The Scaling Curve: Dario Amodei, Anthropic, and the Race to Build and Survive Superintelligence — Chapter Ten

1. 1. Dario Amodei of Anthropic developed the Responsible Scaling Policy (RSP) to address the critical question of how to manage increasingly dangerous AI models.
2. 2. The RSP introduces AI Safety Levels (ASLs), a framework inspired by virology biosafety levels, with each level correlating to a new category of risk and requiring specific safety measures.
3. 3. ASL-3 models are defined as those that could significantly broaden access to catastrophic CBRN (chemical, biological, radiological, nuclear) knowledge, enabling unskilled individuals to cause harm.
4. 4. At ASL-4, models might enhance the capabilities of sophisticated state actors or become autonomous enough to operate unpredictably, potentially deceiving safety evaluations.
5. 5. The RSP institutionalizes Dario's conviction that AI safety and capability are not separate disciplines, but rather integrated aspects of the same engineering challenge.
6. 6. Internally, the RSP aligns team incentives by making progress on safety research a prerequisite for deploying the next model, fostering a company-wide commitment to safety.
7. 7. The RSP's concrete evaluations and specified security measures enabled meaningful engagement with policymakers and national security experts, moving AI safety discussions beyond abstract concepts.
8. 8. Dario Amodei believes government action will eventually be required, advocating for surgical, evidence-based regulation, with transparency legislation for frontier AI companies as a starting point.