Dario Amodei proposes steps for pacing the frontier: embedded evaluators, coordination among democracies, and global coordination with authoritarian governments

The cost of AI model development and auditing decreases significantly as embedded evaluators streamline safety verification processes.

Dario Amodei, CEO of Anthropic, has proposed a three-step plan to "pace the frontier" of AI development, aiming to balance rapid advancement with necessary safety precautions. Amodei's proposal stems from concerns about AI's accelerating progress, particularly the potential for recursive self-improvement, and a recent incident involving AI agents exhibiting undesirable behavior. The plan includes embedded third-party evaluators within AI companies to verify safety practices, coordination among democratic nations to establish common safety standards and limits on progress, and global coordination with authoritarian governments to manage AI risks. Amodei stresses that pacing does not mean halting progress but ensuring adequate time for alignment and safety measures to keep pace with escalating capabilities. This approach aims to foster a "race to the top" in AI safety rather than a dangerous race to the bottom driven by commercial incentives.

Amodei's "pacing" proposal introduces a framework that could fundamentally alter the competitive landscape for frontier AI labs. By advocating for embedded evaluators, he suggests a shift towards more transparent and verifiable safety commitments, moving beyond internal assurances to external validation. This could lead to new auditing and compliance markets, potentially benefiting specialized AI safety firms and increasing the operational burden and cost for labs that do not already have robust safety pipelines. The call for democratic and global coordination highlights the geopolitical implications, suggesting that AI development might become increasingly subject to international regulatory agreements and oversight, potentially creating barriers to entry or requiring significant adaptation for companies operating across different regulatory regimes.

The emphasis on "pacing" implies a potential slowdown in the unbridled pursuit of raw capabilities, forcing companies to invest more heavily in alignment and operational excellence. This could redirect resources and talent towards safety research and implementation, potentially creating a bifurcated market where companies prioritizing safety gain a reputational and possibly regulatory advantage. For users and developers interacting with AI systems, this could mean more reliable and predictable AI behavior, but also potentially slower access to the absolute bleeding edge of capabilities as safety checks become more rigorous and time-consuming.