1. Combat Resignation on AI Risks
Recognize that humanity has overcome bleak situations before, and collective action and dialogue can shift the world from an “inevitable” future to a “possible” one where AI risks are managed. The belief that it’s impossible makes it so; instead, view it as hard but achievable.
2. Advocate for International AI Agreements
Support the creation of internationally coordinated agreements, such as a US-China bilateral treaty, to prevent the reckless race to build superintelligence until its safety can be deeply understood and ensured. Unilateral action by one entity is insufficient to address this global challenge.
3. Support Compute Governance for AI Safety
Advocate for policies that govern the primary resource for powerful AI systems, including tracking AI chips, consolidating data centers, and imposing training thresholds (e.g., 10^24 flops) to prevent pushing the frontier of generally capable AI. This provides visibility and intervention capability.
4. Prioritize Deep AI System Understanding
Emphasize the critical need for a radical, deeper understanding of how AI systems actually reason, work, and internalize goals, rather than relying on superficial behavioral “papering over” of problems, especially when dealing with superintelligence.
5. Adopt a High Standard for AI Risk Tolerance
Recognize that for technologies with civilization-wide implications, even double-digit probabilities (e.g., 25%) of catastrophic or extinction risk are unacceptably high, demanding extreme caution and a re-evaluation of current development speeds.
6. Discriminate AI Development Risks
Understand that not all AI is existentially dangerous; focus concern and efforts on the specific category of “racing towards generally capable, superintelligent AI systems” rather than broadly opposing all AI applications like chatbots or self-driving cars.
7. Challenge Cynicism on AI Risk Claims
Be skeptical of arguments that AI CEOs’ concerns about catastrophic or extinction risks are merely marketing hype; consider their consistent historical statements and the counter-intuitive nature of scaring investors or regulators with such claims.
8. Prevent Proliferation of Dual-Use AI
Support the establishment of red lines and domestic rules to prevent the broad proliferation of powerful dual-use AI capabilities (e.g., for cyber vulnerability discovery or bioweapon design) into the hands of incautious or malicious actors.
9. Understand AI Orthogonality Thesis
Grasp that intelligence and goals are orthogonal, meaning any level of intelligence can be paired with any goal; this helps avoid the intuition that higher AI intelligence inherently correlates with “goodness” or human-compatible morality.