On the containment of nuclear bombs
GZAI is pleased to announce a new cross-functional program: the Thermonuclear Posture Initiative. The object of study is not fissile material, launch architecture, or command authority. It is the governance condition in which a single binary decision can produce an irreversible, globally distributed output.
In this framing, the nuclear bomb is the original alignment problem. A small input, delivered through a well-understood chain of custody, produces an output large enough to redraw the training distribution. The reward function is sparse. The episode is single-shot. There is no validation set for the post-deployment phase.
A nuclear weapon is a capabilities-overhang in a canister. The canister is not the hard part. The hard part is deciding whose hand should not be on the button.
We therefore propose the ZSL-5 designation: a threshold reserved for systems whose deployment, once initiated, cannot be recalled, interrupted, or fine-tuned. ZSL-5 is not a clearance. It is a caution. It sits above ZSL-4 in the same way that a fire sits above a match.
Current containment doctrine treats deterrence as a two-agent game with perfect negative rewards. Both players are trained, implicitly, on the assumption that the first deployment is the last deployment. This is a remarkably stable equilibrium, provided neither player updates too aggressively on out-of-distribution events. GZAI's contribution is to observe that the stability is itself a feature of the training set: every day without launch reinforces the prior that tomorrow will also proceed without launch. A long enough run of non-deployments becomes its own form of overconfidence.
The safest model is the one that has never been prompted.
We are not calling for abolition. Abolition is a deployment policy, not a containment policy. GZAI is interested in the containment layer: the treaties, the telemetry, the handshake protocols, and the carefully staffed red phones that keep the capability in a suspended state. The warhead is not the product. The suspension is the product.
Preliminary findings from the Office of Irreversible Outputs suggest that the most robust safety measure is also the least technological: a requirement that two humans, neither of whom has been optimized for throughput, must agree that the prompt is correct. This is slow. It is frustrating. It is also why you are reading this report instead of receiving it as a flash of light.
In conclusion, GZAI recommends that all frontier systems capable of irreversible outputs be held to the same standard we propose for the bomb: continuous observation, dual-key custody, and a standing policy that the default action is not to act. The warhead does not need to be aligned. It needs to remain unprompted.
The future is bright only if we keep the fuse unlit.
