What question did this study set out to answer?

This research questions the assumption that achieving AGI is primarily a scaling problem, asserting instead that multiple interrelated factors contribute to its infeasibility.

April 30, 2026Open Access

The Compound Infeasibility of AGI: Why Replicating and Exceeding Human Cognition Is Not a Scaling Problem

Key Points

This research questions the assumption that achieving AGI is primarily a scaling problem, asserting instead that multiple interrelated factors contribute to its infeasibility.
Propose a compound infeasibility thesis organized around four axes: energetic-computational, developmental-cultural, causal-embodied, and domain-general.
Ground the axes in empirical evidence from 2024-2026 data.
Engage with existing literature to highlight limitations in prevailing arguments regarding AGI potential.
Identified that 76% of AI researchers disagree with the dominant scaling consensus.
Demonstrated that addressing one axis of the thesis introduces demands on others, complicating the path to AGI.
Outlined three conditions necessary for disproving the infeasibility claim regarding AGI exceeding human intelligence.

Abstract

The dominant formal case against the feasibility of artificial general intelligence (AGI) is complexity-theoretic (van Rooij et al., 2024) and has been answered by the observation that the proof relies on adversarial-distribution assumptions real human behavior does not satisfy (Guerzhoy, 2024). The literature now sits at an impasse: a formal impossibility argument that proves too much, alongside an empirical scaling consensus that 76% of surveyed AI researchers no longer share (AAAI, 2025). We argue this impasse is methodological. The question of AGI feasibility has been treated as single-axis (computational, architectural, or definitional), and single-axis arguments admit single-axis rebuttals. We propose instead a compoundinfeasibilitythesisorganizedaroundfourreinforcingaxes—threeempirical, one conceptual—none individually fatal, which jointly preclude any plausible path from current machine-learning paradigms, on the timeline the inevitability literature describes (single-digit years), to a system matching and exceeding human general intelligence. The operative claim is infeasibility-on-current-paradigms, not in-principle impossibility (Section 7.9). The central methodological move is a cross-axis dependency (Section 3.5): any plausible mitigation of one axis re-imports demands on at least one other, so single-axis rebuttals do not aggregate into a rebuttal of the conjunction. The four axes are (i) energetic-computational, (ii) developmental and cumulative-cultural, (iii) causal-embodied generalization, and (iv) the absent domain-general specification of exceeding human general intelligence. We ground each axis in measured 2024–2026 evidence (BabyLM, the ARC-AGI-2 closure and ARC- AGI-3 launch, GSM-Symbolic, Apple’s “Illusion of Thinking,” Epoch AI data-exhaustion projections, theAAAI2025panel). We engage Hendryck et al. (2025) directly: the framework substantially advances the matching specification but does not address exceeding, and its own “jagged-profile” diagnosis is the Goodhart pattern our argument predicts. We conclude with three falsifiable refutation conditions: a sample-efficiency threshold, a transfer-without- retraining threshold, and a documented specification-and-realization breakthrough on the exceeding problem. Until any is met, the burden of proof sits with the inevitability claim, not with the skeptic.

Read Full Paperexternally

AIに質問

Bookmark

View Full Paper