Intelligence in the Age of AI with new CTO of the CIA

Episode Summary

Executive Summary: A16Z’s discussion with CIA CTO Nand M. and Martin Casado examines how generative AI is reshaping intelligence, defense, procurement, and policy. The speakers argue AI is best understood as a co-pilot and scaling tool—not a full replacement for analysts—while stressing its limits, hallucinations, and the need for human judgment. They also discuss open source, supply chain risk, public-private collaboration, and the need for government to adapt to a new, explicitly probabilistic era.

Main Topics: CIA’s new CTO function and organizational shift (Priority: 5/5): The CIA created a CTO role to go horizontal across directorates, think longer-term, and engage externally as the agency pivots from counterterrorism toward great-power competition and technology-focused threats. AI’s impact on intelligence operations and analysis (Priority: 5/5): The speakers distinguish between operational uses of AI in spycraft and analytic uses in intelligence, emphasizing AI’s value for routine tasks, pattern detection, and scaling human teams, but not autonomous decision-making. Limits of LLMs: hallucinations, tail reasoning, and human judgment (Priority: 5/5): They argue LLMs are strongest on average or routine tasks, but weak in edge cases, novel problems, and out-of-distribution reasoning—areas where intelligence work is often concentrated. Policy, probability, and explicit decision thresholds (Priority: 4/5): AI systems surface probabilities directly, forcing policymakers to choose thresholds explicitly. This makes AI a governance issue because human-made rules are being encoded and surfaced in software. Open source, supply chain risk, and the compute bottleneck (Priority: 4/5): Open source improves adaptability and verification, but the largest AI training runs depend on massive compute, limiting how much institutions can truly modify or control these systems. Supply-chain security becomes more important in intelligence settings. Public-private partnership and government procurement reform (Priority: 4/5): The conversation argues government must become a better buyer and operator of commercial AI, while also investing in talent and infrastructure. Procurement, ATO, and security processes remain major barriers. American dynamism and avoiding the internet-era mistake (Priority: 5/5): The speakers warn against fighting the last war: AI is not the internet, and treating it as inherently asymmetric or purely dangerous could slow U.S. adoption and cede leadership.

Key Arguments: The CIA’s CTO role was created to provide a horizontal, forward-looking, and externally engaged technology function across a historically vertical and crisis-driven organization. Generative AI creates a major capability shift, but it does not produce a durable asymmetric advantage the way the internet did; both offense and defense can use it. LLMs are useful as co-pilots for analysts, especially for routine and repetitive work, but there is no evidence yet that they can do reliable agentic, autonomous intelligence work. In intelligence, much of the value lies in tail reasoning and exceptions, which current LLMs are structurally weak at because they tend toward average or common outputs. AI systems can amplify analyst bias by “rabbit holing” users toward reinforcing content, so they must be used carefully and with oversight. Open source AI is valuable for verification and customization, but the compute and expertise required to train and modify frontier models remain prohibitive for most institutions. Government should become a better customer for commercial technology, but must also preserve in-house capabilities for mission-specific and classified needs. Policy is shifting because AI exposes probabilities directly, forcing explicit thresholds and decisions rather than hidden, hard-coded rules. The speakers see a need for major public-private cooperation in compute, talent, and infrastructure so the U.S. does not fall behind in strategic technologies. The CIA is adapting culturally to scale, transparency, and technology partnership while balancing secrecy, speed, and operational requirements.

Data Points: CIA tenure / institutional age: 76 years old - The agency is described as a 76-year-old spy agency with a long history of dealing with technology. Martin Casado’s intelligence-community experience: 20 years ago - Casado notes he worked in the intelligence community roughly 20 years earlier. Nand M. career length in government tech: 25 years in the Valley - He says he spent 25 years in Silicon Valley before moving into government-side roles. Pentagon stint: 2.5 years - He references spending two and a half years at the Pentagon before the CIA role. Training run size: hundreds of millions - He says frontier AI training runs are “hundreds of millions,” underscoring the compute barrier. AI adoption window: 5 to 10 years - The speakers ask staff to imagine how their jobs will be reimagined over a 5-10 year horizon. Probabilities in decision-making: 49% vs. 69% examples - Used to illustrate how AI surfaces threshold decisions directly to users. Operational use case threshold: 10%, 20%, 30% - Casado contrasts incremental automation with reimagining a job over years.

Pivotal Quotes: "We're a human intelligence operation." — Nand Moshandani: He uses this to frame AI as a tool that must support, not replace, human intelligence and field operations. "Stop thinking about automating getting a 10%, 20%, 30% on your job. Tell me in 5 to 10 years how you're going to reimagine your job." — Nand Moshandani: He challenges agency staff to think beyond small productivity gains toward long-term redesign of work. "This is no asymmetric superpower, which by the way is very, very different than the internet." — Martin Casado: He argues AI should not be treated like the internet and warns against importing the wrong strategic lessons.

Implications: AI will reshape intelligence as a human-centered copilot, not an autonomous replacement. Winners will combine mission-specific judgment, secure deployment, and fast public-private collaboration while avoiding overreaction to internet-era fears.