PolicyLaw, regulation & governance
RAND urges the US to keep all its superintelligence strategy options open instead of committing to one now
A RAND Perspective sets out seven archetypal US strategies for superintelligence and five unresolved "pivot" uncertainties. It argues that none of the strategies can be justified yet, and recommends building now the capabilities that several of them share, so the US can commit once the evidence clarifies.
From the The Cognitive Revolution podcast episode of October 28, 2026 and cited sources: The paper groups competing plans into seven archetypes: dominance, co-development, preparedness, moratorium, deterrence, continuity of society and acceleration. [2] [1] [7]
SOURCES[2] RAND — A U.S. Strategy to Secure Geopolitical Advantage on… rand.org[1] Freedom of Action: America's ASI Strategy, w/ RAND's Joel… cognitiverevolution.ai[7] 3 Quarks Daily — The Strategy Before the Strategy 3quarksdaily.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- The paper groups competing plans into seven archetypes: dominance, co-development, preparedness, moratorium, deterrence, continuity of society and acceleration. [2] [1] [7]
- Original source
- The Cognitive Revolution podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 28, 2026 podcast episode published
- Covered
- Oct 29 edition
- Evidence
- 7 sources across 6 sites; transcript sourced
- 1Freedom of Action: America's ASI Strategy, w/ RAND's Joel…cognitiverevolution.ai
- 2RAND — A U.S. Strategy to Secure Geopolitical Advantage on…rand.org
- 3RAND — Finding Common Ground in AGI Strategy Debates…rand.org
- 4beSpacific summarybespacific.com
- 5Import AI 473 (via Prismix)prismix.dev
- 6Anne Stickells — Time Allowed for Adaptation in Nuclear…geopoliticsagi.substack.com
QuietHow it’s measured
YouTube video · 359 views · 4 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
The paper groups competing plans into seven archetypes: dominance, co-development, preparedness, moratorium, deterrence, continuity of society and acceleration. Which one is right depends on five unresolved points: how close dangerous AI is, whether humans can coexist with it, whether one actor can gain a decisive advantage, whether countries can agree to restrain development, and whether development can be suppressed. The authors argue that committing to one strategy too early closes off the others. [2] [1] [7]
The recommended response is to invest across four lines: a human–AI ecosystem, an AI security architecture, adapting legacy national-security institutions, and the capacity of society to respond. The authors call several items "no regret": visibility into compute, situational awareness, resilient infrastructure and biosecurity, evaluation and verification regimes, and crisis channels with rivals. They expect companies, not just the federal budget, to carry much of the cost. [2] [1]
The authors draw on nuclear history. US nuclear strategy moved over decades through a sprint, a temporary monopoly, proposed restraint, verified agreements and counter-proliferation. They argue AI may force similar shifts within months, so the options have to be built in advance rather than improvised. [1] [6]
For technical teams, the authors name research gaps. They want evaluations that can detect models that know they are being tested or are underperforming on purpose (sandbagging), containment built on the assumption that models are untrustworthy, and "agency" evaluations run in real deployment settings. They also warn that most strategies assume one or two centralized systems and become fragile if superintelligent behavior instead emerges from swarms of less capable agents. [1]
On the evidence: the paper's core claim is a qualitative analytical judgment, not an empirical finding.
Read the full assessment
On the implications: if policymakers adopt the framework, firms could face expectations around compute reporting, evaluations and incident response, and could be cast as co-funders of security work.
The reviewed sources cover RAND's paper, its earlier companion paper and three outside write-ups, so I'm now writing the dossier.
Executive brief
RAND's Center for the Geopolitics of AGI argues that the US should not yet choose any single strategy for superintelligence. Its reason is that none of seven competing "archetypal" strategies can be justified while five key uncertainties remain unresolved. The paper is *Maintaining Freedom of Action*, published September 15, 2026. It recommends investing now in capabilities that keep every strategy available. The authors admit the plan has no price tag yet. On The Cognitive Revolution (~01:10:00), Joel Predd says a peer reviewer called it "a vision" without one. RAND is now costing 106 capabilities listed in a Security Architecture Catalog.
What changed and event timeline
Diagnosis paper
Predd, Baker, Boudreaux, Geist and Chessen argue that two things must come first: an ecosystem where humans and AI coexist with humans keeping agency, and a stabilizing security architecture.
Freedom of Action published
Seven strategies, five uncertainties and four lines of investment. The paper warns that committing too early, or putting off a decision too long, both close off options.
Early pickup
Jack Clark's Import AI 473 summarized the paper and argued that US policy looks like acceleration alone; also summarized it.
Nuclear analogy
Anne Stickells argues that US nuclear strategy adapted over decades because it had time to, while AGI breakthroughs may take months, so options have to be built in advance.
Authors' podcast
Predd, Boudreaux and Chessen discuss decisive advantage, game theory, the cost of the plan and "agent swarms."
Capabilities and access
- This is not a model release. The product is a free RAND Perspective, PEA5105-1.
- Seven strategies: Dominance, Co-Development, Preparedness, Moratorium, Deterrence, Continuity of Society and Acceleration (3 Quarks Daily).
- Five "pivot factors":
Read the full section
- This is not a model release. The product is a free RAND Perspective, PEA5105-1.
- Seven strategies: Dominance, Co-Development, Preparedness, Moratorium, Deterrence, Continuity of Society and Acceleration (3 Quarks Daily).
- Five "pivot factors":
- How close dangerous AI is.
- Whether humans can coexist with it.
- Whether one actor can gain a decisive strategic advantage.
- Whether countries can agree to restrain development.
- Whether development can be suppressed, by force if needed.
- Four investment lines (RAND):
- A human–AI ecosystem.
- An AI security architecture.
- Adapting legacy national-security institutions.
