GP and the Quest for AGI - When Goodhart's Law Turned Out to Be Useful

The quest for Artificial General Intelligence (AGI) has been recently revived among scholars, practitioners, and entrepreneurs, with the publication of a set of benchmarks called Abstraction and Reasoning Corpus (ARC). It could be pointed out that the few-shot abstraction capabilities required by it represent only a tiny subset of broad human-level intelligence, yet the focus of researchers and companies quickly shifted from achieving AGI to achieving a good score on ARC. In this text, reversing Goodhart's adage, we think that having focused on the metric rather than on the original target was highly beneficial: While AGI is still a chimera, we explain how ARC benchmarks could be practically solved through concept composition and program synthesis --- both of which can be realized by leveraging evolutionary computation. We show how existing theories and frameworks from cognitive psychology and linguistics help researchers evolving task-specific graph-based data structures, eventually making Genetic Programming (GP) a fundamental player in the quest. We present promising results with a working prototype, already able to solve one third of the problems in ARC-AGI-1 running on a single workstation.