Anthropic CEO Dario Amodei has proposed a strategy to slow the development of artificial intelligence. He cited the recent OpenAI-HuggingFace hack and the rapid acceleration of AI capabilities as primary reasons for adopting a more cautious approach to model improvement.
The company is unilaterally committing to the use of embedded evaluators from third-party organizations like METR. These evaluators will receive company badges, desks, and access to internal systems to verify safety commitments and report incidents, similar to regulatory oversight in banking.
Amodei advocates for coordination among AI companies in democratic nations to establish common safety standards. He suggests that the US government should provide antitrust waivers to enable these discussions, addressing concerns that such cooperation might otherwise trigger regulatory scrutiny.
To address the competitive landscape, Amodei suggests that the US government restrict the sale of advanced chips and semiconductor equipment to Chinese firms. He believes these measures, combined with restrictions on model distillation, could maintain a lead over China for the next 3–5 years.
The proposal also includes a call for global coordination with authoritarian regimes, including China. Amodei suggests focusing on narrow agreements, such as prohibiting the use of AI for the production of biological weapons, despite acknowledging the limitations of such cooperation.
