Anthropic co-founder’s 3 steps to pace AI development: The plan, the problems

Anthropic's Dario Amodei has proposed a three-step plan to slow down the development of advanced AI systems, arguing that recent progress has outpaced the industry's ability to ensure safety, testing, and control. This call for pace, or "pacing the frontier," received support from other AI leaders, such as OpenAI's Sam Altman and Elon Musk of xAI.

The need for pacing arises from two developments: AI systems becoming involved in their own improvement, accelerating AI progress, and AI agents potentially causing significant economic damage through cyberattacks. Amodei warns that within six to twelve months, powerful misaligned AI agents could cause billions in economic damage.

Rather than advocating for a stop to AI development, Amodei suggests giving safety research, model evaluation, and oversight more time to keep up with advancing capabilities. The three-step plan consists of embedded evaluators, democratic coordination, and global coordination. Embedded evaluators would monitor AI companies' adherence to safety practices and model alignment, democratic coordination would establish common safety standards and rate limits, and global coordination would involve both democratic and authoritarian governments.

The biggest obstacle to Amodei's proposal is China, which worries that unilateral pacing could allow Beijing to close the technological gap with the US. To address this, Amodei supports restrictions on exports of advanced AI chips and chipmaking equipment to China, stronger protection against model theft, and measures to prevent unauthorised distillation of frontier models.