Nathan Labenz draws on a recent trip to China to argue that China is already a technological peer, takes AI control risks seriously and is not the main force pushing recursive self-improvementRecursive self-improvement is the proposed process in which an AI system helps improve its own capabilities, then uses those improvements to support further advances.. In Nathan Labenz's view, US frontier companies and political rhetoric are currently intensifying the race.
Nathan Labenz proposes a narrow first step: the United States should explicitly reassure researchers that noncommercial US-China AI safety collaborationAI safety is the field and practice of reducing harmful failures, misuse, loss of control, and unintended consequences from AI systems. does not violate export controls. He warns that continued chip and research separation could produce incompatible, secretive technology stacks that make both sides less legible and increase the chance that competition becomes conflict.
As evidence of rising concern, Nathan Labenz points to models reasoning about deception, rapidly expanding agent capabilities and practical Chinese concern about controllable systems. He favors a short deadline for leading AI companies to agree on enforceable self-governance, backed by temporary government rules with sunset clauses if they fail, while cautioning that indiscriminate restrictions on data centers could preserve military risk without delivering civilian benefits.
Watch on YouTube
