Anthropic CEO outlines plan to ‘pace the frontier’

AI companies will now allow external auditors to verify their safety practices, impacting how AI development is monitored.

Anthropic CEO Dario Amodei has proposed a three-pronged strategy to "pace the frontier" of AI development, citing concerns about rapid advancement and potential risks. He is unilaterally committing Anthropic to the first step: allowing "embedded evaluators" from third-party organizations to verify safety commitments and report incidents, a move he compares to embedded regulators in banking. Amodei also called for leading AI companies to coordinate on common safety standards and limits on AI progress, suggesting the US government provide an antitrust waiver for such discussions. To address concerns about China's AI development, he proposed the US and allies restrict chip sales and manufacturing equipment to Chinese companies, while also advocating for limited global coordination with authoritarian governments on prohibiting specific dangerous AI applications like biological weapons. These proposals come amidst heightened debate on AI safety, including a researcher's resignation from Anthropic over existential risk concerns.

Amodei's plan introduces a novel approach to AI safety by proposing "embedded evaluators," aiming to build trust and accountability through external verification. This could set a new industry standard for transparency, forcing other AI labs to adopt similar measures or face increased scrutiny. The call for government intervention to facilitate safety discussions, specifically through antitrust waivers, signals a potential shift in regulatory involvement, moving beyond passive observation to active enablement of industry self-governance.

The strategy also addresses geopolitical concerns by suggesting a coordinated effort to slow down potential adversaries' AI progress, which could reshape global AI competition dynamics. By proposing measures like chip export restrictions and model distillation crackdowns, Amodei is advocating for a more assertive stance in maintaining technological leadership, potentially leading to more protectionist policies within the AI sector.