Proof of AGI is the impossibility of evals (thewatershed.markpesce.com)

In a thought-provoking analysis, Mark Pesce argues that the increasing difficulty of evaluating artificial intelligence (AI) serves as a compelling indicator of the existence of artificial general intelligence (AGI). He draws parallels between AI assessments and human intelligence testing, stating that both face challenges stemming from the inability to decompose the constructs into measurable components, benchmark saturation, and the continual shifting of goalposts. These issues reflect a core truth: both AI and human intelligence exhibit a generality that standard tests cannot adequately capture. Pesce suggests that the very inadequacy of evaluation instruments acts as evidence for AGI's emergence. As AI systems begin to outperform traditional benchmarks and evaluations become as complex as testing human intelligence, it signals that AGI may have arrived without requiring formal proof. The implication is significant for the AI/ML community, as it challenges the efficacy of existing evaluation methods and prompts a reevaluation of what it means to measure intelligence, both artificial and human. The recognition that testing AGI poses similar difficulties to those faced in measuring human intelligence marks a pivotal moment in the discourse around AI capabilities.