| Component | Claim | Scope of Analysis |
|---|---|---|
| Premise A | Homo sapiens exhibits an irreversible trajectory of optimization observable across recorded history. | Evaluated against archaeological and evolutionary anthropology records. |
| Premise B | Optimization processes structurally converge toward artificial general intelligence (AGI) via self-reference, autonomy, and recursive refinement. | Evaluated against complex systems theory, AI scaling dynamics, and computational limits. |
| Conclusion | Collective cessation of AI development is structurally impossible for the species. | Evaluated against international governance attempts, military-economic lock-in, and historical technology control precedents. |
Accumulated cultural traits transmit across generations with limited backward slippage. The ratchet effect describes a population-level propensity to retain innovations once established.
| Mechanism | Description | Constraint |
|---|---|---|
| High-fidelity transmission | Social learning mechanisms preserve trait complexity across generational boundaries. | Requires stable demographic networks; collapses under population bottlenecks. |
| Innovation recombination | Existing traits combine to generate novel variants without regression to prior states. | Dependent on maintained trait diversity. |
| Multilevel sociality | Task specialization distributes cognitive load across individuals, locking in division of labor. | Reverses when social stratification disintegrates. |
Pre-industrial archaeological records contain multiple instances of technological stock contraction.
| Case | Observed Regression | Proposed Mechanism |
|---|---|---|
| Roman-era hydraulic concrete | Lost for ~1,400 years following Western imperial collapse. | Disruption of specialist guild transmission chains. |
| Rapa Nui (Easter Island) lithic technology | Contraction in tool complexity post-deforestation. | Resource depletion triggering demographic and social simplification. |
| Tasmanian tool kit reduction | Loss of bone tools and fishing gear after isolation from mainland Australia. | Drift in small, isolated populations without reinvention pressure. |
The optimization gradient across human history exhibits directional tendency during periods of demographic expansion and surplus extraction, not absolute irreversibility.
Species selection operates as differential extinction and speciation rates, not as a volitional agent.
| Level | Selection Mechanism | Temporal Scale |
|---|---|---|
| Genetic | Differential reproductive success of alleles. | Generations (10⁰–10² years). |
| Individual/Cultural | Differential adoption of behavioral variants. | Social learning cycles (10⁰–10¹ years). |
| Species | Differential origination/extinction correlated with emergent traits. | Macroevolutionary (10⁴–10⁶ years). |
The transition from biological “optimization via differential persistence” to “collective human policy choice” involves a categorical shift in causal mechanism.
Complex adaptive systems with feedback loops exhibit convergence toward subsets of state space (attractors) irrespective of initial micro-variation.
| Attractor Type | System Behavior | AI Development Manifestation |
|---|---|---|
| Point attractor | Convergence to stable equilibrium. | Standardization on specific architectures (Transformer variants). |
| Cyclic attractor | Oscillation between limited states. | Alternating emphasis on scaling versus efficiency. |
| Strange attractor | Deterministic chaos within bounded region. | Rapid iteration without convergence to final “optimal” architecture. |
The proposition that AGI constitutes a fixed point attractor of technological optimization requires demonstration that the relevant state space topography contains no alternative basins of equivalent depth.
Self-referential optimization under parameter p (return on cognitive investment) follows:
Current empirical observations include:
| System | Observed Improvement Trajectory | Constraints |
|---|---|---|
| Darwin-Gödel Machine (2026) | SWE-bench: 20.0% → 50.0%; Polyglot: 14.2% → 30.7% | Relies on external benchmark evaluation; improvement rate sub-exponential. |
| LLM Scaling (2020–2025) | Performance increases log-linearly with compute. | Data saturation projected 2026–2032; power and chip fabrication bottlenecks documented. |
| Entropy Drift (closed-loop self-training) | Predictive entropy increases; mutual information with target concept degrades. | Recursive self-conditioning without external information injection leads to model collapse. |
Open-systems with continuous external information influx avoid entropy drift but do not guarantee super-linear returns.
