FTX Future Fund

Stated Theory of Change

FF's stated theory of change was pure longtermist expected value maximization: the far future contains astronomically more potential value than the present, so actions that shape humanity's long-run trajectory -- especially reducing existential risk from AI and pandemics -- dominate any utilitarian calculation. FF was the vehicle to deploy unprecedented capital at unprecedented speed into this space.

The specific mechanism was threefold:

1. Staff-led grants for large, established longtermist orgs (Longview, CEA, Lightcone) -- strengthening infrastructure
2. Open call for specific projects (AI safety research, biosecurity, forecasting) -- discovering new talent and ideas
3. Regranting via 100+ trusted individuals -- testing whether philanthropic capital deployment could be parallelized through a network of decision-makers

The regranting model was the genuinely novel element. Traditional funders have centralized decision-making; FF was testing whether longtermist grantmaking could scale through distributed trust networks. In Beckstead's 2017 formulation: the bottleneck in longtermist philanthropy isn't money, it's "finding the right people and projects." Regranting was an attempt to parallelize that search.

Revealed Theory of Change

FF's actions reveal a theory of change that was less about rigorous impact evaluation and more about flooding the zone with money. The operational reality:

• Speed over rigor: $160M committed in ~5 months, with a team of 5. That's $32M per person, or roughly $6.4M per person per month. For comparison, Open Phil has ~100+ staff deploying ~$650M/year. FF was operating at 10x the per-person grantmaking rate with a fraction of the infrastructure.

• EA ecosystem funding, not just longtermist research: The largest single grants went to CEA ($13.94M) and Longview ($15M) -- organizations that are EA infrastructure, not direct existential risk reduction. This suggests FF's revealed priority was growing the EA movement itself, with the implicit theory that a larger EA movement would produce more existential risk reduction.

• Network-based trust, not independent evaluation: The regranting model relied entirely on the judgment of individuals within the EA network. The regrantors were not publicly identified, their decisions were not reviewed, and there was no formal accountability mechanism. This is a trust-based system, and it worked only as long as the trust was warranted.

• Political ambitions alongside philanthropy: The simultaneous SBF-funded Carrick Flynn campaign, Guarding Against Pandemics, and Building a Stronger Future reveal that FF was one piece of a broader EA political project. The goal was not just research funding but building political power for longtermist priorities.

The divergence from stated theory: FF claimed to be making the world's most important grants through careful consequentialist reasoning. In practice, it was making fast grants through an insular network, with substantial funds going to organizations run by FF's own advisers, and no public accountability for decisions.

Key Assumptions

Assumption 1: Longtermism correctly identifies the highest-priority interventions.

• For: Standard expected value arguments; the potential scale of the far future genuinely dwarfs the present.
• Against: The American Prospect critique that multiplying small probabilities by astronomical future populations can justify anything; the observation that actual longtermist grantmaking looks remarkably similar to "fund our friends' organizations."
• Testable: Somewhat -- the longtermist portfolio's impact can be retrospectively evaluated as AI development unfolds.
• If wrong: FF's entire grantmaking portfolio was systematically misdirected.

Assumption 2: Speed of capital deployment matters more than rigor of evaluation.

• For: In a neglected field with genuine funding constraints, fast money fills gaps. The regranting model enabled hundreds of small grants that no traditional funder would make.
• Against: Tyler Cowen's observation that the team failed to identify their own existential risk. The regrantor system produced some clearly low-value grants alongside genuinely important ones.
• If wrong: FF's speed may have wasted money and created false dependencies in the ecosystem.

Assumption 3: The funding source is separable from the funded work.

• For: Pragmatic consequentialism -- a dollar's impact doesn't depend on its origin. FF team members appear genuinely not to have known the money was stolen.
• Against: EA specifically ignored warnings about SBF because the money was too valuable. The "earn to give" framework structurally discouraged scrutiny of earning methods. Clawbacks imposed additional harm on grantees.
• If wrong: The entire FF project was, from the start, morally compromised by its funding source, and the team's failure to investigate was a foreseeable consequence of EA's framework.

Assumption 4: Concentrated donor funding is an acceptable risk.

• For: In practice, most philanthropic fields depend on a few major donors.
• Against: FF's collapse demonstrated that the AI safety field's funding base was fragile. A single fraud could and did disrupt hundreds of projects.
• If wrong: The AI safety ecosystem needs to prioritize funding diversification over total dollar amounts.

Strengths

1. The regranting model was genuinely innovative. FF tested whether philanthropic grantmaking could be parallelized through trusted intermediaries. This model survived FF's collapse in modified form (Open Phil's Regranting Challenge, various community-based grantmaking). The small grants ($30K-$200K) to individual researchers were a niche no other funder was filling well.

2. Speed enabled real projects. AI Safety Camp, individual researchers, career transitioners, and small research orgs received funding within weeks of applying. For a field moving as fast as AI, this responsiveness had genuine value.

3. Scale signaled legitimacy. FF's ambition ($1B/year target) signaled to the broader world that AI safety was a serious field worth serious investment. This demonstration effect may have influenced subsequent funding decisions by other philanthropists.

4. Some grants had outsized impact. The $200K Anysphere investment became a $10B+ company. The $290K to AI Safety Camp helped build the field's talent pipeline. The $5M to Ought/Elicit funded a product still in use. Individual regrants supported researchers who went on to do important work.

Weaknesses and Risks

1. Governance was a sham. A 501(c)(3) with a board of three convicted felons (Ellison, Wang, Singh), no independent directors, no conflict-of-interest policies, no external audit, and no public financial disclosure. This is not "room for improvement" -- it is the absence of governance.

