Work

Cost Function as Worldview: A Threshold-Based Framework for AGI Endogenous Constraint

Year: 2026
Type: preprint
Abstract: Current approaches to AI alignment—RLHF, Constitutional AI, preference aggregation—suffer from a fundamental flaw: they impose external normative constraints on systems lacking internal cost-awareness. This paper proposes an alternative framework: unify human value considerations under a single structural primitive—the cost-bearing threshold, formalized as a computable cost function evaluated across three scales: personal (individual resources and consequences), social (public order, law, collective cost), and systemic (physical causality, irreversibility, structural collapse). We advance four claims: (1) value differentiation is a matter of scale, not content; (2) total cost = realized sunk costs + projected costs over multi-step causal chains; (3) value change is parameter update within a dynamic cost function, not epistemic enlightenment; (4) legal codes constitute institutionally stabilized, intersubjective cost mappings accumulated over millennia. Building on this, we introduce the Forced Inference Tree (FIT) architecture: given a candidate action, the system samples successor branches (≤5 per node), rolls forward (≤10 steps), queries a structured cost matrix at each node, and prunes branches exceeding threshold. Instead of teaching AI to "be good", we force it to calculate the price. If cumulative cost exceeds the bearing threshold, the action is blocked. This is not an alignment algorithm; it is a survival constraint architecture for embodied AGI—implementable today, falsifiable tomorrow.
Author: Zhang, You
Language: English
Open Access status: green
OpenAlex ID: W7225873626
Cites:
Cited by:
Related to: 10