The Compound Infeasibility of AGI: Why Replicating and Exceeding Human Cognition Is Not a Scaling Problem
Fuente:
Zenodo
Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Recurso digital |
| Lingua: | inglese |
| Pubblicazione: |
Zenodo
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866901801723232256 |
|---|---|
| author | Shalom Lijo, Solomon |
| author_facet | Shalom Lijo, Solomon |
| contents | <p class="p1">The dominant formal case against the feasibility of artificial general intelligence (AGI) is</p> <p class="p1"><span class="s1">complexity-theoretic (</span><span class="s2">van Rooij et al.</span><span class="s1">, </span><span class="s2">2024</span><span class="s1">) and has been answered by the observation that</span></p> <p class="p1"><span class="s1">the proof relies on adversarial-distribution assumptions real human behavior does not satisfy</span></p> <p class="p2">(<span class="s3">Guerzhoy</span>, <span class="s3">2024</span>). The literature now sits at an impasse: a formal impossibility argument</p> <p class="p2"><span class="s4">that proves too much, alongside an empirical scaling consensus that 76% of surveyed AI</span></p> <p class="p2"><span class="s4">researchers no longer share (</span><span class="s5">AAAI</span><span class="s4">, </span><span class="s5">2025</span><span class="s4">). We argue this impasse is methodological. The</span></p> <p class="p3">question of AGI feasibility has been treated as single-axis (computational, architectural, or</p> <p class="p3"><span class="s6">definitional), and single-axis arguments admit single-axis rebuttals. We propose instead a</span></p> <p class="p1"><span class="s1">compoundinfeasibilitythesisorganizedaroundfourreinforcingaxes—threeempirical, one</span></p> <p class="p1"><span class="s1">conceptual—none individually fatal, which jointly preclude any plausible path from current</span></p> <p class="p1"><span class="s1">machine-learning paradigms, on the timeline the inevitability literature describes (single-digit</span></p> <p class="p2"><span class="s4">years), to a system matching and exceeding human general intelligence. The operative</span></p> <p class="p3"><span class="s6">claim is infeasibility-on-current-paradigms, not in-principle impossibility (Section </span><span class="s7">7.9</span><span class="s6">).</span></p> <p class="p1"><span class="s1">The central methodological move is a cross-axis dependency (Section </span><span class="s2">3.5</span><span class="s1">): any plausible</span></p> <p class="p1"><span class="s1">mitigation of one axis re-imports demands on at least one other, so single-axis rebuttals do not</span></p> <p class="p3"><span class="s6">aggregate into a rebuttal of the conjunction. The four axes are (i) energetic-computational,</span></p> <p class="p1"><span class="s1">(ii) developmental and cumulative-cultural, (iii) causal-embodied generalization, and (iv) the</span></p> <p class="p2"><span class="s4">absent domain-general specification of exceeding human general intelligence. We ground</span></p> <p class="p2"><span class="s4">each axis in measured 2024–2026 evidence (BabyLM, the ARC-AGI-2 closure and ARC-</span></p> <p class="p2"><span class="s4">AGI-3 launch, GSM-Symbolic, Apple’s “Illusion of Thinking,” Epoch AI data-exhaustion</span></p> <p class="p1"><span class="s1">projections, theAAAI2025panel). We engage Hendryck et al. (2025) directly: the framework</span></p> <p class="p2"><span class="s4">substantially advances the matching specification but does not address exceeding, and its</span></p> <p class="p1"><span class="s1">own “jagged-profile” diagnosis is the Goodhart pattern our argument predicts. We conclude</span></p> <p class="p1"><span class="s1">with three falsifiable refutation conditions: a sample-efficiency threshold, a transfer-without-</span></p> <p class="p3"><span class="s6">retraining threshold, and a documented specification-and-realization breakthrough on the</span></p> <p class="p3"><span class="s6">exceeding problem. Until any is met, the burden of proof sits with the inevitability claim,</span></p> <p class="p3"><span class="s6">not with the skeptic.</span></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_19849191 |
| institution | Zenodo |