2. The 2018 warnings were ignored. Beckstead and MacAskill received specific, detailed warnings about SBF's dishonesty from people with direct knowledge. They chose to dismiss those warnings and, three years later, to lead SBF's philanthropy. This is the single most damaging fact for their credibility.

3. EA's framework was the root cause, not just SBF's character. The "earn to give" doctrine, the utilitarian tolerance for high-risk bets, the conflation of intellectual brilliance with moral trustworthiness, the insular social network, and the movement's financial dependence on a few wealthy individuals all created the conditions for this outcome. MacAskill's reflection that EA was "too trusting" understates the structural problem.

4. Conflicts of interest were pervasive and undisclosed. MacAskill (FF adviser) directing funds to organizations he founded; the EA Forum post assessing FF's grants published on a platform funded by FF; the regrantor network consisting of the funder's social circle. The entire system ran on personal relationships with no formal safeguards.

5. The collapse caused lasting damage. Beyond the direct financial losses (~$25M+ in unfulfilled grants, ~$70M+ in clawback proceedings), the reputational damage to AI safety funding persists. New donors face a higher bar for trust. Researchers who depended on FF learned that the field's funding was less stable than it appeared.

Cross-References

• Open Philanthropy / Coefficient Giving: FF was essentially an accelerated, less-rigorous version of OP's longtermist portfolio. Beckstead came from OP; the focus areas were identical. OP stepped in with emergency bridge funding after the collapse. The contrast in governance (OP has extensive public documentation, professional staff, and board independence) makes FF's failures more stark.

• SFF (Survival and Flourishing Fund): Jaan Tallinn (SFF's principal donor) was also a seed investor in Alameda. After FTX's collapse, SFF and Tallinn increased AI safety giving, partially filling the gap. The irony: Tallinn's money helped launch both Alameda and the recovery from its collapse.

• LTFF (Long-Term Future Fund): FF's regranting program was a scaled-up version of what LTFF does. Post-collapse, LTFF's role as a small-grants funder for individual AI safety researchers became more important.

• Anthropic: The $500M SBF/Alameda investment (stolen funds) was a separate but connected action. Anthropic's governance isolation of SBF (non-voting shares, no board seat) proved prescient. The FTX estate's $1.34B recovery from Anthropic stake sales was the single largest creditor recovery. The EA network connections between FF staff and Anthropic (Balwit as Amodei's CoS, Aschenbrenner engaged to Balwit) persist.

• Redwood Research: FF's $6.6M grant was committed but never paid -- the most-documented instance of a real project killed by FF's collapse. This likely delayed Redwood's research program by months.

• CAIS (Center for AI Safety): Received $6.5M from FF, then faced a clawback probe. This dual harm -- losing future funding AND risking return of past funding -- exemplifies the worst-case outcome for grantees.

What Would Change This Assessment

• Discovery that FF team members had knowledge of the fraud: No evidence supports this, but if it emerged, the assessment would shift from "negligent trust" to "complicity."
• Publication of FF's internal grant evaluation documents: If these showed rigorous, thoughtful analysis, it would strengthen the case that FF was valuable despite its governance failures.
• Evidence that FF-funded AI safety research produced breakthrough results: If specific FF grants led to major AI safety advances, the utilitarian calculus tilts more favorably.
• Beckstead publishing a detailed post-mortem: A candid accounting of what he knew, when, and what he'd do differently would significantly update the assessment of FF's leadership.
• The EV independent investigation results being published in full: This could either exonerate or further implicate EA leadership.

Self-Critique

What's weakest in this analysis:

• I could not access EA Forum discussions directly, which means I'm missing the community's real-time deliberation -- the most nuanced discourse about FTX/FF occurred there, not in journalism.
• The counterfactual analysis (what would have happened if FF hadn't collapsed) is necessarily speculative. I can list the grants and projects, but I can't quantify the counterfactual impact of $25M+ in unfulfilled grants.
• My assessment leans heavily on the TIME investigation and New Statesman reporting. If those sources have significant errors, my narrative about the 2018 warnings would need revision.

Where this analysis could be biased:

• Hindsight bias is severe in this case. The 2018 warnings look damning in retrospect, but many people receive warnings about associates that turn out to be overblown. I've tried to note that no one predicted the specific fraud, but the narrative inevitably reads as "they should have known."
• I may underweight the genuine good FF's grants accomplished. The regranting model funded real research by real people. Some of those people are now doing important AI safety work. The utilitarian case that FF produced net positive expected value -- even accounting for the collapse -- is not obviously wrong.

What a thoughtful person who disagrees would say: "You're retroactively condemning people who made reasonable decisions under uncertainty. Beckstead was the most respected longtermist grantmaker in the world; SBF appeared to be a legitimate crypto billionaire; the 2018 warnings were about management style, not fraud; and FF's grants funded genuinely important work. The real failure was SBF's criminality, not FF's grantmaking. Your structural critique of EA is unfalsifiable -- you'd say the same things whether or not SBF committed fraud."

Single weakest claim: That the regranting model was "genuinely innovative." Open Phil had already experimented with regranting, and the EA community already had informal giving networks. FF scaled an existing practice; calling it "innovative" may overstate its novelty.

What information would most change my view: Beckstead's private account of what he knew about the 2018 warnings and why he chose to join FF anyway. If he had conducted genuine due diligence and reached a reasonable (if ultimately wrong) conclusion that SBF had changed, the assessment shifts significantly. If he simply chose not to investigate because the opportunity was too good, it confirms the worst version of the narrative.