On The Definition of Intelligence (2024) AGI-25 (arxiv.org)

A recent paper proposes a species-agnostic, operational definition of intelligence called "entity fidelity": an agent is intelligent with respect to a concept if, given exemplar entities that instantiate that concept, it can produce new entities that are indistinguishable from the exemplars to any allowed evaluator. The authors formalize this as ε-concept intelligence: an agent is ε-intelligent for a concept if no admissible distinguisher can separate generated entities from original ones by more than ε. The framework treats concepts as sets or distributions of entities (images, behaviors, plans, etc.), defines admissible distinguishers and tolerances, and sketches empirical protocols for measuring fidelity across modalities and tasks. This formulation matters because it unifies many AI paradigms—reinforcement learning, generative modeling, classification, analogical reasoning, and goal-directed decision-making—under a single, testable criterion. Practically, it gives a concrete path to benchmark concept learning and generalization (measure transfer to new exemplars), to design safety evaluations (by specifying powerful distinguishers to reveal failures or deceptive behavior), and to compare architectures through measurable ε thresholds. Key technical implications include careful choice of admissible distinguishers and ε to balance realism and rigor, and using distributional tests to operationalize AGI evaluation across modalities.