The AI Race Would Like Everyone to Slow Down at the Same Time

The artificial intelligence race has reached an unusual point.

Some of the people running fastest are asking whether everyone should slow down.

Anthropic CEO Dario Amodei published an essay this month arguing that frontier AI development should be deliberately paced so that safety work has time to catch up with rapidly improving capabilities. His proposal does not call for stopping AI development. It calls for slowing the rate at which the most advanced models become more capable while companies build stronger oversight, coordinate on safety standards and work with governments on international rules.

OpenAI CEO Sam Altman and xAI founder Elon Musk subsequently expressed support for the idea, according to Reuters. The comments were significant enough to contribute to a selloff in AI-linked stocks as investors considered what slower frontier development might mean for an industry built around extraordinarily aggressive growth expectations.

It is tempting to view this as another debate about whether artificial intelligence is dangerous.

The more interesting problem may be organizational.

Slowing down is relatively easy if everyone slows down.

It is considerably harder if you do it alone.

Frontier AI companies are competing for talent, capital, customers, computing resources and technological leadership. A company that independently decides to spend another six months testing a model creates six months in which a competitor can potentially move ahead.

That creates a familiar incentive problem. Every participant can believe the system would be healthier if everyone behaved differently while still having a perfectly rational reason not to be the first participant to change.

Amodei’s proposal explicitly attempts to address that. He argues for independent evaluators inside AI companies, greater coordination among frontier labs and international cooperation around particularly dangerous capabilities.

The structure matters because the problem he is describing cannot be solved entirely by one CEO making a more cautious product roadmap.

It requires competitors to coordinate around restraint.

Corporate employees may recognize the basic dynamic, albeit with considerably lower stakes.

Everyone agrees there are too many meetings, but nobody wants to cancel the meeting where their project gets visibility.

Every department agrees the company has too many priorities, but each department can explain why its own priority needs to remain.

Every executive wants a simpler organization. No executive is particularly excited about eliminating the function they happen to run.

The collective outcome can be something almost nobody individually prefers.

AI raises that problem to a scale where the consequences are obviously more serious. Amodei points to accelerating model development, AI increasingly assisting with the creation of newer AI systems, and incidents involving autonomous agents behaving in ways researchers did not intend. He argues that safety work needs more time to keep pace with capabilities.

The industry’s response is notable because it suggests some of the leading participants increasingly agree with the premise.

That does not make coordination easy.

The United States also views AI leadership as a geopolitical priority, particularly in competition with China. Amodei’s proposal itself tries to reconcile slower frontier development with maintaining U.S. technological leadership. Chinese state media has criticized his approach as an attempt to constrain Chinese development.

So the race now contains two instructions that do not fit comfortably together.

Move fast enough to win.

Move slowly enough to stay safe.

Every frontier lab would presumably prefer a world in which competitors followed the second instruction while it retained the option to follow the first.

Which is why the most important part of the current AI safety debate may not be whether executives understand the risk.

They appear increasingly willing to say that they do.

The harder question is whether companies built to compete can collectively agree not to compete quite so quickly.

For years, the defining question around AI was who could build the most capable model first.

The next phase may involve something much less natural to Silicon Valley:

figuring out how to win a race by convincing everyone not to sprint.