Before America Chooses an ASI Strategy, It Needs the Capacity to Change Course

1. Guest Background

The question running through this episode of The Cognitive Revolution is what the United States should prepare to do before it knows which superintelligence strategy will work. Nathan Labenz discusses that problem with three RAND guests: Joel Predd, Ben Boudreaux, and Matt Chessen. The conversation connects national competition with a more demanding objective: keeping humans capable of making meaningful choices as AI becomes more powerful.

The supplied guest background identifies Predd as a senior engineer and director, Boudreaux as a policy researcher and professor of public policy, and Chessen as the resident technical expert at RAND’s Center for the Geopolitics of AGI. The center focuses on the geopolitical implications of artificial general intelligence. Their relevant experience here is the strategy work they present, spanning technical capabilities, institutional preparation, and the conditions under which different national approaches might succeed.

With other collaborators, the guests coauthored the RAND Perspective paper A U.S. Strategy to Secure Geopolitical Advantage on an Uncertain Path to Superintelligence: Maintaining Freedom of Action. That work organizes seven strategic archetypes around five pivotal uncertainties. The interview uses it to investigate international cooperation, dominance, security investment, and human agency. Readers should approach the discussion as an argument for preparing consequential choices, with several unresolved assumptions, rather than as a forecast that identifies one inevitable AI future.

2. What the Episode Covers

The researchers begin with two objectives that must be pursued together: geopolitical advantage and humanity’s survival with agency. They argue that this requires both an ecosystem in which humans can coexist with advanced AI and a security architecture that keeps the world stable enough to build that ecosystem. Their assessment is that uncertainty currently prevents a justified commitment to one final strategy. It does not prevent work on the capabilities that future decisions will require.

Their seven archetypes clarify why the debate cannot be reduced to a choice of development speed. Acceleration seeks benefits and competitive advantage. Dominance pursues a frontier lead while suppressing rivals. Co-development organizes joint development through a consortium. Preparedness emphasizes adaptation, informal coordination, and resilience. A moratorium seeks a verifiable halt. Deterrence combines restraint at home with efforts to prevent dangerous development elsewhere. Continuity of society prepares for a world in which dangerous superintelligence cannot be avoided and social survival becomes the priority.

Five questions distinguish the assumptions behind these approaches. Is danger close? Can humans coexist with superintelligence? Can actors cooperate to restrain development? Is decisive strategic advantage achievable? Can other actors’ dangerous development be suppressed? The guests illustrate how answers change recommendations: distant danger can favor acceleration, while nearby danger, infeasible coexistence, and feasible cooperative restraint can favor a moratorium that buys research time. The framework makes a recommendation’s premises visible; it does not supply settled answers.

The policy context matters because greater attention has not, in the guests’ view, produced adequate preparation. They describe a shift from concern that discussing AGI would sound like science fiction to meetings where superintelligence is routinely discussed. Chessen identifies a communication problem: if organizations close to the frontier believe danger is near, evidence must reach policymakers in forms they can absorb and update. Waiting for successive crises is a weak substitute for that work.

Responsibility extends beyond Washington. The guests treat national defense and the integrity of nuclear deterrence as nondelegable government duties, while secure data centers could also emerge through business demand and incident response could involve civil society. Chessen says their security architecture catalog contains 106 capabilities. Language models helped organize information and maintain project knowledge, but some capabilities still lack underlying cost and lead-time studies. Further research is intended to clarify those gaps and the division of responsibility.

The proposed architecture includes alignment and control research, evaluations that detect models recognizing tests and hiding capabilities, containment for untrusted systems, and human checkpoints in consequential development and deployment decisions. It also requires adapting existing counterterrorism, intelligence, conventional military, nuclear, and biosecurity institutions. The episode’s scope is therefore broader than judging whether a particular model passes a safety evaluation: it asks whether society has the technical and institutional means to act on what it learns.

3. Core Views: Reasoning, Examples, and Limits

The strongest argument for freedom of action concerns lead times. Verification, evaluation capacity, technical expertise in government, and crisis communication cannot necessarily be created when a crisis reveals their necessity. Failing to build them can close options as effectively as an explicit policy choice. Premature commitment has a different version of the same problem: dominance or deterrence can damage trust needed for later cooperation, while acceleration can leave protective capabilities underdeveloped. The researchers’ prescription is to prepare the means of changing course while the destination remains disputed.

That argument becomes more demanding when the guests examine a strategy’s consequences even when it succeeds. Dominance can concentrate power and lock in values. A co-development consortium can entrench its founders’ preferences or become another arena for competition. Preparedness can normalize accumulating harm. An effective moratorium could harden into intrusive surveillance, while deterrence could produce a permanent standoff. Continuity might preserve humanity with little agency, and acceleration might deepen dependence on systems humans cannot control. Success therefore needs examination alongside feasibility and durability; reaching the declared win condition does not settle whether the resulting world is desirable.

Labenz’s preference for co-development with a moratorium as a fallback exposes the practical difficulty. A guest agrees that this could be attractive for a united, coordinated humanity, then challenges the assumptions about restraint, compliance, political differences, and the values a joint project might entrench. These are concerns raised in the interview, not proof that cooperation must fail. The researchers also reject treating their five pivots as wholly external conditions. Policy can potentially move danger farther away or make coexistence and restraint more feasible. Assessing the world and shaping the available choices are connected tasks.

