This document discusses the challenges of evaluating and testing progress toward human-level artificial general intelligence (AGI). It proposes two approaches for testing early-stage AGI systems: 1) having AGIs complete online educational tests and 2) developing a series of "AGI Preschool Tests" (AIP Tests) inspired by theories of multiple intelligences and interactive learning. The document cautions that general intelligence is complex and multifaceted, so AGI testing should involve complex, multifaceted evaluations rather than simple tests. It argues against focusing on testing individual AGI system components in isolation, since components may have different roles and meanings depending on the overall AGI system.

Intelligence Assessment For Early-Stage Software Systems Aimed at Human-Level, Roughly Human-Like AGI PDF

The list of books you might like