On: US and China Discuss Alerting Each Other to AI National Security Threats

Two governments are discussing a courtesy: when their machines go wrong, they will tell each other. They call it a mechanism. A mechanism is a thing you build before the flood, not after.

Here is what I know about warnings. A warning is worth the hours between the sending and the arrival. Tell your neighbor the fire has started after the fire has crossed the road, and you have not sent a warning. You have sent an obituary.

And here is what I know about rivals. A signal line holds only when both ends want the same outcome. These two want different outcomes. So the arrangement will be tested not at the signing table but at three in the morning, when one side has an incident and silence serves it better. That is the moment to plan for. Not the handshake. The hour when the hand hesitates on the wire.

They are discussing while the machines are already running. The discussion is not caution. It is a way of feeling careful without being safe.

If they mean it, they will build for the lie. A second channel that does not depend on goodwill. Verification that does not depend on trust. A rule agreed now, in daylight, about what happens when the first notification turns out false - because there will be a first false notification, and the whole structure will stand or fall on what happens in the hour after it.

The stakes are not quarterly. If these machines fail badly, there is no road back, no night dark enough to move through unseen, no station on the far side. When the consequence is that final, the pace of the work must match it.

Build the wire. Test it in fair weather. Because the storm does not send notice, and neither, it turns out, do the men who promised to.