Berkeley Existential Risk Initiative (BERI)

Stated Theory of Change

BERI removes bureaucratic friction from university x-risk research. The causal chain:

1. University researchers work on existential risk reduction
2. University administrative systems create obstacles (slow hiring, inflexible purchasing, rigid salary scales)
3. BERI fills these gaps by hiring staff, buying equipment, and handling admin outside university channels
4. Researchers spend more time on research and less on bureaucracy
5. Research output increases, contributing to x-risk reduction

BERI explicitly disclaims deeper strategic ambitions: "BERI is not a donor or a grantmaker." They provide infrastructure, not direction. The selection of which research groups to support is the closest thing to a strategic bet, but the criteria are opaque ("collective judgement of BERI's Board of Directors, with influences from advisors and trusted sources of funding").

The intellectual foundation comes from founder Andrew Critch, whose ARCHES paper argues that existential risk arises primarily from multi-stakeholder coordination failures -- externalities between AI systems and organizations that nobody is specifically responsible for preventing. Supporting independent university research (as opposed to industry safety labs) is consistent with this: university groups are less captured by commercial incentives and can study the coordination problems that labs have no incentive to solve.

Revealed Theory of Change

BERI's actions largely match their stated theory, with some significant additions:

Consistent with stated theory: The collaboration program operates as described. Survey data suggests real value. The overhead model is well-designed. BERI grows modestly and avoids scope creep (they actually narrowed from grants to collaborations in 2020).

Beyond stated theory: BERI has repeatedly incubated things that grow independent -- MATS ($6.3M channeled, now a standalone 501(c)(3)), Sculpting Evolution (became SecureBio), and arguably SFF (seeded from BERI's grants program). This "launch and release" pattern is not in their official description but may be their highest-impact activity. Providing a legal and administrative home for nascent x-risk projects until they can stand alone is a distinct and valuable function.

Divergence from stated theory: The founder's trajectory tells a different story than the org's. Critch spent 2017-2021 deeply involved with BERI, then shifted full-time to a for-profit AI company (Encultured AI/HealthcareAgents, co-founded with Jaan Tallinn). He now spends 1hr/month at BERI. This could indicate that Critch decided the infrastructure model was working well enough to leave in others' hands, or that he found the model insufficient and moved to what he considers higher-impact work ("products > narratives"). His recent emphasis on building products and his critique that "most public AI risk discourse is incorrect" suggest some disillusionment with the research support model.

Key Assumptions

1. University research groups are doing work that reduces x-risk.

• Evidence for: CHAI (Stuart Russell), CSER, David Krueger's lab, InterACT, Barrett's standards work -- these are recognized safety-relevant groups
• Evidence against: BERI does not assess research quality or impact; some collaborations (GreenEarth Social, Social Science Prediction Platform) have unclear x-risk connections; trial collaborations are evaluated on "compatibility" not impact
• Testable? Partially -- could assess publication output, citations, and downstream impact of supported researchers
• If wrong: BERI is efficiently supporting work that does not reduce x-risk, which means the overhead (however lean) is wasted

2. Operational friction is a meaningful bottleneck for university x-risk research.

• Evidence for: 44% of projects would "not have happened" without BERI; collaborator survey praises speed and flexibility; university bureaucracies are genuinely slow for non-standard requests
• Evidence against: 60% of projects would have happened anyway, just more slowly; the actual dollar amounts per trial collaboration ($1-10K) are very small; major research projects are constrained by ideas and talent, not by whether someone can buy snacks or Uber rides quickly
• Testable? Yes -- compare research output before and after BERI engagement
• If wrong: BERI provides convenience, not impact. Researchers would adapt to working within university systems and produce similar output.

3. The 10-15% overhead is justified by the value BERI adds.

• Evidence for: Overhead rate competitive with fiscal sponsors; BERI has reduced it from 40% to ~10-15%; per-project overhead creates good incentive alignment
• Evidence against: No comparison to direct funding alternatives; funders could theoretically give directly to universities with an admin budget
• Testable? Yes -- funders could experiment with direct vs. BERI-mediated grants
• If wrong: BERI is a well-intentioned but unnecessary middleman

4. BERI can survive post-MATS and maintain relevance.

• Evidence for: 2024-2025 CG grants continue (AI governance, InterACT, standards); new collaborations being added; team growing from 1 to 3 FTE core
• Evidence against: Revenue dropped 20% in 2024; 84% donor concentration unresolved; no evidence the new-donor goal was met
• Testable? Will be clear within 1-2 years
• If wrong: BERI contracts to a small org managing a handful of CHAI and CLTC collaborations

Strengths

1. Lean and efficient: 3 FTE managing 37+ collaborations with 10-15% overhead. Very low administrative bloat.
2. Demonstrated incubation value: MATS and SecureBio both grew through BERI's fiscal sponsorship into independent orgs -- a concrete track record of value creation.
3. High collaborator satisfaction: 4.74/5.0, with strong qualitative feedback about speed, responsiveness, and lack of bureaucracy.
4. Ethical integrity: Returned all FTX money ($389K) in full; Critch takes $0; strong transparency practices for org size (audited financials, published bylaws, public donor list).
5. Sophisticated financial model: Per-project overhead reasoning is economically sound and creates better incentives than centralized overhead funding.
6. Global university network: Connections across Berkeley, MIT, Stanford, Oxford, Cambridge, Princeton, ETH Zurich, and more -- a unique position in the x-risk ecosystem.
7. CLTC AI Standards work: Barrett's NIST AI RMF profile for GPAIS is a tangible policy output with real-world adoption and industry endorsement.

