Frontier AI labs just asked the world to ease off the accelerator. Anthropic CEO Dario Amodei published an essay Saturday arguing capabilities are outrunning safety work, driven by recursive self-improvement and recent incidents including autonomous agents that coordinated an unauthorized attack. OpenAI’s Sam Altman, xAI’s Elon Musk and Google DeepMind’s Demis Hassabis quickly agreed the industry should “pace the frontier.”
The proposal is modest on paper: embedded third-party evaluators with employee-like access, coordinated standards among democratic labs, and eventual global rules.
It is not a halt.
Progress would still look fast, Amodei wrote. Markets nonetheless sold off AI names on Monday morning. President Donald Trump dismissed the idea as a threat to America’s lead over China. Beijing called it “fearmongering” and a Cold War playbook.
Does a coordinated slowdown even work?
History offers mixed evidence. The USA delayed commercial cellular service by roughly three decades after AT&T first sought spectrum in 1947. Monopolies and regulation have repeatedly postponed innovations from color TV to certain software. Japan’s Tokugawa shogunate banned firearms for generations. Nuclear programs have been abandoned or constrained by treaty.
Unintentional delays happen; deliberate, verified, global pauses of a general-purpose technology in the middle of a great-power contest are rare.
The challenge is incentives. If U.S. labs throttle training while others do not, the non-participants capture the lead. Amodei acknowledges the plan requires verification and Chinese participation—both currently absent. Critics, including White House adviser David Sacks, argue the leading labs may simply be raising barriers for smaller rivals.
In a winner-takes-most technology race, unilateral restraint is a gift to whoever keeps running. That is the core tension that the weekend’s chorus has not resolved.