A Definition of AGI [pdf] (www.agidefinition.ai)

A coalition of researchers proposes a concrete, quantifiable definition of AGI by adapting the Cattell‑Horn‑Carroll (CHC) theory of human cognition. They break general intelligence into ten equally weighted cognitive domains—General Knowledge, Reading & Writing, Math, On‑the‑Spot Reasoning, Working Memory, Long‑Term Memory Storage, Long‑Term Memory Retrieval, Visual Processing, Auditory Processing, and Speed—and operationalize each with psychometrically grounded task batteries (text, visual, auditory). The framework produces a standardized AGIScore (0–100%) where 100% equals a well‑educated adult. Applying it yields a “jagged” profile for current models: GPT‑4 scores 27% and GPT‑5 58%, with notable strengths in knowledge‑intensive areas but persistent, critical shortfalls in foundational capacities—especially long‑term memory storage (MS = 0% for both in the reported table), working memory, multimodal perception, and reliable retrieval (confabulation mitigation). This approach is significant because it replaces vague claims about “AGI” with a diagnostic, multimodal psychometric toolkit that can pinpoint specific architectural and training gaps. By mapping AI capabilities to validated human cognitive constructs, the framework guides research priorities (e.g., continual learning and memory systems, robust retrieval to reduce hallucinations, on‑the‑spot reasoning and working memory), informs safety and policy discussion with measurable progress, and provides a repeatable, human‑centric benchmark distinct from task‑specific or economic‑impact metrics.