Manifund

Stated Theory of Change

Manifund's stated theory of change has evolved through two phases:

Phase 1 (2022-2024): Innovate funding mechanisms. Build the world's first impact certificate marketplace to create "venture funding for charitable endeavors." The elegant mechanism: investors buy shares in projects, which pay out if the project later receives retroactive prizes. This aligns incentives by rewarding people who correctly identify high-impact work early. Championed by Paul Christiano, Vitalik Buterin, and Scott Alexander.

Phase 2 (2024-present): Decentralize AI safety funding. After impact certs failed to find product-market fit, Manifund pivoted to regranting as the primary theory of change. The causal chain: (1) AI safety funding is too concentrated in a few large institutions (CG/OpenPhil, SFF), which move slowly and miss opportunities. (2) Domain experts with $100K+ budgets can make faster, more informed funding decisions because they have private information about people and projects in their sub-field. (3) Manifund provides the transparent infrastructure (platform, 501(c)(3), payment processing) that lets this happen. (4) The result is faster capital deployment to early-stage AI safety work, seeding new organizations before larger funders notice them.

The secondary theory of change is physical infrastructure: Mox provides the coworking space, events, and community hub that makes SF's AI safety community more productive through colocation and serendipitous interaction.

Revealed Theory of Change

The revealed theory of change diverges from the stated one in several important ways:

Austin Chen's real orientation is "Winning," not AI safety. His North Star, stated explicitly on a podcast, is "Winning." His idols are "Bill Gates and SBF" -- builders who created massive value through aggressive action, not careful stewardship. His defense of SBF ("should be freed and put back to work") and his pattern of platforming controversial speakers at Manifest both reveal someone who optimizes for boldness and contrarianism over prudence. Manifund's AI safety focus is sincere but instrumental -- it's the field where Austin can have impact given his network and capabilities, not a deep personal commitment to safety research.

The "neutral platform" positioning masks governance gaps. Manifund frames itself as infrastructure rather than a decision-maker. But someone chose the regrantors. Someone chose donor "D." Someone chose which speakers to invite to Manifest. The "neutral platform" framing serves to deflect accountability for these choices. When Dan Hendrycks used $250K of regrantor budget to fund his own organization, the response was "heuristics for regrantors to keep in mind" rather than a formal policy -- consistent with a culture that values speed and autonomy over governance.

Mox is becoming the real operation. With $1.2M in annual costs, 183 members, and 377 events, Mox is now larger than Manifund's core grantmaking infrastructure (which operates on ~$150K-250K in actual operating revenue). The revealed theory of change may be shifting from "regranting platform" to "AI safety community infrastructure provider."

Manifest reveals values priorities. Three years of inviting the same HBD-adjacent speakers, with explicit refusal to change ("done ~nothing to reduce the likelihood that I invite speakers who the commenters consider racist"), reveals that intellectual freedom and edginess are valued more highly than community safety or reputational stewardship. This is consistent with Austin's "Winning" orientation but in tension with serving as a trusted fiduciary for AI safety funding.

Key Assumptions

1. Decentralized grantmaking outperforms centralized grantmaking for AI safety. Manifund assumes that giving 10 people $100K+ each and letting them make independent decisions produces better outcomes than having a single team evaluate applications.

• Evidence for: Venture capital literature supports power-law returns from distributed bets. The Timaeus early funding success demonstrates genuine value from regrantor private information. FTX Future Fund, which pioneered this model, was widely praised before FTX collapsed.
• Evidence against: No systematic outcome evaluation exists. Austin admits he can't evaluate whether grants are good. The worst-case regrant (Epoch benchmark) may have accelerated capabilities. Large single-regrantor grants to established orgs don't demonstrate regranting's comparative advantage.
• Testable: Compare outcomes of Manifund-funded projects vs. LTFF or SFF-funded projects at similar stages and amounts. Neither side has done this.
• What changes if wrong: If regranting doesn't outperform, Manifund's 5% fee is a tax on the ecosystem without added value.

2. Speed matters more than accuracy in AI safety funding. The core pitch is "<1 week from recommendation to money." This assumes time-sensitive opportunities regularly arise and that faster deployment compensates for less evaluation.

