Long-Term Future Fund (LTFF)

Weakest evidence: Impact of funded projects (no rigorous evaluation exists), and whether the CEA merger helps or hurts.

Stated Theory of Change

LTFF claims to reduce existential risk (primarily from AI) by funding the "long tail" of the AI safety talent pipeline: early-career researchers, career changers, small projects, and unconventional bets that larger funders (OP at $257K median, SFF at $100K median) cannot efficiently evaluate at $25K median grant size.

The specific mechanisms are:

1. Talent pipeline: Fund people to upskill, transition careers, or test their fit for AI safety work. Many LTFF grantees are meant to "graduate" to larger grants from OP/SFF or to positions at established organizations.
2. Epistemic diversity: By staffing the fund with people who do direct work (not professional grantmakers), and by deliberately distancing from OP, LTFF claims to fund projects OP would miss or reject.
3. Speed: As the largest always-open funder (pre-Q1 2025), LTFF caught time-sensitive opportunities other funders missed.
4. Hits-based giving: The "one excited fund manager" model explicitly accepts high variance in exchange for catching outlier opportunities.

Revealed Theory of Change

The actions tell a somewhat different story:

What they actually optimize for: Getting money out the door efficiently. The entire organizational design -- part-time managers, minimal evaluation time per grant, hits-based approval -- is built for throughput. Linch's grantmaker lessons make this explicit: detailed evaluation rarely changes outcomes, so the fund optimizes for volume of good-enough decisions rather than precision of any single decision.

Where the stated and revealed theories diverge:

1. Epistemic diversity vs. network effects: The fund claims to provide diversity from OP, but fund managers are appointed via private process by OP-connected committees, most managers work at OP-funded organizations, and the chair was simultaneously employed at OP until 2023. The "independence" is more structural than epistemic.

2. Open application vs. network-driven: The open application form is real, but the centralized grantmaking critique applies: fund managers use network-based heuristics ("someone in my network can vouch for this person"), and Bay Area / rationalist community members have natural advantages.

3. Impact measurement vs. faith: The fund has disbursed $20M with essentially no systematic impact evaluation. The 2021 retrospective covered 23 grants from 2018-2019 and found 26% had "very little information." Fund managers acknowledge they "cannot provide strong assurances" their work hasn't been net-negative. This is honest but also means the theory of change is largely untested.

4. Institutional capacity vs. volunteer dependence: LTFF's two full-time employees managing eight-figure turnover is remarkably efficient but creates fragility. The persistent inability to hire a permanent full-time chair (2+ years) suggests the institutional model may not be sustainable.

Key Assumptions

1. Early-stage AI safety talent is funding-constrained (not talent-constrained)

• Evidence for: Multiple applicants report difficulty funding career transitions; LTFF's narrowly-rejected-grants list shows promising projects going unfunded
• Evidence against: AI safety positions at labs now attract many strong applicants; the field has grown substantially since 2017
• If wrong: LTFF's primary value proposition collapses. If good talent gets funded regardless, LTFF is funding the marginal talent that wouldn't have succeeded anyway.

2. Hits-based giving is appropriate for this domain

• Evidence for: The top grants appear much better than marginal grants; some LTFF-funded people (SERI MATS, Robert Miles) have had visible impact
• Evidence against: A 30-40% success rate with 26% unknown outcomes means the "hits" are hard to identify even retrospectively
• If wrong: The high-variance approach wastes money on duds without enough hits to compensate

3. Part-time fund managers with direct work experience make better decisions than full-time professional grantmakers

• Evidence for: "Contact with reality" is a real advantage; full-time grantmakers can become disconnected from the field
• Evidence against: Part-time managers are unreliable (Bergal's reflections), spend minimal time per grant, and have high turnover. The "direct work" advantage may be swamped by the "not enough time" disadvantage.
• If wrong: LTFF would need to professionalize its grantmaking team, which would change its character fundamentally

4. Independence from OP is achievable and valuable

• Evidence for: The Aug 2023 reforms were substantive; OP no longer funds grantmaking; chairs no longer dual-hold OP positions
• Evidence against: Personnel networks persist; fund managers work at OP-funded orgs; OP still funds operations; the CEA merger creates new entanglement
• If wrong: LTFF is effectively an OP regrantor with extra steps, providing an illusion of diversity rather than genuine alternatives

Strengths

1. Unmatched transparency: LTFF publishes detailed payout reports, marginal grant analyses, narrowly-rejected-grant descriptions, COI disclosures, scoring thresholds, and self-critical reflections. No comparable funder in AI safety comes close.

2. Genuine niche: The $25K median grant for early-stage individuals fills a real gap. OP and SFF don't efficiently make grants this small. Someone needs to fund the person who needs $30K for a 6-month career transition.

3. Institutional self-doubt: Fund managers publicly question whether their work is net-positive. This epistemic humility is rare for funders and probably makes them better grantmakers.

4. Track record of reform: When the Beckstead era failed, LTFF reformed. When OP entanglement became untenable, LTFF distanced. When rolling applications overwhelmed the team, LTFF returned to rounds. The fund adapts, even if slowly.

5. Cultural willingness to fund contrarians: Habryka funding Turner despite "pretty strong" disagreement demonstrates real intellectual openness. The fund culture appears to genuinely prioritize researcher quality over ideological alignment.