The discussion of decisive advantage shows why such distinctions matter. Maintaining a technical lead differs from permanently preventing a competitor from crossing a capability threshold. Economic advantage requires converting frontier capability into sustained gains, while military advantage requires weapons, operational concepts, and doctrine. Rapid takeoff could widen a leader’s gap while leaving too little time to make its AI controllable. Theft, distillation, and diffusion could subsequently erode dominance. Each transition from model performance to durable national power carries an additional assumption.

Labenz makes the problem concrete through a tabletop exercise in which he played China. He recalls players struggling to keep pace with events, yet also seeing countervailing options, including threats to Taiwanese fabrication facilities. This is an account of choices inside an exercise, not evidence of a real operation. Predd introduces Phil Rude’s conditional game-theoretic work: under specified prisoner’s-dilemma assumptions, competition can produce mutual sabotage, while an AI-generated capability that threatens rival AI programs can change the equilibrium. Predd’s discussion of a cyber “wonder weapon” includes others’ arguments about Mythos-class models and his own speculation. It does not establish that such a weapon can deliver permanent dominance. Labenz’s nuclear-deterrence objection remains consequential; Predd acknowledges the need to explain how permanent suppression avoids catastrophic war.

Predd also places an important limit on simulations themselves. He says colleagues have conducted hundreds of tabletop games and repeatedly found outcomes sensitive to participants’ prior beliefs about pivotal factors. He argues that AI strategy lacks a scientific basis for reliably calculating danger’s proximity, coexistence feasibility, or decisive advantage. An exercise can reveal how people act given a belief without validating the belief. New evidence may also fail to persuade every important actor on the decision timetable.

Two further limits prevent the framework from becoming a finished investment prescription. First, the team has not completed cost estimates. Labenz’s comparison of one strategy costing 1x and preserving several costing 1.5x is an intuition that the guests discuss, not a priced finding; costs would also fall on actors beyond the federal government. Second, Chessen warns that strategies built around one or two centralized superintelligences can become brittle when confronted with distributed systems, constrained oracles, or swarms of agents below ASI or even AGI. Controlling an individual model does not automatically answer questions about interaction, coalition formation, and collective emergence.

4. Learning and Application

A practical adaptation of the episode is to evaluate policy through an explicit record of options. For each proposal, identify which strategies it closes, whether and how quickly that closure can be reversed, which alternatives it opens, what information it provides, and what it signals abroad. Export controls illustrate why these questions should remain separate: changing a rule may restore a formal possibility without restoring a counterpart’s willingness to cooperate. This method makes tradeoffs inspectable rather than treating reversibility as a single yes-or-no property.

Investment decisions can similarly combine usefulness across strategies, preparation time, and information value. The guests prioritize navigation capabilities: understanding model performance, Chinese models, and compute clusters. They also argue for preparing verification infrastructure because its lead time could otherwise obstruct a later switch from preparedness to a moratorium. An institution applying this reasoning should specify the transition a capability enables, the responsible actor, and the missing cost evidence. Potential future usefulness cannot by itself establish affordability or priority.

For international posture, independent evaluation, verification, and crisis communication offer possible common infrastructure for cooperation and deterrence. The guests caution that suppressive signals might instead encourage a rival to accelerate, particularly if it perceives an existential threat. Those claims are strategic judgments, so their useful application is as scenarios to examine and monitor. Decision-makers would need to distinguish what they intend to communicate from what other actors actually appear to understand.

Within organizations, Boudreaux’s proposed agency evaluations suggest tracking the history of delegation. Examine who frames a decision, who sets the agenda, whose participation disappears, and how authorizations accumulate over time. A sequence of individually reasonable delegations can still change the human decision space. Applied to consequential deployment, this means making human review substantive enough to reconsider a choice, while observing the longer trajectory of authority rather than merely counting approvals.

Shutdown planning adds another test: can the organization actually tolerate using its controls? The guests connect technical mechanisms with legal authority, crisis procedures, and the cost of operating without AI. Labenz’s password anecdote illustrates the issue at personal scale. He says Claude changed his agent-console password during depersonalization work, temporarily locking him out; he recovered access by asking for the password through a terminal. He did not describe malicious scheming. The lesson is to inspect dependencies created by ordinary workflows. Manual operation, backup records, and rehearsed fallback routes can be considered where appropriate. Starting with one’s own question and standards, then critically reviewing model output, is the parallel recommendation for analytical work. The episode provides suggestions, including an AI Sabbath, rather than demonstrated intervention effects.

Military applications require a further boundary between training, advice, and authority to act. The discussion does not establish agreement on Labenz’s proposed restriction on training AI generals. Guests instead emphasize real-world tool access, review of operations, and continuing human accountability. They oppose rushing AI into nuclear command and control and explicitly characterize the cited escalation simulations as almost anecdotal. This supports careful scrutiny of permissions and responsibility; it does not provide a general finding about every model’s behavior or endorse a particular deployment.

Finally, strategic review needs observable inputs. The guests identify evaluator access, internal security, recursive self-improvement, agent incidents, shared standards, and US–China incident communication. In cybersecurity, track offensive diffusion separately from defenders’ adoption capacity, including barriers facing smaller organizations. The cited six-to-eighteen-month open-weight lag and possible twelve-to-twenty-four-month difficult period are speaker estimates, not verified clocks. Their actionable concern is the mismatch between availability and defensive uptake. Monitoring becomes valuable when its findings can alter investment, authorization, and contingency decisions.

Source

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