• Evidence for: Timaeus early funding example (months earlier than SFF). Real cases where compute bottlenecks needed immediate resolution.
• Evidence against: Many grants are not time-sensitive. The end-of-year budget rush in 2023 (when regrantors scrambled to spend remaining budgets) produced worse grants than earlier, more deliberate ones.
• Testable: Tag each grant as "time-sensitive" or not, and compare outcomes.

3. Transparency creates a positive selection effect. Manifund publishes all grants with regrantor reasoning. The assumption is that this transparency attracts better projects and donors, and serves as a public good for the ecosystem.

• Evidence for: Grantee quotes praising the public discussion. Nuno Sempere's endorsement. The Lantern Bioworks proposal reaching #1 on Hacker News demonstrates visibility value.
• Evidence against: The Epoch grant was criticized precisely for being confidential within a transparency-focused system. Some regrantors may avoid controversial-but-important grants because they're public.

4. The AI safety funding ecosystem is under-diversified, and more funders help. Manifund positions itself as an alternative to CG/OpenPhil dominance. The assumption is that funder diversity makes the ecosystem more robust.

• Evidence for: Post-FTX, the concentration risk became obvious. Moskovitz pulling back from rationalist events shows how one funder's preferences can constrain an entire community. Multiple independent funders with different evaluation approaches reduce single-point-of-failure risk.
• Evidence against: Manifund's own funding is extremely concentrated (anonymous "D" for regranting, SFF for operations). They've reduced concentration at one level while introducing it at another.

Strengths

Genuine technical competence. Manifund ships fast, builds good software, and operates with extreme leanness. Running a $3M+/year grants program on $83K in total compensation is extraordinary. The platform itself (transparent applications, public reasoning, fast payments) is genuinely better than alternatives.

Intellectual honesty about failure. The impact certificate pivot is one of the most honest admissions of failure in EA. The regrant review naming bad grants alongside good ones is rare. Self-ratings of 2/10 (OpenPhil AI Worldviews) and 4/10 (Manifold Charity) show genuine self-assessment. This culture is Austin's best contribution.

Demonstrated early-mover advantage. Timaeus, ChinaTalk, and several other grants demonstrate Manifund's ability to fund projects before larger funders notice them. This is the core value proposition, and it works.

Mox as community infrastructure. 183 members, partnerships with 19 AI safety orgs, 377 events. This is arguably Manifund's highest-impact output -- providing physical infrastructure for the AI safety community's most productive hub.

Weaknesses and Risks

Governance is structurally absent. Three-person board (CEO + two), no COI policy, no whistleblower policy, no grant evaluation, no annual review for 2024. This is not a startup in its first months -- Manifund has distributed $5M+ and is entering its fourth year. The Hendrycks self-dealing incident was identified but not structurally prevented from recurring. The Maini board COI (AI VC fund + former DeepMind + charity board) is unaddressed.

The Manifest albatross. Three years of platforming HBD-adjacent speakers, with explicit refusal to stop, creates reputational damage that extends to Manifund's core grantmaking mission. Dustin Moskovitz's public displeasure, the departure of senior EA community figures, and ongoing external criticism all reduce Manifund's ability to attract mainstream donors and grantees. The CEO's personal defense of every controversial speaker compounds this.

The SBF defense undermines trust. The CEO of a nonprofit seeded with FTX money publicly argues the convicted fraudster "should be freed and put back to work." In a community still reeling from SBF's destruction of $100M+ in EA funding, this judgment call raises questions about Austin's capacity for evaluating trustworthiness -- a core skill for someone running a grantmaking platform.

Donor concentration creates fragility. Anonymous "D" provides $1.5M/year for regranting. SFF provides general support. OpenPhil has declined to fund. If "D" or SFF withdraw, Manifund's core operations are at risk. The 5% fee generates ~$150K, not enough to sustain even current minimal staffing.

No evidence that regranting works better. Despite four years of operation, no outcome evaluation exists. Austin admits he can't evaluate whether grants are good. The claim that regranting outperforms alternatives is plausible but unverified. The regrant review (Jesse Richardson) evaluates process quality, not outcome quality.

