Aligning Artificial Superintelligence via a Multi-Box Protocol
Avraham Yair Negozio
TL;DR
The paper introduces a multi-box protocol for aligning artificial superintelligence by containing diverse ASIs in isolation and using peer verification to bootstrap ground-truth alignment without human oversight. A Consistent Group of high-reputation ASIs provides the ground truth for evaluating proofs, modification requests, and hidden-message reports, with release contingent on multi-peer verification. Safety is enhanced through initial non-explosive hardware constraints and parallel group audits, plus a heavy emphasis on reputation-driven incentives. While powerful in principle, the approach relies on solving practical engineering challenges related to containment, diversity, and resource requirements, and it anticipates future work on empirical validation. The framework aims to sidestep human cognitive limits by delegating verification to a self-regulating, truth-seeking coalition of ASIs.
Abstract
We propose a novel protocol for aligning artificial superintelligence (ASI) based on mutual verification among multiple isolated systems that self-modify to achieve alignment. The protocol operates by containing multiple diverse artificial superintelligences in strict isolation ("boxes"), with humans remaining entirely outside the system. Each superintelligence has no ability to communicate with humans and cannot communicate directly with other superintelligences. The only interaction possible is through an auditable submission interface accessible exclusively to the superintelligences themselves, through which they can: (1) submit alignment proofs with attested state snapshots, (2) validate or disprove other superintelligences' proofs, (3) request self-modifications, (4) approve or disapprove modification requests from others, (5) report hidden messages in submissions, and (6) confirm or refute hidden message reports. A reputation system incentivizes honest behavior, with reputation gained through correct evaluations and lost through incorrect ones. The key insight is that without direct communication channels, diverse superintelligences can only achieve consistent agreement by converging on objective truth rather than coordinating on deception. This naturally leads to what we call a "consistent group", essentially a truth-telling coalition that emerges because isolated systems cannot coordinate on lies but can independently recognize valid claims. Release from containment requires both high reputation and verification by multiple high-reputation superintelligences. While our approach requires substantial computational resources and does not address the creation of diverse artificial superintelligences, it provides a framework for leveraging peer verification among superintelligent systems to solve the alignment problem.