as defined, does not necessarily require consciousness or even an inherent capacity for learning. A computer program designed to solve calculus problems, for example, can exhibit high ε-category intelligence by consistently producing correct solutions, even if it operates deterministically without any form of consciousness. Similarly, a fixed, non-learning language model that consistently generates correct answers within a specific domain would still be considered intelligent by this definition. Learning, in this framework, is seen as a mechanism that *changes* a system’s intelligence over time, rather than a prerequisite for intelligence itself.
The definition also highlights the “locality” of intelligence. A homing pigeon, while not capable of human language, can classify breast-cancer slides with high accuracy, demonstrating intelligence specific to that visual classification category. General intelligence, then, would imply the ability to generate samples across a much wider range of categories.
A Dynamic View of General Intelligence
The paper introduces a dynamic framework for general intelligence, distinguishing between two capabilities:
- Synchronic Capability: An entity’s breadth and capacity to generate high-fidelity samples for a given set of categories at a specific point in time.
- Diachronic Capability: The ability of an entity to adapt and transition its competence from one set of categories to another over time. This essentially measures its adaptive capacity.
Concepts like efficiency, cost, and compression gain their importance by contributing to these two capabilities. For instance, a low-cost generation process can enhance synchronic capability when dealing with many categories, while compressed knowledge representation can boost diachronic capability when categories change frequently.
Also Read:
- Defining Neural Network World Models: A Framework for Clarity in AI Research
- The Adaptive Frontier: Exploring Self-Evolving AI Agents
Implications for AI Development
This new definition has significant implications for the evaluation, safety, and generalization of AI systems. By focusing on the indistinguishability of generated samples from target categories, it offers a unified yardstick for comparing diverse intelligent behaviors. For safety, it suggests that curating high-quality, safety-compliant synthetic data during initial training can mitigate harmful outputs at their source. Furthermore, by training on multiple known categories, AI systems can learn abstract category structures, enabling them to generalize and generate consistent samples for entirely new, unseen categories, thus fostering genuine generalized intelligence.
The paper concludes by suggesting that this framework could be further refined using Category Theory, a branch of mathematics, to formally model categories and their relationships, potentially providing a deeper structural account of generalizable intelligence.