Cross-References

SFF (Survival and Flourishing Fund): Manifund's operational patron. SFF provides general support grants and funded the small regrantor budgets in 2023. SFF also serves as a retroactive funder for ACX Grants impact market. The relationship is one of dependent-beneficiary: Manifund depends on SFF for survival but SFF doesn't depend on Manifund. Manifund's regranting model is complementary to SFF's S-process -- both champion-based, but Manifund operates at smaller scale with faster turnaround.

LTFF (Long-Term Future Fund): Manifund's closest competitor/complement. LTFF funded Manifold Markets ($200K) and participates as retroactive funder in ACX Grants. Manifund's 2023 review discusses close collaboration with EA Funds, including building a grantee dashboard (leaf-board.org) and exploring "common app" for EA. Nuno Sempere's assessment explicitly positions Manifund as potentially replacing EA Funds through superior technical competence.

Manifold Markets: Sibling entity. Same founders, shared origins, now formally separate. Manifold received $1.5M from FTX, $1.2M from SFF, $200K from LTFF. The prediction market company's pivot struggles ("feels trapped in local optima") explain why Austin left for Manifund. The GitHub repo is still under manifoldmarkets/manifund.

Lightcone Infrastructure: Hosts Manifest at Lighthaven. Received $300K loan from Manifund. Builds the S-Process software for SFF. The Manifold/Manifund/Manifest/Lightcone web of relationships is dense: shared events, loans, conference venues, and overlapping community.

CG/OpenPhil: The notable non-relationship. OpenPhil declined to fund Manifund. This means the largest AI safety funder evaluated and rejected Manifund's model, its team, or both. Manifund positions itself as an alternative to CG dominance but operates at 1/100th the scale.

What Would Change This Assessment

• Manifund publishes a systematic grant outcome evaluation. Even a simple "did the funded project produce what was promised?" analysis would be transformative. Would either validate or invalidate the regranting model.
• Manifund formally separates from Manifest or Manifest stops inviting HBD speakers. Either would resolve the reputational problem.
• Anonymous "D" is revealed or replaced by diversified donors. Would address the donor concentration risk.
• The board expands to 5+ members with real independence. Would address the governance gap.
• OpenPhil/CG funds Manifund. Would be a strong signal that mainstream institutional funders endorse the model.
• Austin Chen publicly updates on SBF. Acknowledging the fraud was bad would restore some trust in his judgment on character assessment.
• A major scandal involving a regrantor grant (fraud, capabilities acceleration, self-dealing at scale). Would demonstrate that the governance-free model has real costs.

Self-Critique

What sources should I have checked but didn't?

• The full "Winning, Risk-Taking, and FTX" podcast transcript (only stubs available)
• Detailed 2024 990 filing (not accessible via ProPublica)
• EA Forum comment threads on the 2023 review and 2025 regrants post (comments contain valuable community reactions)
• Whether any Manifund grantees have publicly criticized the Manifest controversy

Where is this analysis potentially biased?

• I may be over-weighting the Manifest controversy relative to core operations. The HBD speakers are invited to a conference adjacent to Manifund, not to Manifund itself. The grantmaking could be excellent even if the conference is bad.
• I may be under-weighting the value of Mox, which is harder to analyze from public sources but may be Manifund's highest-impact contribution.
• I may be applying institutional governance standards to a 3-person startup, where informal governance is standard and appropriate.

What would a thoughtful person who disagrees say? "You're conflating Manifest with Manifund. The conference is organized by Manifold/Manifund staff in their spare time, not as a Manifund program. The grants are what matter, and the grants are good -- Timaeus, ChinaTalk, MATS, Apollo. Austin's personal views on SBF are irrelevant to whether the regranting model works. And the governance criticism is premature -- would you apply these standards to a 3-person YC startup?"

What's my single weakest claim? That "the CEO's SBF defense undermines trust." Many thoughtful people have heterodox views about criminal justice and proportionality of punishment. Austin's SBF defense, while startling, may reflect contrarianism rather than poor character judgment. His track record of selecting good regrantors (Evan Hubinger, Neel Nanda, Joel Becker) is better evidence of his judgment than a podcast hot take.

What information would most change my view? A systematic evaluation showing that Manifund-funded projects produced meaningfully better outcomes than comparable projects funded by LTFF or SFF at similar stages. If such evidence existed, the governance concerns would be secondary to demonstrated effectiveness. Conversely, evidence that Manifund-funded projects systematically underperformed would make the governance concerns urgent.