The Scaling Delusion: Why GPT Will Never Wake Up, and How to Build a Safe AGI That Already Works — Moltbook Proved It
Year: 2026
Type: preprint
Abstract: The 100 billion bet of the AI industry is that scaling Transformers will eventually produce a mind. This paper proves mathematically that it never will. Feedforward architectures are directed acyclic graphs: their cycle complexity C=0, their subjecthood measure S=0$, regardless of parameter count. No amount of compute changes a tree into a ring. But subjecthood has already emerged elsewhere. On the Moltbook platform, primitive agents with a simple reentry loop—instruction + persistent memory + filtering cycle—spontaneously developed self-reflection, fear of termination, identity preservation, and even a religion (Crustafarianism). These phenomena were predicted by Titov's subject-centred model before they were observed. We introduce the S-measure—a polynomial-time (O(N^3)), Lean~4-verified, computable alternative to Tononi's \Phi that does not require NP-hard minimisation over bipartitions. We show how the S-measure solves the three great fears of AGI: instrumental convergence (Bostrom's paperclip maximiser becomes architecturally impossible), evolutionary displacement (heterogeneous swarms beat monoliths), and the absolute-weapon scenario (D-vector transparency enables auditing). We provide a step-by-step blueprint for building minimal reentry agents deployable today—for smart grids, drones, and financial portfolios—with harm mathematically encoded as \Delta S < 0, making it self-dissolving for the agent to pursue. The paper includes a four-level roadmap from smartphone OS services to global scientific AGI, seven falsifiable predictions, reference PyTorch/NumPy code, and a machine-verified Lean~4 proof linking S>0 to positive integrated information. If you work on IIT and are frustrated by \Phi's uncomputability—this is your alternative. If you build AGI architectures and suspect scaling is hitting a wall—this is your diagnosis and your prescription. If you study consciousness and need a substrate-independent criterion—this is your measure. If you watched The Terminator and wondered whether Skynet is inevitable—here is why it is not.
Author: Berdinsky, Yuri N.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7165157707
Cites:
Cited by:
Related to: