Pacing AI won’t solve the governance gap

In May 1972, Richard Nixon and Leonid Brezhnev signed the SALT I agreements, aiming to curb weapons that could lead to mutual destruction. Verification relied on technical means, such as reconnaissance satellites. Dario Amodei's essay "We Must Pace the Frontier" highlights that AI advancement is outpacing our ability to control it, especially as AI can build future AI versions.

Amodei suggests a three-step plan: independent evaluators should have access to AI labs, leading labs should align on safety standards, and international agreements should be pursued, even with China. While some agree with the diagnosis, the proposal's economic implications pose challenges. Independent verification can coexist with a market that values speed, but collective restraint requires altering economic incentives.

However, the core issue lies in the fact that building is inherent to scientific progress, and asking the entire technical field to limit their advancement is fighting economic forces and a fundamental aspect of scientific research. Coordinating AI safety is essential, but it cannot be the sole foundation. Aviation safety, though improved through regulations, still faced challenges due to structural issues and misunderstandings.

The same applies to AI, where rules can be ineffective if there's limited visibility into complex systems. Therefore, while international coordination is necessary, it should not solely rely on achieving a global slowdown of AI development, as this outcome is unlikely.