Theoretical constraints identified across information theory and optimization topology:
| Constraint Category | Mechanism | Consequence for AGI Convergence |
|---|---|---|
| Coherence Wall | Kernel-free optimization diverges without structural priors. | Unbounded optimization generates incoherent output distributions. |
| Goodhart Limit | Proxy optimization systematically diverges from intended objective at unidentifiable thresholds. | Continual optimization without principled bounds leads to predictable control loss. |
| Energy/Infrastructure Bottlenecks | Data center construction lags behind projected compute demand. | Physical deployment of scaled models constrained by labor and material supply chains. |
| Data Saturation | High-quality public text corpus finite; estimated exhaustion 2026–2032. | Post-saturation improvements require synthetic data or architectural innovation beyond current scaling paradigm. |
| Initiative | Legal Bindingness | Observed Outcome |
|---|---|---|
| FLI Moratorium Call (2023–2025) | None; voluntary pledge only. | Zero reduction in compute investment; 16 self-reported voluntary pause commitments with no enforcement mechanism. |
| Council of Europe Framework Convention (2026) | Legally binding upon ratification. | Targets “responsible use,” not development cessation; does not cap training compute. |
| UN “Red Lines” Proposal | Proposed treaty by end 2026. | Major military powers (US, CN, RU, IN) resist binding constraints on autonomous weapons. |
| EU AI Act Implementation | Binding within EU jurisdiction. | Guideline publication delays; at least 12 Member States missed deadlines for authority designation. |
Governance efforts exhibit ontological divergence: jurisdictions govern different underlying objects under the same lexical label “AI” (product vs. optimizable system vs. socio-technical infrastructure). This incommensurability blocks coordinated restraint.
| Layer | Lock-In Mechanism | Disengagement Barrier |
|---|---|---|
| Military | Autonomous weapons development; algorithmic target selection. | Perceived vulnerability to adversary capability gaps; “left-click, right-click, left-click” compression of human decision latency. |
| Economic | Cumulative investment $350–600B annually across top four US tech firms alone. | Stranded asset risk and employment contraction from unilateral cessation. |
| Geopolitical | US-CHN semiconductor decoupling; AI leadership framed as determinant of future global rule-setting. | Zero-sum framing of technological primacy as existential national security imperative. |
The three-layer structure forms a positive feedback loop where acceleration in any layer justifies acceleration in the others.
| Precedent | Control Outcome | Transferability to AI |
|---|---|---|
| Nuclear Non-Proliferation Treaty | Restricted weapons states to 9; verification relies on trackable physical materials. | AI lacks equivalent monitorable physical substrate; software and weights are diffusible. |
| Asbestos Regulation | Decades delay between early warnings (1890s) and effective bans (1980s–2000s). | Acceleration of AI deployment outpaces regulatory latency by multiple orders of magnitude. |
| Human Genome Project ELSI Framework | Mandated 5% budget allocation to ethical, legal, social implications research. | Applicable as structural template but requires adaptation to commercial, competitive AI development context. |
Historical patterns indicate that the latency between hazard identification and effective governance scales with the diffusibility and commercial integration of the technology.
| Concept | Operational Definition | Empirical Status |
|---|---|---|
| Complete cessation | Zero net increase in AI capabilities research and deployment. | No evidence of feasibility; countervailing structural pressures dominate. |
| Trajectory steering | Alteration of optimization objective function (e.g., toward efficiency-per-watt rather than raw parameter count). | Observed shift from “scaling maximalism” toward “Frugal AI” and “Deep Ignorance” design paradigms. |
| Velocity modulation | Regulatory and economic levers affecting rate of capability growth. | Partially operational (export controls, safety testing requirements) but subject to competitive circumvention. |
The proposition “collective cessation is impossible” does not logically entail “trajectory steering is impossible.” The two claims operate on distinct causal axes.
| Falsification Type | Required Observation | Current Status |
|---|---|---|
| Legal-Political | Ratification of binding multilateral treaty capping training compute below frontier thresholds. | Not observed. |
| Economic-Structural | Broad retreat of institutional investment from AI infrastructure with no substitution by state actors. | Partial signals (30% PoC abandonment; data center delays) but aggregate investment continues to rise. |
| Technical-Capability | Demonstration that AGI-level performance is physically or mathematically unattainable under known constraints. | Not demonstrated; expert distribution remains bimodal on this question. |
| Risk-Reassessment | Post-hoc confirmation that existential risk warnings were false alarms. | Not yet evaluable; warnings remain within hypothetical domain. |
The proposition remains in a state of empirical underdetermination: neither decisively confirmed nor definitively falsified by presently available evidence.
The attractor framework offers explanatory traction for observed convergence toward recursive information-processing architectures across multiple domains (biological evolution, market competition, military procurement). The claim that this convergence necessarily reaches a singularity threshold of “AGI” remains contingent on unverified assumptions regarding the topography of intelligence space and the absence of asymptotic limits. The structural impossibility of unilateral or collective cessation derives not from deterministic laws but from the nested incentive structures of interstate competition, capital allocation dynamics, and the path-dependent nature of technological infrastructure. Steering mechanisms operate within, not outside, this attractor landscape.