Pacing the AI frontier needs energy and coordination

Dario Amodei's call to "pace the frontier" is serious, specific and more honest than the usual AI safety theatre. He is right that today's models give us something worth studying, that an extra year or two could materially improve alignment, and that independent evaluators need access deep enough to test training pipelines rather than merely inspect finished products.

Yet the proposed industry-wide coordination leaves me uneasy. When the companies building and selling frontier models negotiate among themselves about how quickly they should advance, which risks count and who verifies compliance, that is a cartel conversation with better branding.

OpenAI's reported approach to Congress about whether a coordinated slowdown would breach antitrust law exposes the institutional problem: the features that make pacing credible - common triggers, independent verification and consequences for defection - are also the features that make an agreement among competitors look like an output restriction.

That does not make pacing the wrong objective. It means the labs cannot be allowed to own the mechanism.

President Trump has now rejected calls to slow development or regulate it more heavily, arguing that America leads China and that "whoever wins AI wins". I understand the national-security logic, and Amodei himself acknowledges it. But reducing a technology that its own builders say could threaten humanity to a US-versus-China race is dangerously reductive zero-sum thinking. If the risk crosses borders, a national lead is not the same thing as a human win. Winning a race over a cliff is still losing.

The answer is neither a private pact among laboratories nor political permission to accelerate without a credible brake. Governments should set capability thresholds, require independent evaluation and create narrow legal protection for genuine safety coordination, while preserving antitrust scrutiny everywhere else. They should also use the power they already possess as buyers. Procurement conditions tied to testing, incident disclosure and pacing would move laboratory behaviour faster than another voluntary pledge, because frontier companies can ignore critics more easily than customers.

Amodei has moved the argument in the right direction. The next step is to take control of the pace away from the people with the strongest commercial incentive to set it themselves. We already have successful models for how to achieve transparent global coordination on critical issues, including the International Atomic Energy Agency.

#ArtificialIntelligence #AIGovernance #AISafety

https://lnkd.in/g545vvp4