Key Concepts
- Applied AI/ML in National Security
- Agentic AI for Scientific Advancement
- High-Performance Computing (HPC)
- Data Security and Governance (FedRAMP, NIST 800-53)
- Partnerships (Academia, Commercial Industry, Frontier Labs)
- Explainability, Isolation, and Governance in AI Architecture
- Export Compliance
AI at Los Alamos: A Historical Perspective
Los Alamos National Laboratory has been involved in applied AI and ML for approximately 70 years, dating back to the 1950s. An early example is the "Los Alamos chess" program in 1956, running on the Maniac 1 supercomputer. Due to memory limitations, the program was designed without bishops, showcasing early resourcefulness in applied statistics and machine learning. The lab's early work involved developing Monte Carlo methods after the Manhattan Project, which are still in use today. The advent of AI agents has presented significant opportunities for accelerating scientific progress.
Agentic AI for Inertial Confinement Fusion (ICF) Capsule Design
Los Alamos is exploring generative AI and agentic AI to accelerate scientific discovery. A demonstration project involved an AI agent designing an ICF capsule for Lawrence Livermore National Laboratory. The agent was tasked with reading relevant papers and generating a fusion capsule design. Uniquely, the agent executes code on high-performance computing assets to run thermodynamic and hydrodynamic tests. This approach leverages decades of mathematical and scientific knowledge related to nuclear stockpile management. The agent produced a design that optimized yield in simulations.
National Security AI Office: Mission and Partnerships
The National Security AI office at Los Alamos is tasked with advancing AI science, not just consuming existing tools. The lab develops its own models but also recognizes the need for partnerships with commercial industry and academia. They have partnerships with the UC family of schools and collaborations with frontier labs like OpenAI for chem-bio safety work. Los Alamos provides a safe environment for conducting potentially dangerous research.
Venado Supercomputer and Classified AI
Los Alamos houses the Venado supercomputer, featuring over 2500 nodes of Grace Hopper superchips, developed in partnership with Nvidia and HPE. This supercomputer is used to push the boundaries of AI research. OpenAI's models have been integrated into the classified networks at Los Alamos to address challenging problems specific to their data and mission.
Trust and Responsibility in AI Partnerships
Partnerships require trust, especially when sharing access to models and model weights. The responsibility for AI tool and service outcomes is crucial. The speaker references OM memorandum M25-21 (April), which codifies US government concerns about fielding AI systems. This memorandum emphasizes the need to move faster in adopting AI but also highlights the real-world impacts of government workloads. Data breaches can have geopolitical and kinetic consequences.
Navigating Security and Compliance: FedRAMP, NIST, and DoD Requirements
Software as a service (SaaS) companies must understand the responsibilities that come with handling sensitive data. While SOC 2 reports are common, the US government requires adherence to NIST 800-53 (Rev 4), which includes over a thousand security controls. FedRAMP aims to simplify compliance by vetting a subset of these controls, but the process can be challenging. The DoD adds further layers of security requirements through its Cloud Computing Security Requirements Guide (CCSRG) and CNSSI 1253, specifying impact levels for different data types (PII, mission data, operational data, finance data).
AI Governance and Risk Management
AI governance is still under development, with agencies creating strategies for AI implementation based on the April 3rd memorandums. NIST's AI Risk Management Framework (2023) provides guidance. Collaborating with government entities offers the opportunity to shape the future of AI governance.
Collaboration Opportunities with Los Alamos
Los Alamos offers unique collaboration opportunities due to its access to petabytes of data that has never been on the internet, subject matter expertise in various scientific fields, and capabilities in designing high-performance computing systems. The lab seeks partnerships to advance national security and competitive advantage.
Architectural Considerations for Agentic AI in Government
Four key considerations for bringing agentic tools and services to the federal government:
- Build for Explainability: Ensure transparency in decision-making processes to maintain trust and accountability.
- Build for Isolation: Leverage open-source tools and services to create isolated environments, especially for sensitive data.
- Build for Governance: Provide software bills of materials (SBOMs) and address open-source dependencies and patching plans to facilitate compliance.
- Keep Up the Speed: Maintain up-to-date services and address export compliance issues to ensure timely deployment.
Conclusion
Los Alamos National Laboratory views AI as both a significant opportunity and a threat to national security. The lab is committed to developing AI solutions, fostering partnerships, and addressing the ethical and security challenges associated with AI deployment. Their work in areas like nuclear non-proliferation has led to innovations such as the ChemCam sensor on Mars, demonstrating the broader scientific impact of their research. The lab seeks collaboration to push the boundaries of knowledge and address critical national security challenges.
AI summaries can miss context or contain errors. Check important details against the original video.