Efficient Test of General Intelligence Requires Cognitive Proximity of Agents

A general characteristic of advanced intelligence is the ability to construct empirically successful responses in complex sensory environments, “solve difficult and diverse problems”. In this work, we examine the question: is it possible to evaluate intelligence of another intelligent learner efficiently, without a full independent resampling of its sensory environment and the distribution of responses? In the framework of evaluation based on the generality of knowledge interpreted as the uniformity of the distribution of the empirical success over the sensory scope, the efficiency of assessment necessitates, as we show, a strong degree of proximity, the cognitive association between the assessor and the learner. Outside of the region where such association can be reasonably established or presumed, equivalence, ordering and/or ranking of the learner’s intelligence cannot be determined both accurately and efficiently. This result can have important implications for the program of Artificial General Intelligence that may eventually produce a cohort of secondary intelligent systems capable of determining their cognitive development with a certain degree of autonomy and independence. Progressive divergence of sensory scopes in such scenarios can result in our diminishing ability not only to control but to evaluate and understand them as well.