Weaknesses and Risks

1. No impact evaluation after $20M disbursed: This is the most serious weakness. The fund's theory of change -- that small grants to early-career people reduce existential risk -- is essentially unfalsified. The 2021 retrospective was a start but covered a fraction of grants at minimal depth.

2. Persistent leadership vacancy: The permanent full-time chair position has been open for 2+ years. This is the single most important role in the organization. Interim leadership during major institutional transitions (CEA merger) is a serious governance risk.

3. Extreme operational fragility: Two full-time employees managing $5-12M annually. If Caleb or Linch left, there is no obvious successor and no documented process to ensure continuity.

4. CEA merger conflicts: Post-merger, the entity running EA Global conferences, community building grants, and the EA Forum will also control the fund that evaluates competing programs. This is a structural conflict of interest that disclosure alone cannot resolve.

5. Policy blind spot: The documented pattern of underfunding policy work while overfunding technical AI safety research reflects the fund managers' backgrounds, not an evidence-based assessment of comparative impact.

6. Declining transparency: Optional public reports since 2021, growing fraction of anonymous grants, no payout report for 11+ months. The fund's distinctive transparency advantage is eroding.

7. Donor base ceiling: The OP matching mobilized ~$1M from individual donors. Without OP matching, the fund may struggle to sustain $8-10M annual disbursements. Financial sustainability post-OP is unproven.

Cross-References

• Open Philanthropy: LTFF's primary historical funder and the entity it's trying to become independent from. OP still funds operations. The relationship is evolving toward genuine independence but is incomplete.
• Survival and Flourishing Fund (SFF): Second-largest AI safety funder. Operates at larger grant sizes ($100K median) and uses the S-Process for decision-making. Complementary rather than competitive with LTFF.
• Manifund: Decentralized regranting platform. Former LTFF fund manager Adam Gleave became a Manifund regrantor. Manifund's model (individual regrantors with discretion) directly challenges LTFF's institutional model.
• Lightspeed Grants: Designed to address LTFF's speed issues with even faster turnaround. Funded by Jaan Tallinn (also funds SFF). Partially overlaps with LTFF's niche.
• SERI MATS: Major LTFF grantee ($316K). The MATS talent pipeline is one of LTFF's most visible successes -- but MATS has grown into an established program that can raise funds independently.
• CEA: Merging with EA Funds. Will become LTFF's institutional home. The relationship is symbiotic but creates governance concerns.
• FTX Future Fund (defunct): Briefly operated in LTFF's space with 10x the funding. Its collapse validated the need for funding diversity but also showed how fragile the ecosystem is.

What Would Change This Assessment

Positive updates:

• Publication of a rigorous, independent evaluation of 50+ LTFF grants tracking outcomes 3+ years post-funding
• Appointment of a permanent full-time chair with grantmaking experience and operational capability
• Post-merger governance structures that demonstrably prevent CEA-LTFF conflicts of interest
• Evidence that LTFF-funded career transitions led to sustained AI safety careers at measurably higher rates than baseline

Negative updates:

• Discovery that a significant fraction of LTFF funding went to projects or people with undisclosed conflicts of interest
• Evidence that the CEA merger caused LTFF to systematically favor CEA-aligned projects
• Continued absence of payout reports for 18+ months (suggesting institutional decay)
• Departure of both Caleb Parikh and Linch Zhang without clear succession
• Evidence that Manifund/Lightspeed/individual regrantors achieve comparable outcomes at lower overhead

Self-Critique

What's weakest in this analysis:

1. I'm relying heavily on LTFF's own self-reported information. The fund's unusual transparency makes this better than average, but there is no external check on the narrative.
2. I may be too credulous of the "epistemic diversity" argument. The Aug 2023 reforms were real, but I'm uncertain whether they changed actual funding decisions or just the appearance of independence.
3. The CEA merger analysis is necessarily speculative since it was recent at the time of evidence collection and outcomes weren't yet visible.

What would a thoughtful person who disagrees say: "LTFF has disbursed $20M to hundreds of people in AI safety. Even with minimal formal evaluation, the field has grown enormously, and LTFF grantees like MATS alumni are now staffing every major alignment lab. The demand for rigorous evaluation is itself a form of inappropriate formalism for a hits-based fund in a pre-paradigmatic field. LTFF's real contribution is maintaining a low-friction pipeline that gets money to people quickly, and the 'weaknesses' you identify -- minimal evaluation time, high turnover, part-time managers -- are features of a lean, adaptive system, not bugs."

That's a reasonable counterargument. My response: it may be true that LTFF's informality is a feature, but without any measurement, we genuinely don't know. A fund that cannot distinguish between "we funded 100 people and 30 became productive AI safety researchers" and "we funded 100 people and 5 became productive" is flying blind -- and asking donors to fly blind with it.

What information would most change my view: A well-designed retrospective evaluation covering 100+ LTFF grants that tracked career outcomes, publications, and field retention 3+ years post-funding. If the success rate were substantially above baseline career transition rates (say, 40%+ of funded individuals working productively in AI safety 3 years later), I would update strongly positive on LTFF's theory of change. If it were below 15%, I would update strongly negative.