Anthropic chief calls for slower advances in AI capabilities

Dario Amodei called on AI laboratories on September 12 to slow improvements in model capabilities to allow more time for safeguards. Reuters reports that Sam Altman and Elon Musk publicly expressed agreement with his intervention. [2]

His essay proposes independent evaluators embedded in companies, coordination on safety standards and international cooperation with verification. He does not call for all training to stop: he wants capability advances to leave time to check safeguards. [1]

The damage scenarios Amodei discusses are risks he describes, rather than established future events. His commitment to outside evaluators does not yet demonstrate that every oversight arrangement has been implemented. [1]

Reuters highlights support from other executives, alongside the economic pressures of competition between laboratories. Its report connects the safety proposal with commercial incentives that could complicate implementation. [2]

The Daily Caller gives more space to U.S. congressional debate and positions expressed by Republican and Democratic officials. It also reports Anthropic’s request for national testing requirements. The cited bills and statements should not be described as an already adopted system. [3]

The issues to follow are concrete: evaluators’ actual access, publication of their findings and requirements ultimately adopted by government. Agreement in principle among executives does not yet answer those implementation questions. [3]