Two of the loudest voices on AGI risk—Stuart Russell (Human Compatible, 2019) and Leopold Aschenbrenner (Situational Awareness, 2024)—agree on the diagnosis but differ on the cure. Russell says AI should defer to human preferences. Aschenbrenner says AGI arrives by 2027 and state control is inevitable. BeTrueCore, a proposal just posted on ethresear.ch, argues both share a structural gap: they assume preferences can be observed, aggregated, and managed top-down—by AI systems, corporations, or states.
The Shared Blind Spot in Russell and Aschenbrenner
Russell's assistance game asks which preferences to defer to, but runs into the preference falsification problem (Kuran, 1995): people don't express true preferences under social pressure. Aschenbrenner's "AGI realism" hands control to the state, but that's control over AGI, not sovereignty for citizens. Both implicitly treat values as something that can be extracted and centrally managed. BeTrueCore sees this as a category error: you can't learn genuine human values from AI alone—you need infrastructure where humans express those values freely, verifiably, and continuously.
BeTrueCore's Three Structural Differences
The proposal lists three concrete departures. First, energy posture: AGI infrastructure demands gigawatts of centralised compute; BeTrueCore processes biometrics locally (FaceID on device), generates ZK-proofs client-side, and archives proofs cheaply via Celestia DA. Sovereignty without the energy race. Second, preference authenticity: MACI + ZK-SNARKs eliminate the preference falsification problem by letting individuals vote with mathematically weighted intuition (Vote Weight Unit / VWU) while maintaining anonymity and verifiability. Third, direction of control: Aschenbrenner's model is top-down state control; BeTrueCore is bottom-up sovereign collective intelligence. AI becomes notary, not judge.
Why This Matters for AGI Alignment Infrastructure
Russell's core question—how do AI systems learn genuine human values?—cannot be answered by AI alone. BeTrueCore doesn't claim to have built the answer; it's a proposal for the missing infrastructure layer. The full specification lives on Zenodo, and the authors are inviting discussion on GitHub—especially around the canonical observation schema for Vote Weight Unit audit trails.
If this line of thinking holds, the AGI alignment debate shifts from “how do we control the AI” to “how do we build the feedback channel that makes control meaningful.” That's a question the industry has mostly ignored. BeTrueCore is forcing it into the open.
Source: Beyond AGI Control: Sovereign Collective Intelligence as Complementary Infrastructure
Domain: ethresear.ch
Comments load interactively on the live page.