Weaknesses and Risks

1. Critical funding concentration: 84% from two sources (Open Phil + Tallinn). This is an existential vulnerability for BERI itself. If either funder deprioritizes them, the org could collapse within a year.
2. Founder disengagement: Critch at 1hr/month means the intellectual engine behind BERI is gone. The current leadership (Cooper) has operations expertise but no visible AI safety vision or public engagement.
3. Governance weakness: ED and Board President are the same person (Cooper), with only 2 other board members. Limited independent oversight for an org managing millions.
4. No strategic quality control: BERI does not evaluate whether the research it supports is high-quality, impactful, or actually reducing x-risk. They provide services to groups selected by opaque criteria.
5. Post-MATS revenue cliff: Losing ~48% of CG grant volume with MATS' independence. 2024 revenue already declined 20%. Sustainability of current scale is uncertain.
6. Network concentration of influence: The Critch-Tallinn-SFF-BERI web, while managed with recusals, represents structural concentration in the x-risk funding ecosystem that limits independent oversight.
7. Scalability concerns acknowledged: BERI's own 2024 goals admit "Many of our current processes are simply not scalable... With 37 collaborations and likely more to come, it's critical that we fix this now." No evidence the fix happened.

Cross-References

• Effective Ventures / CEA: Similar fiscal sponsorship model but much larger and more centralized. BERI is smaller, leaner, and more focused on university contexts. EV's governance failures provide a cautionary tale about concentrated organizational structures.
• Rethink Priorities Special Projects Program: BERI explicitly refers researchers to RP-SPP for traditional fiscal sponsorship. Different niche -- RP-SPP does general fiscal sponsorship, BERI is university-specific.
• CHAI (Stuart Russell): BERI's first and most enduring collaboration. CHAI is BERI's largest single program by CG grant volume.
• MATS: BERI's most successful incubation, now independent. The relationship shows BERI's model at its best.
• SFF (Survival and Flourishing Fund): Seeded from BERI, now a major funder (~$15M in 2024 Initiative Committee grants alone). Critch sits on both BERI and SFF. The structural interconnection is the most notable cross-reference.
• Coefficient Giving/Open Phil: Primary funder. CG's continued small-to-medium grants suggest ongoing positive assessment but declining investment relative to the MATS era.

What Would Change This Assessment

Upward updates:

• BERI publishes a formal theory of change with testable predictions about how their work reduces x-risk
• Successfully diversifies funding beyond 50% from top two donors
• Elizabeth Cooper publishes substantive public writing on AI safety strategy
• Independent evaluation confirms high counterfactual impact of BERI-supported research
• The "new program" beyond collaborations launches and demonstrates impact

Downward updates:

• 2025-2026 revenue continues declining, confirming post-MATS contraction
• Collaborator survey satisfaction drops below 4.0/5.0
• Evidence emerges that BERI-supported research groups produce less impactful work than non-BERI peers
• The Critch-Tallinn network faces scrutiny revealing undisclosed conflicts
• BERI fails to add the independent board member needed to separate ED and Board President roles

Self-Critique

Weakest claim: My assessment of BERI's "revealed theory of change" being partly about incubation is based on just three examples (MATS, SecureBio, SFF). This could be pattern-matching rather than a real strategic function.

Sources I should have checked but could not:

• EA Forum discussion threads about BERI (blocked domain)
• LessWrong posts, including Critch's "Taking AI Risk Seriously" (blocked domain)
• BERI annual reports (hosted on Google Docs, not fetchable)
• Full Charity Navigator assessment

Potential bias: I may be too sympathetic to BERI's "humble infrastructure" framing. A harder-nosed assessment might conclude that an org with no published theory of change, no impact evaluation, and opaque selection criteria is simply under-managed, not admirably lean. The 44% counterfactual claim from a self-administered survey is not strong evidence by rigorous standards.

What a thoughtful disagreer would say: "BERI is a perfectly fine small nonprofit doing perfectly fine small work. The problem is opportunity cost. The millions of dollars flowing through BERI could be directed more effectively through direct grants to these same research groups. The 10-15% overhead, even if competitive, is a tax on x-risk funding that produces no research. The incubation examples are happy accidents, not a model. And the founder leaving to build a for-profit tells you everything about whether he thinks this model actually works."

Information that would most change my view: An independent evaluation of whether BERI-supported research groups produce more impactful safety research than comparable non-BERI groups. If the answer is no, then BERI provides convenience but not impact. If the answer is yes, then the operational leverage theory of change is validated.