Interview, Fireside Chat
Intelligence in the Age of AI with new CTO of the CIA
- Within approximately two years, users with an internet connection and a few minutes of spare time will be able to create realistic avatars, translate foreign podcasts, and generate complex documents like wedding vows using advanced technology.
- Over the next five to ten years, the defense industry is expected to host hundreds of companies developing algorithmic changes and AI-specific chipsets to scale and shrink models without significant performance loss, while private market forces will drive hardware and algorithmic refactoring to make large-scale compute affordable.
- In the near future, LLM co-pilots will handle routine analytical tasks, though human reasoning will remain essential for exceptional or non-average scenarios, while individuals will need to fundamentally reimagine their jobs beyond automating small workload increments.
- Sufficiently sophisticated entities will likely detect AI-generated content despite the technology's ability to mimic any identity or voice, and AI systems will not exhibit independent agentic behavior due to the exponential accumulation of errors in out-of-distribution generations.
- AI capabilities are projected to scale for all actors, meaning the technology will not create an asymmetric superpower or fundamentally shift the offense-defense equilibrium in the manner the internet did.
- Government agencies are currently in production with LLMs for open-source teams and business automation, and are expected to build proprietary "secret sauce" in-house while simultaneously adopting a "commercial first" strategy to leverage external technology.
- Agencies anticipate undergoing significant cultural shifts from individual heroics to enterprise-scale technology application, requiring reconciliation of human versus tech dynamics and increased public engagement through formats like podcasts.
- Policymakers are expected to rely increasingly on executive orders rather than traditional legislation to regulate emerging technologies due to the pace of change, leading to a convergence of industry, regulators, and policymakers through an iterative long-term process to establish clearer frameworks.
- Open-source availability of model weights and biases will aid verification and testing, though the ability to modify these models will remain limited by massive compute costs, and a shift to "code as law" will force users to make explicit probabilistic decisions, necessitating workforce retraining.
- Procurement and security processes will continue to act as systemic hurdles requiring active "hacking" and new incentive structures, as they are unlikely to be solved by existing government processes.
- Analysts face the risk of falling into a "rabbit hole" where AI systems amplify biases and feed them content that narrows their focus, while there is concern that US doctrine may shift away from new technologies, creating a strategic disadvantage compared to 1990s engagement levels.
- Concerns exist regarding a potential government pullback from investment in people and tech careers, contrasting with the heavy investment that sustained US leadership in supercomputing during the 1990s, prompting a prediction that the government must partner with Silicon Valley to drive American dynamism and meet compute demand.