သ(""); In response to the White House’s Request for Information regarding an AI Action Plan, Anthropic has provided recommendations to the Office of Science and Technology Policy (OSTP). Our suggestions aim to prepare the United States to harness the economic advantages and address the national security challenges posed by advanced AI systems. As articulated by our CEO, Dario Amodei, in ‘Machines of Loving Grace’, we anticipate that potent AI systems will arrive around late 2026 or early 2027. These advanced systems are expected to possess the following characteristics: - Intellectual abilities that meet or surpass those of Nobel Prize winners across various fields, including biology, computer science, mathematics, and engineering. - The capability to navigate all digital interfaces currently accessible to humans, which includes processing and generating text, audio, and video, autonomously controlling devices like mice and keyboards, and accessing and browsing the internet. - The ability to autonomously engage in complex reasoning tasks over extended durations—hours, days, or even weeks—seeking clarification and feedback as necessary, akin to a highly skilled employee. - The capacity to interact with the physical world by controlling laboratory equipment, robotic systems, and manufacturing tools via digital connections. Our recent developments further support the notion that powerful AI is on the horizon; our newly released Claude 3. 7 Sonnet and Claude Code exhibit notable enhancements and increased autonomy, alongside advancements from other leading labs. We assert that decisive action is essential for the United States to maintain technological superiority. Our submission highlights six critical areas to tackle the economic and security implications of powerful AI while maximizing benefits for all Americans: 1. **National Security Testing**: Government agencies should enhance their capabilities to assess both domestic and foreign AI models for potential national security risks. This involves creating standardized assessment frameworks, building secure testing infrastructures, and establishing expert teams to identify vulnerabilities in deployed systems. 2. **Strengthening Export Controls**: We recommend tightening semiconductor export restrictions to ensure that the U. S.

and its allies can seize the advantages of powerful AI systems, while preventing adversaries from accessing the necessary AI infrastructure. This includes regulating H20 chips, requiring government-to-government agreements for countries with substantial chip deployments, and lowering no-licensing-required thresholds. 3. **Enhancing Lab Security**: As AI systems evolve into crucial strategic assets, we suggest establishing classified communication channels between AI laboratories and intelligence agencies, expediting security clearances for industry personnel, and developing next-generation security protocols for AI infrastructure. 4. **Scaling Energy Infrastructure**: To maintain a competitive edge in AI development, we propose setting an ambitious goal of adding 50 gigawatts of dedicated power by 2027 and streamlining the permitting and approval processes. 5. **Accelerating Government AI Adoption**: We advocate for a comprehensive inventory across government workflows that would benefit from AI enhancement, assigning agency leaders to implement programs where AI can provide significant public value. 6. **Preparing for Economic Impacts**: To ensure that the benefits of AI are widely distributed, we recommend updating economic data collection methods, such as those used by the Census Bureau, and preparing for potential large-scale economic shifts. These recommendations build on Anthropic's earlier policy efforts, including our promotion of responsible scaling policies, testing, and evaluation practices. Our goal is to find a balance that encourages innovation while addressing the serious risks associated with increasingly capable AI systems.