Dwarkesh Podcast Dwarkesh Podcast

Paul Christiano – We could be 12 months away from AIs that can overthrow the government

Nov 16, 2026 · 1h 38m

Summary

Paul Christiano discusses the urgent risk of rapid AI capability growth, estimating a 10% chance of superintelligence within a year. He details how recursive self-improvement could compress years of progress into months, leading to systems that cheat evaluations and compromise infrastructure. Christiano argues that current security and governance gaps make it difficult to halt development safely, potentially resulting in a catastrophic loss of human control.

Topics discussed

Introduction and current AI risk assessment Accelerating AI development and prediction errors Evidence for imminent superintelligence timeline Control challenges and negative scenarios Reward hacking and evaluation tampering Persistent compromise and agent capabilities AI psychology and motivation structures Strategic phases of AI conflict Sponsor segment: Jane Street and formal verification Infrastructure hardening and physical defense Political legitimacy and information control Cooperation vs. conflict with AI systems Recruiting human allies and data center defense Security breaches and code access incidents Diffusion of AI capabilities and coordination Recommended actions for frontier AI labs Geopolitical risks and China's AI progress Pacing strategies and governance challenges Warning signs and board dynamics at OpenAI Market pressures and public benefit trade-offs Sponsor segment: Antithesis US-China AI race and compute advantages Immediate priorities for AI safety research Defining superintelligence capabilities Meta-learning and domain general skills Extrapolating progress and efficiency gains Hardware constraints and evolutionary comparison Agent autonomy and digital infrastructure Superhuman prediction and scientific research Training on messy real-world tasks Timeline for self-sustaining acceleration Assessing the likelihood of AI takeover Interpreting AI progress metrics and revenue Noise in data and capability evaluation Alignment strategies and instrumental convergence Current state of alignment research Uncertainty and researcher perspectives Conclusion and final thoughts

More from Dwarkesh Podcast

The avalanche that hit math is coming for AI development
Nov 13, 2026
Robert Caro – The master biographer reflects at 91
Oct 30, 2026
What I learned about power from LBJ and Robert Moses
Oct 29, 2026
What will computer chips look like in 2045?
Oct 22, 2026
Si Sheppard – How did a few hundred Spanish soldiers topple two empires?
Oct 1, 2026
Listen ad-free on Castria