Dario Amodei posted his essay “We Must Pace the Frontier” on X. OpenAI, xAI, and senior figures at Google DeepMind have indicated alignment with the core direction. Sam Altman committed OpenAI to the embedded evaluators step. Elon Musk endorsed the post. Several DeepMind researchers signed the related July statement calling for tools to pace automated AI development.
The underlying argument is straightforward. Current alignment, monitoring, and interpretability methods are not yet sufficient for the rate of capability improvement we are seeing. Recursive self-improvement amplifies this gap. If models can accelerate their own development while those foundations remain incomplete, the risk of losing control increases. A few additional years focused on improving alignment techniques, monitoring systems, and interpretability research would allow meaningful progress on these bottlenecks.
Dario’s three-step framework is a practical response. The embedded evaluators proposal is particularly useful. Giving independent teams ongoing, employee-level access creates a clearer mechanism for verifying safety claims than relying solely on internal reporting.
China remains the primary constraint, which Dario acknowledges. Any coordinated pacing among democratic labs must account for the risk of ceding technical leadership. That does not eliminate the case for buying time to strengthen the underlying safety infrastructure.
Sources: