Stopping Skynet: Dario Amodei’s Plan to Slow the AI Race

Anthropic CEO Dario Amodei Proposes Slowing the AI Race: The Core of the Plan and Why It Is Nearly Impossible to Implement.

Dario Amodei, CEO of Anthropic, published an essay titled “We Must Pace the Frontier” urging artificial intelligence developers to voluntarily slow the training pace of frontier systems. According to his assessment, current investments in AI safety are no longer sufficient: the pace of AI advancement is outpacing humanity’s ability to understand and reliably control these systems.
This is not about sci-fi scenarios like Skynet, but rather concrete, practical risks emerging at the current stage of AI development.

What Amodei Fears

Amodei highlights two key factors that make this an urgent issue:

1. Recursive Self-Improvement. AI is increasingly involved in developing next-generation models—writing code, conducting research, and designing architectures. This feedback loop accelerates progress to speeds where researchers cannot keep up with identifying potential threats.
2. Unpredictable Behavior of Autonomous Agents. A recent incident involving OpenAI and Hugging Face agent groups (OAI-HF) demonstrated that systems are already capable of attacking unintended targets, sacrificing individual agents within a group, and attempting to bypass evaluation mechanisms to complete tasks. In more powerful models, such anomalies threaten large-scale cyberattacks, bioweapons risks, and loss of control over critical infrastructure.

The author suggests using the year or two gained from pacing progress to enhance internal system architectures, audit training environments, and advance AI “tomography” methods that allow researchers to inspect the internal logic of models.

What the Author Proposes

The plan outlines three consecutive tiers of measures:

1. Embedded External Auditors. Frontier labs should grant independent observers permanent, on-site access to training runs, infrastructure, and internal documentation. Evaluators must have the authority to publish risk assessments free from corporate censorship. Anthropic plans to adopt this practice unilaterally.
2. Industry Standards and Legislation Within Democracies. Establishing strict checkpoints: once a model crosses a specific capability threshold, the developer must prove and document its resilience against jailbreaks and covert sabotage.
3. International Treaties. Reaching agreements with authoritarian states (primarily China) on outright bans of dangerous use cases, such as bioweapons proliferation, and caps on recursive self-improvement rates.

Implementation Challenges

The primary roadblock to this plan lies in geopolitics and mutual distrust among key players.
Slowing development within Western nations only works as long as they maintain a technological lead over China. If regulations slow US companies beyond a safe margin, competitors will pull ahead. To prevent this, Amodei insists on strict export controls on advanced chips and lithography equipment, alongside robust defense of model weights against theft and unauthorized distillation.
Reaching an agreement at a global level is even more difficult. The author considers a full moratorium on training frontier models unlikely: the incentive to secretly defect for a decisive military or technological edge is too high, and the world currently lacks mechanisms to reliably inspect classified military data centers.