General intelligence disentangled via a generality metric for natural and artificial intelligence

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Comparative Psychology

Background:

• General intelligence is traditionally viewed as success in diverse situations.
• Resource limitations necessitate understanding intelligence as performance up to a certain difficulty level.

Purpose of the Study:

• To dissect general intelligence into quantifiable measures of generality and capability.
• To apply these measures to individuals and groups across species (humans, animals, AI).
• To investigate the relationship between generality and capability in different agents.

Main Methods:

• Development of two non-populational metrics: generality and capability.
• Application of these metrics to cognitive and perceptual tests for individuals and groups.
• Analysis of how capability is distributed as a function of task difficulty.

Main Results:

• Generality and capability can decouple at the individual level; specialized agents can exhibit high capability and vice versa.
• These metrics also decouple at the population level, with high-capability groups often showing no diminishing returns in generality.
• Individual generality measure correlates with traditional notions of general intelligence and cognitive efficiency.

Conclusions:

• A new quantitative framework for general intelligence based on task difficulty is proposed.
• This framework allows for a nuanced evaluation of general intelligence in humans, animals, and AI systems.
• The findings provide tools for understanding the evolution of general intelligence and assessing progress in Artificial General Intelligence (AGI).