Anthropic CEO Dario Amodei is calling for slower AI development and has already granted third-party evaluator METR broad access to Anthropic models to audit safety practices. This is not a hypothetical position. It is happening now, unilaterally, without waiting for industry consensus.
Amodei's plan has three steps. Step one is external evaluation, which Anthropic is doing today. Step two requires industrywide coordination, likely involving government. Step three is not yet detailed in the summary, which is exactly why the full essay matters. The sequencing and the conditions Amodei sets for each phase are where the real argument lives.
The phrase 'pace the frontier' is Amodei's term for deliberately slowing training runs to let safeguards catch up with capabilities. Whether competitors follow or use this as an opportunity to accelerate is the open question. Read the original essay at darioamodei.com to see how he proposes to make slowing down a coordinated act rather than unilateral surrender.
[READ ORIGINAL →]