New framework for defining and objectively measuring AGI, based on 87 skills and abilities, visualising progress over time

A new framework for objectively measuring progress toward Artificial General Intelligence (AGI) repurposes a 30-year-old taxonomy of human capabilities originally built to describe jobs. Researcher Anton Korinek discovered that the O*NET database—maintained by the US Department of Labor and containing 87 validated skills and abilities like Deductive Reasoning, Programming, and Manual Dexterity—serves as a perfect AGI scorecard. By benchmarking frontier AI models (Gemini 3.1 Pro, Claude Opus 4.6) and robotics systems against specific tests at three time points (2020, 2023, 2025), the analysis translates AI performance into human percentiles.

The resulting visualizations are striking spider charts showing rapid capability expansion. The data reveals that as of 2025, AI exceeds the 25th percentile of human ability in 83 of 87 dimensions. The only remaining gaps are in physical domains like Stamina and Manual Dexterity, illustrating Moravec's paradox—that AI masters cognitive tasks before sensorimotor skills. This framework provides the first empirical, longitudinal tracking system for AGI progress, with a clear finish line: when AI saturates all 87 dimensions, it has reached human-level general capability.

The methodology anchors each assessment to established benchmarks: SWE-bench for programming, ARC-AGI for reasoning, Mobile ALOHA for manipulation, and KITTI for spatial orientation. Because the O*NET taxonomy was developed independently of AI for occupational psychology, it avoids bias toward current AI strengths. The open dataset allows researchers to challenge scores and track the "frontier filling in" across cognitive, sensory, and physical domains simultaneously, providing a concrete visualization of AI approaching human substrate.