Institute for AI Policy and Strategy (IAPS)

Stated Theory of Change

IAPS believes transformative AI is coming within years to a decade, that the US government is catastrophically underprepared, and that technically grounded policy research can steer outcomes toward safety. The specific mechanism: produce research at the intersection of AI and national security that translates into legislation, executive action, and institutional capacity.

The core insight is that compute is the most governable input to AI development. AI chips are physically scarce, concentrated in a few supply chains, and potentially trackable -- making them the most practical chokepoint for preventing dangerous proliferation while allowing beneficial development. By making export controls technically enforceable (through location verification, whistleblower programs, and supply chain monitoring), IAPS aims to give the US government real tools rather than unenforceable rules.

The broader theory encompasses three pillars: (1) frontier security -- ensuring AI systems are safe from cyberattack, misuse, and loss of control; (2) compute policy -- governing the hardware layer to control who can build powerful AI; (3) international strategy -- navigating US-China competition to avoid both an uncontrolled arms race and catastrophic conflict.

Underlying the public "nonpartisan national security" framing is a worldview rooted in EA/longtermist thinking. Wildeford has stated the "default outcome of AGI is doom" and estimates transformative AI within years. IAPS's Strategic Visions paper explicitly treats "takeover by powerful misaligned AI systems" as a first-class risk. The national security framing is not dishonest -- it is a translation of x-risk concerns into language that opens doors in Washington.

Revealed Theory of Change

IAPS's actions closely align with its stated theory, with some revealing emphases:

Compute governance dominates. The research page shows 20 compute governance papers vs 18 policy/standards and 16 international governance. The flagship research (location verification) and highest-profile policy achievement (Chip Security Act) are both compute governance. This isn't the balance of the three-pillar framing -- it's one-pillar dominance, which makes sense because compute governance is where IAPS has the strongest competitive advantage and the most tractable policy levers.

Playing the inside game. IAPS's engagement with the Trump administration (OSTP RFI response, AI Action Plan analysis, constructive framing) reveals a pragmatic, non-ideological approach. They frame AI governance as enabling national competitiveness, not restricting it. This is smart DC strategy: work with whoever is in power.

Field-building through the fellowship. The 29-fellow program (0.5% acceptance) with placements across government, think tanks, and academia is a long-term influence strategy. IAPS isn't just producing research -- it's training the people who will be in senior policy positions in 10-20 years.

Comfortable with extreme scenarios. Publishing on ASI deterrence and MAIM dynamics is unusual for a "nonpartisan think tank." This reveals IAPS takes catastrophic AI risk seriously at the institutional level, not just through Wildeford's personal views.

The EA-to-DC translation. The zero EA Forum presence, the "nonpartisan" branding, the hiring of former NSC/State/BIS professionals -- all point to a deliberate strategy of translating EA-originated AI safety concerns into mainstream national security language. The revealed theory of change includes "escape the EA brand" as an operational requirement.

Key Assumptions

1. Compute governance is the highest-leverage AI governance intervention.

• Evidence for: Chips are physically scarce and concentrated; export controls are the most powerful existing policy tool; location verification is technically feasible.
• Evidence against: Software innovations (distillation, algorithmic efficiency) can circumvent hardware constraints; China is developing domestic chip manufacturing; open-source models reduce the importance of frontier compute.
• Testable: If Chinese AI capabilities advance rapidly despite export controls, this assumption weakens.
• If wrong: IAPS's core research agenda becomes less policy-relevant.

2. Research can translate into legislation and executive action quickly enough to matter.

• Evidence for: Location verification research informed the Chip Security Act within ~18 months. OSTP RFI responses were incorporated into the AI Action Plan.
• Evidence against: Legislation can die in committee; the Trump administration is relaxing export controls; Nvidia's lobbying may block the CSA.
• Testable: Track whether the CSA passes and whether location verification is actually implemented.
• If wrong: IAPS becomes an academic institution rather than a policy-shaping one.

3. The US government will continue to care about restricting China's AI chip access.

• Evidence for: Bipartisan support for the CSA; intelligence community concern about Chinese AI military applications.
• Evidence against: Trump administration is relaxing some controls; commercial interests (Nvidia, etc.) push for fewer restrictions; Carnegie Endowment's Winter-Levy flags this risk explicitly.
• Testable: Watch export control policy over the next 1-2 years.
• If wrong: IAPS's entire compute policy pillar loses its policy audience.

4. National security framing is the best vehicle for AI safety concerns in the current political environment.

• Evidence for: Bipartisan appeal; doors open in both Democratic and Republican administrations; avoids the "AI doomer" label.
• Evidence against: National security framing can be co-opted for purposes unrelated to safety (surveillance, hegemony); risks alienating the technical AI safety community; may not survive a political environment that deprioritizes national security competition.
• Testable: Whether IAPS maintains influence across political transitions.
• If wrong: IAPS may find its framing captures attention but not the right kind of policy action.

5. Hardware controls do not create more problems than they solve.

• Evidence for: Location verification is lightweight, privacy-preserving, and uses existing chip capabilities.
• Evidence against: Cybersecurity center's Clipper Chip analogy; STAIR's surveillance/function creep concerns; Nvidia's customer trust argument; risk of pushing countries toward Chinese alternatives.
• Testable: If location verification is implemented, observe whether cybersecurity incidents result or whether foreign customers defect.
• If wrong: IAPS's flagship policy achievement actively harms security.

Strengths

Concrete legislative impact in record time. Most think tanks operate for years before their research shapes legislation. IAPS went from founding (Sept 2023) to having research incorporated into bipartisan bills (May 2025) in less than two years. The Chip Security Act and Stop Stealing Our Chips Act are real policy outputs, not just papers.

