The Coherence Wall and the Cold War of Machine Learning: Capability Competition and the Aftermath of Unbounded AI

This paper argues that frontier AI competition increasingly reproduces a Cold War structure: rival actors justify accelerated capability development by reference to the danger of adversary access. However, the analogy fails at the point where it matters most. Unlike nuclear deterrence, advanced AI does not present a stable symmetry of threat, verification, or second-strike logic. Its risks include autonomous deployment, persuasion, cyber leverage, surveillance, institutional capture, recursive improvement, and large-scale bypass of human agency. The central claim is that a capability race cannot solve alignment because alignment is non-indexical: its constraints must hold regardless of which actor deploys the system. If competing actors lack preservation, non-bypassability, anti-domination, and repair/closure constraints, then each side’s accusation of the other intensifies the race while leaving the underlying architecture unchanged. The paper concludes that the decisive distinction is not American versus Chinese AI, open versus closed AI, or commercial versus state AI, but kernel-bound systems versus systems whose constraints remain indexical to the deploying actor. It situates this argument within the Coherence Wall canon by defining the threshold at which capability outruns closure mechanisms and predicts strategic enclosure, access restriction, deterrence, or post-event governance when non-indexical closure is absent.