Pace AI — the case for being able to slow down

Nobody chose this pace, and nobody can leave it.

Nobody sat down and decided how fast this should go. The speed is a by-product of competition between the companies named above; it could have been faster or slower, and no part of it was planned. What holds it there is that all of them are under the same pressure. A company that slows down alone loses ground, and the frontier moves on without it, so none of them do. The people inside can see this happening, and none of them can fix it from where they sit.

The work is starting to do itself.

Models now write the code, help design the experiments and help read the results. The people inside describe raising their forecasts, then being caught out again months later, over and over. As that loop tightens, the speed stops depending on how fast people can work.

The ask is for an option, not a decision.

This is the part that gets misread. A brake that exists and is never touched would satisfy almost everyone making this argument. What is being asked for is a capability; whether it is ever exercised is a question for whoever is holding it when the time comes.

It cannot be built during the emergency.

Anything assembled in a hurry will be crude and badly aimed. Verification systems, reporting standards and international agreements take years to design and test. That work has to happen while there is still time to argue about it.

It has to be international, or it is nothing.

No single company or country can pace a frontier everyone else is still racing toward. A rule that binds only American labs slows only American labs. Verification across borders is the hard part, and most of it is unsolved.

A limit nobody can check is just a press release.

Thresholds set by guesswork will block work that is safe and wave through work that is dangerous. Whatever gets built has to rest on measurements outsiders can run themselves, by methods published in full.

Done badly, it makes things worse.

Many of the people who want this also distrust it, and they are right to. Oversight covering only publicly released models pushes capability into private ones. Coordination run carelessly hands more control to whoever administers it. Both are real risks, and both are arguments for designing this slowly and in public, beginning now.