Strong technical talent on compute governance. The team of Grunewald, Aarne, King, Roberts, and Brass combines technical expertise with former BIS/government experience. This makes their research credible to both technologists and policymakers.

Effective DC positioning. The Jenny Marron hire (NSC, State), the "nonpartisan" branding, the constructive engagement with the Trump administration, and the CNAS collaboration all demonstrate institutional sophistication unusual for an EA-incubated org. IAPS knows how to play the Washington game.

Rapid response capability. The distillation attacks memo (weeks after industry reports), the AI Action Plan rapid analysis (9 researchers), and the team's ability to produce timely policy-relevant work distinguish IAPS from slower academic think tanks.

Fellowship as force multiplier. 29 fellows trained per year, placed across government and think tanks, creates a growing network of IAPS-trained policy professionals. This is a long-term investment in institutional influence.

Weaknesses and Risks

Extreme funder concentration. Open Philanthropy appears to provide 70-90%+ of IAPS's budget. This creates existential dependency on a single funder and undermines the editorial independence claim. If OP pivots its AI policy funding, IAPS could collapse. The lack of earned revenue or diversified funding is a structural vulnerability.

Governance gap. No independent board, no separate legal entity (until 2026), no public financial reporting. For a $10M+/year organization influencing legislation, this is below the standard of comparable think tanks. The 2026 spin-off could fix this, but the transition is opaque.

Cybersecurity critique is substantive. The Clipper Chip analogy from the Center for Cybersecurity Policy is not dismissible. If location verification mechanisms create vulnerabilities in US military and critical infrastructure chips, the consequences would be severe. IAPS has not publicly engaged with this critique in depth.

Global South critique is unaddressed. Vipra's STAIR argument that hardware controls entrench US hegemony, erode trust in international relations, and promote technological delinking deserves engagement. IAPS's "nonpartisan" claim rings hollow if it only considers US interests in its compute governance recommendations.

Political fragility. IAPS's compute policy pillar depends on the US government caring about export controls. The Trump administration's relaxation of some controls and Nvidia's lobbying power create real political risk. If the CSA dies, IAPS's flagship achievement evaporates.

EA stigma management is a tightrope. The complete absence from EA community spaces is strategically rational but creates a disconnect. IAPS's funding, intellectual origins, and many staff come from the EA ecosystem, but it presents as a mainstream think tank. This works as long as journalists don't dig too deeply.

Cross-References

Complementary to: GovAI (academic compute governance research, theory-focused -- IAPS is the applied/legislative arm), CNAS (established national security think tank -- IAPS collaborates directly), Epoch AI (quantitative AI development tracking -- also formerly RP-sponsored).

Different approach from: MIRI (IAPS works within mainstream policy; MIRI advocates for extreme measures), CAIS (IAPS avoids the "pause" framing), ARC (IAPS is policy, not technical alignment research), PauseAI (IAPS is the opposite of confrontational advocacy).

Potential competition with: CSET (Georgetown -- similar compute governance focus), AIPI (similar DC positioning but more EA-branded), ARI (Americans for Responsible Innovation -- similar space).

Funding ecosystem: Part of the Open Philanthropy network alongside GovAI, CAIS, FLI, 80K Hours, and others. OP also funds fellows in Congress and CNAS, creating an interconnected influence infrastructure.

What Would Change This Assessment

• If the Chip Security Act passes and location verification is implemented successfully: Would strongly validate IAPS's theory of change and compute governance approach.
• If location verification creates a cybersecurity incident: Would seriously damage IAPS's credibility and validate the Clipper Chip critique.
• If the Trump administration abandons export controls entirely: Would undermine the policy relevance of IAPS's core research agenda.
• If IAPS publishes research that clearly contradicts an OP position: Would validate the editorial independence claim. Absence of such evidence to date is notable.
• If the 2026 board is dominated by EA insiders or OP-connected individuals: Would confirm the governance concern. An independent, diverse board would alleviate it.
• If China achieves frontier AI capability despite export controls: Would suggest IAPS's core bet on compute governance is wrong.
• If IAPS diversifies funding significantly away from OP: Would substantially reduce the funder concentration risk.

Self-Critique

What I'm least confident about:

• The non-OP funding picture. "Allied government" funding, SFF amounts, and individual donors could significantly change the funder concentration story if they're larger than I estimate.
• Whether IAPS's editorial independence from OP is genuine or cosmetic. I have no evidence either way -- the claim is untested.
• Whether the 2026 spin-off will improve or weaken governance. The transition is opaque.

Where this analysis may be biased:

• I may be over-weighting the EA origin story. IAPS has hired substantial non-EA talent (Marron, King, Roberts, Ee, Brass) and may genuinely operate as a mainstream think tank despite its origins.
• I may be under-weighting IAPS's actual policy impact. The Chip Security Act and AI Action Plan incorporation are real achievements that most think tanks never accomplish.
• The cybersecurity critique may be technically overstated -- I'm not a chip security expert and am relying on the Center for Cybersecurity Policy's framing.

What a thoughtful disagreer would say: "IAPS is one of the most effective AI safety organizations in the world. In two years they went from zero to having their research shape bipartisan legislation. The EA origins are irrelevant -- what matters is the output. The national security framing isn't deception; it's correct that AI risk IS a national security issue. The funder concentration concern is valid but applies to half the think tanks in Washington. And the cybersecurity critique of location verification is an implementation concern, not a fundamental flaw."

What information would most change my view: Detailed financial data showing funding diversification beyond OP, or evidence that IAPS's board (post-spin-off) has genuine independence from the EA funding network.