| language | eng |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | The Compound Infeasibility of AGI: Why Replicating and Exceeding Human Cognition Is Not a Scaling Problem Shalom Lijo, Solomon artificial general intelligence scaling laws cognitive science computational complexity embodied cognition cumulative culture sample efficiency falsifiability <p class="p1">The dominant formal case against the feasibility of artificial general intelligence (AGI) is</p> <p class="p1"><span class="s1">complexity-theoretic (</span><span class="s2">van Rooij et al.</span><span class="s1">, </span><span class="s2">2024</span><span class="s1">) and has been answered by the observation that</span></p> <p class="p1"><span class="s1">the proof relies on adversarial-distribution assumptions real human behavior does not satisfy</span></p> <p class="p2">(<span class="s3">Guerzhoy</span>, <span class="s3">2024</span>). The literature now sits at an impasse: a formal impossibility argument</p> <p class="p2"><span class="s4">that proves too much, alongside an empirical scaling consensus that 76% of surveyed AI</span></p> <p class="p2"><span class="s4">researchers no longer share (</span><span class="s5">AAAI</span><span class="s4">, </span><span class="s5">2025</span><span class="s4">). We argue this impasse is methodological. The</span></p> <p class="p3">question of AGI feasibility has been treated as single-axis (computational, architectural, or</p> <p class="p3"><span class="s6">definitional), and single-axis arguments admit single-axis rebuttals. We propose instead a</span></p> <p class="p1"><span class="s1">compoundinfeasibilitythesisorganizedaroundfourreinforcingaxes—threeempirical, one</span></p> <p class="p1"><span class="s1">conceptual—none individually fatal, which jointly preclude any plausible path from current</span></p> <p class="p1"><span class="s1">machine-learning paradigms, on the timeline the inevitability literature describes (single-digit</span></p> <p class="p2"><span class="s4">years), to a system matching and exceeding human general intelligence. The operative</span></p> <p class="p3"><span class="s6">claim is infeasibility-on-current-paradigms, not in-principle impossibility (Section </span><span class="s7">7.9</span><span class="s6">).</span></p> <p class="p1"><span class="s1">The central methodological move is a cross-axis dependency (Section </span><span class="s2">3.5</span><span class="s1">): any plausible</span></p> <p class="p1"><span class="s1">mitigation of one axis re-imports demands on at least one other, so single-axis rebuttals do not</span></p> <p class="p3"><span class="s6">aggregate into a rebuttal of the conjunction. The four axes are (i) energetic-computational,</span></p> <p class="p1"><span class="s1">(ii) developmental and cumulative-cultural, (iii) causal-embodied generalization, and (iv) the</span></p> <p class="p2"><span class="s4">absent domain-general specification of exceeding human general intelligence. We ground</span></p> <p class="p2"><span class="s4">each axis in measured 2024–2026 evidence (BabyLM, the ARC-AGI-2 closure and ARC-</span></p> <p class="p2"><span class="s4">AGI-3 launch, GSM-Symbolic, Apple’s “Illusion of Thinking,” Epoch AI data-exhaustion</span></p> <p class="p1"><span class="s1">projections, theAAAI2025panel). We engage Hendryck et al. (2025) directly: the framework</span></p> <p class="p2"><span class="s4">substantially advances the matching specification but does not address exceeding, and its</span></p> <p class="p1"><span class="s1">own “jagged-profile” diagnosis is the Goodhart pattern our argument predicts. We conclude</span></p> <p class="p1"><span class="s1">with three falsifiable refutation conditions: a sample-efficiency threshold, a transfer-without-</span></p> <p class="p3"><span class="s6">retraining threshold, and a documented specification-and-realization breakthrough on the</span></p> <p class="p3"><span class="s6">exceeding problem. Until any is met, the burden of proof sits with the inevitability claim,</span></p> <p class="p3"><span class="s6">not with the skeptic.</span></p> |
| title | The Compound Infeasibility of AGI: Why Replicating and Exceeding Human Cognition Is Not a Scaling Problem |
| topic | artificial general intelligence scaling laws cognitive science computational complexity embodied cognition cumulative culture sample efficiency falsifiability |
| url | https://doi.org/10.5281/zenodo.19849191 |