A pause only buys time

No realistic pause ends the danger from AI. It can only postpone it. That is not an argument against pausing, but it changes how a pause should be judged: by what the time buys, not as a solution in itself.

Why every pause leaks

• The limit erodes. Algorithms keep getting more efficient and chips keep getting cheaper, so a fixed cap on the computing power used to train AI lets through more capable systems every year. A cap that holds has to keep tightening and reach ever smaller players. Algorithmic research is also much harder to monitor than large data centers. See Can AI progress be paused.
• One lab stopping hands the lead to the next. Anthropic has said as much: a pause by one lab alone would shift the lead without making anyone safer. See Conditional pause.
• One country stopping hands it to others. See Can the US and China agree to limit AI.
• Nothing is forgotten. Even a global agreement leaves the knowledge in place, so whatever stops can restart, and has to be kept stopped for as long as the pause lasts. If chips and algorithms keep improving meanwhile, the restart can be more abrupt than steady progress would have been: an overhang.

So a pause that holds needs every frontier actor, a limit that keeps tightening, and enforcement that never lapses. That is a great deal to keep up forever.

What would end the problem

A pause ends well only if something else happens during it. Alignment could be solved, so that building becomes safe. Verification could become good enough to keep a limit in place indefinitely at low cost. Or people could become capable enough, through better institutions or augmented intelligence, to handle what comes next. A pause is a bridge to one of these, never a destination.

This has two consequences. First, a proposal for a pause should say what its exit is: what has to happen during the pause, and what happens when it ends. Second, the value of pausing depends on progress elsewhere, which is why when to pause matters as much as whether. Time is worth most when it can be used best, and that may be late, when AI can help with the research, or early, before anyone can be sure where the edge is. See When should AI development pause.

The strongest objections

• Delay is valuable in itself. Every year without catastrophe counts, and the world changes in ways no plan foresees: new science, new governments, new ideas. A delay with no planned exit may still find one.
• There are chokepoints. A few companies make the most advanced chips and the machines that make them, so hardware progress can be controlled at its source. Algorithmic research needs computing power for its experiments, so a limit on compute slows algorithms too. See Compute governance.
• Some bans have held. The treaty banning biological weapons has held as a norm for fifty years, despite secret Soviet violations, and the ozone treaty phased out the chemicals it targeted. A delay that holds long enough is hard to tell apart from a solution.

What would change this

Evidence that algorithmic progress stalls without large amounts of computing power, or that control of the chip supply chain can keep capable hardware out of reach indefinitely, would make a lasting stop look feasible. Steady efficiency gains on small budgets would make the case that a pause only delays stronger.