We Need AI Treaties. This is How We Get Them
Summary
Host Tristan Harris discusses AI verification with experts Tim Fist and Janet Egan, drawing parallels between nuclear non-proliferation and current AI governance challenges. They argue that international coordination on AI risks, such as recursive self-improvement, requires robust verification technologies to build trust between adversaries like the US and China. The conversation explores leveraging existing chip hardware features, cryptography, and physical inspections to monitor compute usage in frontier AI data centers. By focusing on the concentrated supply chain of AI chips, they propo…
Topics discussed
Introduction: Oppenheimer, nuclear verification, and the AI analogy
Why global AI coordination is urgent and necessary
Lessons from nuclear treaties: CTBT and INF verification
The 'Trust but Verify' principle and current US-China dynamics
Compute as the choke point: Monitoring AI chips and telemetry
Technical solutions: Trusted Execution Environments and cryptography
Challenges in tracking unaccounted chips and data centers
Comparison of AI chip inspections to IAEA nuclear safeguards
Overcoming skepticism: Why verification is possible now
Industry role and the need for imperfect but actionable solutions
Timeline pressures: 18-month window for recursive self-improvement
Diplomatic engagement and the need for independent experts
Policy recommendations: Conditional export controls and incentives
A roadmap for safe AI development and international cooperation
Conclusion: Hope for a humane future and call to action
Listen ad-free on Castria