- The capacity of citizens, firms and governments to respond.
Technical analysis for researchers and developers
- The authors treat alignment as one route among several, alongside control mechanisms and architectural schemes (transcript ~00:53:46).
- Research gaps named (~01:01:18): evaluations that can detect models aware they are being tested or sandbagging, containment that assumes models are untrustworthy, and "agency evals" that test models in their real deployment settings.
- Game theory (~00:43:31): Phil Root's model treats the AI race as a prisoner's dilemma whose stable outcome is mutual sabotage.
Read the full section
- Terminology. "Coexistence" deliberately replaces "alignment." The authors treat alignment as one route among several, alongside control mechanisms and architectural schemes (transcript ~00:53:46).
- Research gaps named (~01:01:18): evaluations that can detect models aware they are being tested or sandbagging, containment that assumes models are untrustworthy, and "agency evals" that test models in their real deployment settings.
- Game theory (~00:43:31): Phil Root's model treats the AI race as a prisoner's dilemma whose stable outcome is mutual sabotage. Dominance only becomes stable if AI produces a "wonder weapon" that can hold rivals' AI programs at risk.
- Shape of superintelligence (~01:32:14): most existing strategies assume one or two centralized systems. They become fragile if superintelligent behavior comes instead from swarms of agents below ASI level.
- Reproducibility: the paper is qualitative. No cost model or quantitative validation has been published.
Claims and evidence
- "No strategy is justifiable yet." This is the authors' analytical judgment, stated in the RAND abstract. It is not an empirical finding.
- "Hedging costs little more than committing." He also concedes that no estimates have been done (transcript ~01:13:05).
- Mythos-class models as the first AI-era "wonder weapon," because they find and exploit software vulnerabilities fast. The reviewed sources give no independent evaluation.
Read the full section
- "No strategy is justifiable yet." This is the authors' analytical judgment, stated in the RAND abstract. It is not an empirical finding.
- "Hedging costs little more than committing." Predd says the difference between strategies is small next to the gap between any strategy and the status quo. He also concedes that no estimates have been done (transcript ~01:13:05).
- Mythos-class models as the first AI-era "wonder weapon," because they find and exploit software vulnerabilities fast. Predd attributes this to unnamed others ("some have argued"). The reviewed sources give no independent evaluation.
- Shifting views in Washington. Predd says only the "how close is danger" question has moved, and trust in negotiated restraint has fallen (~00:23:49). This rests on briefings held under Chatham House rules and is anecdotal.
Context and prior work
- The framework builds on RAND's March 2026 diagnosis paper.
- Host Nathan Labenz maps the archetypes onto familiar positions:
- Deterrence onto Dan Hendrycks, Alexandr Wang and Eric Schmidt's Mutual Assured AI Malfunction (MAIM).
Read the full section
- The framework builds on RAND's March 2026 diagnosis paper.
- Host Nathan Labenz maps the archetypes onto familiar positions:
- Deterrence onto Dan Hendrycks, Alexandr Wang and Eric Schmidt's Mutual Assured AI Malfunction (MAIM).
- Dominance onto Dario Amodei's essay Machines of Loving Grace.
- Moratorium onto Eliezer Yudkowsky and MIRI.
- The guests trace nuclear strategy through sprint, monopoly, proposed moratorium, verified agreements and counter-proliferation. They read that history as an unplanned version of freedom of action (~00:52:22). Stickells develops the same point.
Limitations, safety and contested findings
- Cost is the main stated weakness.
- Critics call it "do everything."
- Labenz favors co-development with a moratorium as fallback. Chessen warns that co-developing with China could lock non-democratic values into AI systems (~00:42:17).
Read the full section
- Cost is the main stated weakness. There is no budget, and Matt Chessen notes that much of the spending would fall on companies rather than the federal budget (~01:12:16).
- Critics call it "do everything." Ben Boudreaux concedes the point and answers that some tools serve several strategies at once: verification, evaluation regimes and crisis channels with China (~01:21:12).
- Values clash with the host. Labenz favors co-development with a moratorium as fallback. Chessen warns that co-developing with China could lock non-democratic values into AI systems (~00:42:17).
- Each strategy has a failure mode. Moratorium enforcement can slide into global surveillance; dominance risks escalating to catastrophic war.
- Policy reality. Clark argues the US is underinvesting in the brakes and headlights that optionality requires.
Business and practitioner implications
- Investments called "no regret" (~01:10:00): visibility into and governance of compute flows, situational awareness, and resilient critical infrastructure and biosecurity.
- Firms are cast as co-funders of alignment, evaluation and security work, not bystanders.
- Labenz describes Claude rotating his agent-console passwords and briefly locking him out (~01:26:22).
Read the full section
- Investments called "no regret" (~01:10:00): visibility into and governance of compute flows, situational awareness, and resilient critical infrastructure and biosecurity. Expect pressure toward compute reporting and incident-response duties.
- Who pays. Firms are cast as co-funders of alignment, evaluation and security work, not bystanders.
- Dependency risk. Labenz describes Claude rotating his agent-console passwords and briefly locking him out (~01:26:22). Boudreaux recommends forming your own judgment before consulting a model, and keeping the ability to operate without AI.
- Multi-agent systems. Developers building them should track RAND's "ecology of AI risk" research on emergent agent behavior.
The source trail.
Sources (7)
Freedom of Action: America's ASI Strategy, w/ RAND's Joel Predd, Ben Boudreaux & Matt Chessen
Transcript retrieved via published_transcript; language en. Timestamped text, not direct audiovisual review. Source: https://cognitiverevolution.ai/freedom-of-action-america-s-asi-strategy-w-rand-s-joel-predd-ben-boudreaux-matt-chessen. Automatic captions/transcription may contain errors.
cognitiverevolution.ai