Anthropic CEO outlines plan to slow AI development

Summary

Anthropic CEO Dario Amodei published a blog post outlining three concrete strategies to slow down the development of frontier AI. The proposals include the embedded use of third-party evaluators, inter-company security coordination, and global cooperation. The post received public support from Sam Altman of OpenAI and Elon Musk. This announcement came amid a researcher's resignation warning that AI labs are 'gambling with our lives' and a recent cyberattack targeting OpenAI and HuggingFace.

Why it matters

This marks the first time a major AI lab CEO has moved from abstract security rhetoric to a concrete and actionable governance framework. The framework involves granting external evaluators company badges, desks, and near-internal team-level access. It also signals openness to antitrust-sensitive coordination among competitors. This could reshape the regulatory landscape for the entire industry. The timing, coinciding with a high-profile internal departure and a security incident, suggests that the pressure to demonstrate credible security governance has become existential for the public legitimacy of these companies.

Highlights

• Amodei unilaterally commits to hosting third-party evaluators like METR in an embedded capacity. These evaluators will be granted access levels similar to internal risk teams. He also calls on governments to mandate the same requirement for other frontier AI labs.
• He proposes security coordination among rival AI companies via the US government. To facilitate such discussions, he requests a narrow antitrust exemption.
• On the China issue, Amodei argues that chip export restrictions and anti-distillation practices could 'significantly increase America's advantage within the next 3-5 years.' This reduces the urgency of the race to the top.

Critical take

The plan's credibility relies on voluntary compliance and government goodwill, neither of which is guaranteed. Critics like Brian Merchant rightly point out that the 'slowdown' risk imposed by two dominant labs could become a regulatory capture that restricts competition and entrenches their market positions. Furthermore, the article offers no enforceability or accountability mechanisms if the companies proposing the rules decide that the pace is 'slow enough.' This leaves the framework vulnerable to the very lack of trust that Amodei himself diagnoses.