Interview, Fireside Chat
Intelligence in the Age of AI with new CTO of the CIA
- The CIA has pivoted from a CTE (Counter-Terrorism Era) focus to "great power competition," driven by Director Burns' business review which identified technology as a central, amorphous threat and opportunity.
- The agency established a new CTO function alongside a China Mission Center and a Transnational Technology Mission Center (T2MC) to address external technological threats and opportunities.
- CIA CTO Nand Mulchandani emphasizes a shift from inward-focused operations to external engagement, including public dialogues to better understand the technology landscape.
- Martin Casado notes that while Generative AI allows for deepfakes and impersonation, detecting AI-generated content is becoming increasingly feasible, suggesting it serves as a "cover in chaos" rather than an asymmetric superpower.
- Unlike the internet, which created asymmetric vulnerability for the US, AI is viewed as a technology that benefits all sides without fundamentally shifting the balance of power.
- The operational side of intelligence involves a "cat and mouse" dynamic where both adversaries and the CIA scale AI capabilities rapidly, treating it as a tool to enhance case officers and operations teams.
- The analytic side is shifting from a "pull model" (analysts querying data) to a "push model," where AI algorithms identify patterns and surface relevant data without explicit human prompts.
- A specific risk identified in AI analytics is the "rabbit holding problem," where algorithms amplify analyst biases by tailoring information to what pleases the user, potentially narrowing their perspective.
- Current Large Language Models (LLMs) are described as effective "co-pilots" for routine tasks and pattern recognition within training data distributions but lack "agentic behavior" due to the exponential accrual of errors when operating out of distribution.
- Intelligence work is characterized by "tail reasoning" (handling exceptions, new problems, and non-routine scenarios), which remains a distinctly human capability that LLMs cannot replicate.
- The CIA currently has LLMs in production, primarily within the Open Source team and for business automation, focusing on immediate, high-value use cases rather than long-term speculation.
- Mulchandani challenges analysts to reimagine their jobs over a 5-to-10-year horizon rather than simply seeking incremental 10-30% productivity gains through automation.
- The transition from "code as law" (where human decisions are hardcoded) to probabilistic AI systems requires policymakers to explicitly define threshold values for decision-making, as AI outputs now present probabilities rather than binary answers.
- The CIA faces a "buy side" culture clash, where rapid innovation in Silicon Valley conflicts with the rigorous, slow-paced requirements of national security procurement and security clearance processes.
- Open-source models present a trade-off: they allow for data customization but may inherit biases from the general internet, whereas proprietary models offer stability but less flexibility.
- The scale of modern AI requires massive compute resources, creating a "supercomputer" dynamic where only large entities (government or major corporations) can train models, limiting the utility of open weights for smaller entities without significant compute cycles.
- There is a strategic concern that the US is pulling back from direct investment in compute infrastructure, unlike the 1990s government-led investment in supercomputing that established US leadership in that era.
- Mulchandani advocates for "American Dynamism," a public-private partnership where Silicon Valley and the government "lean in" together, with the government acting as a sophisticated customer and supplier.
- The CIA is adopting a "commercial first" strategy, prioritizing the acquisition of off-the-shelf commercial technology before building bespoke solutions, provided security and ATO (Authorization to Operate) hurdles are cleared.
- The agency is undergoing significant cultural shifts to reconcile the individual-centric "heroic" spy tradition with the need for enterprise-scale technology, as well as moving from clandestine secrecy to public engagement.
- Casado warns against applying "internet-era" lessons of asymmetry and exponential risk to AI, arguing that AI is a new class of technology that requires distinct policy approaches focused on adoption rather than containment.
- Policymaking in emerging tech areas like AI, 5G, and crypto currently relies heavily on executive orders rather than legislation due to the rapid velocity of technological change and the difficulty of forecasting specific outcomes.
- The CIA's role in policy formation involves gathering intelligence on these emerging tech sectors to provide objective, "by-the-book" analytic support to policymakers who must navigate the balance between regulation and incentive.