Freedom of Action: America's ASI Strategy, w/ RAND's Joel Predd, Ben Boudreaux & Matt Chessen

What It Covers

RAND researchers Joel Predd, Ben Boudreaux, and Matt Chessen present their framework for US superintelligence strategy, outlining seven archetypal strategies—from acceleration to moratorium—and five pivotal uncertainties that prevent committing to any single approach. Their core recommendation: build capabilities for all strategies now to preserve decision-making flexibility as AI advances.

Key Questions Answered

• •Seven Strategy Archetypes: RAND identifies seven distinct superintelligence strategies, each with a coherent win condition and a pathology: Acceleration, Dominance, Co-development, Preparedness, Moratorium, Deterrence, and Continuity of Society. For example, a Moratorium strategy's win state buys time for alignment research, but its pathology is that enforcement mechanisms could harden into a global surveillance regime. Mapping these archetypes prevents policymakers from talking past each other due to unstated assumptions.
• •Five Pivot Factors as Decision Criteria: Before committing to any strategy, five binary questions must be resolved: Is danger close? Is coexistence feasible? Is restraint feasible? Is a decisive strategic advantage reachable? Is suppression feasible? Answers to these questions mechanically determine which strategy is rational. Currently, no evidence base exists to produce consensus on any of them, making premature commitment to one strategy a strategic liability.
• •Freedom of Action as Meta-Strategy: Because the five pivot factors remain unresolved, RAND recommends a meta-strategy of building 106 catalogued capabilities—spanning verification technology, evaluation infrastructure, crisis response channels, and resilient compute—that underpin all seven archetypes simultaneously. Capabilities with long lead times, like verification regimes, should be prioritized even if not required by every strategy, since they cannot be built quickly once needed.
• •Decisive Strategic Advantage Requires a Wonder Weapon: A dominant strategy only becomes viable if AI produces a discontinuous military capability—analogous to nuclear weapons—that rapidly shifts the balance of power. RAND's game theory modeling, led by colleague Phil Root, shows that under competitive conditions, mutual sabotage is the default equilibrium, not dominance. Offensive counter-AI tools or autonomous cyber weapons represent the most plausible candidates, but their existence remains speculative.
• •Signaling Discipline Prevents Strategic Foreclosure: Any action that signals a dominant or suppressive intent risks hardening a rival's resolve and accelerating their program. Chip export controls, for instance, may have catalyzed China's domestic GPU industry by making semiconductor independence existential. RAND recommends building capabilities—like independent evaluation and verification regimes—that credibly signal both deterrence and openness to cooperation simultaneously, preserving optionality without provoking escalation.
• •Coexistence Replaces Alignment as the Core Concept: RAND deliberately uses "coexistence" rather than "alignment" because alignment implies a specific technical solution set, while coexistence encompasses control mechanisms, architectural schemes, and unknown future methods. Practically, this means building agency evaluations that track longitudinal trends in human decision-making authority—mapping which high-stakes decisions are being delegated to AI systems over time—rather than only evaluating discrete task performance or dangerous capability thresholds.
• •Individual and Institutional Resilience as Strategic Assets: Maintaining the capacity to operate without AI makes shutdown threats credible and reduces systemic vulnerability. Concrete practices include: establishing AI-free operational environments for critical infrastructure, maintaining paper backups of essential records, preserving manual override capabilities for key systems, and practicing navigation or analysis without AI assistance. At the personal level, forming independent judgment before querying a model and applying critical review after preserves analytical capacity.