← AI Safety Orgs

AI Risk Mitigation Fund (ARM Fund)

Funding

Non-EA funder. Different source.

Founded
2023
HQ
Berkeley, CA
Team
5
Structure
fiscally sponsored
Model
Donations

Theory of Change

ARM's stated theory of change: independent philanthropic funding for AI safety is essential because (1) AI labs have incentive misalignment -- "they may downplay important catastrophic risks, or overestimate their ability to mitigate risks, in order to promote their products," (2) the field needs viewpoint diversity -- "external researchers may be more comfortable voicing dissent with lab orthodoxy," (3) labs neglect foundational research areas like "mechanistic interpretability on small models, adversarial robustness, and AI alignment theory," and (4) labs don't invest in field-building or upskilling newcomers.

The founding rationale, from Linch Zhang's LTFF newsletter: "The idea is to create an excellent donation option for people solely concerned about AI Risk. They might not really buy the case for existential biosecurity, forecasting, longtermism, etc." ARM strips away EA/longtermist framing to capture donors motivated by AI catastrophic risk specifically.

What They Do

ARM launched in December 2023, spun out of the Long-Term Future Fund (LTFF). Three focus areas: technical AI safety research, AI policy, and capacity building / field-building.

Grants track record (all LTFF-era): 12 grants totaling ~$1.35M (Feb 2021 - Oct 2023) made by ARM team members while serving on the LTFF. Range $2,500-$300,000, median ~$43K. Notable grantees include SERI MATS ($300K, now a major training pipeline), Alexander Turner ($220K, shard theory), David Krueger ($200K, Cambridge compute), and Leap Labs ($195K, interpretability seed funding). The LTFF overall distributed $20M+ over five years, including $6.67M across 197 grants in 2023.

Post-launch grants: none publicly disclosed. ARM's website has stated since launch that "detailed announcements are being prepared." Over two years later, no specific ARM-era grants have been published.

Grant model: 100% pass-through -- all donations (minus processor fees) go directly to grantees. Overhead funded separately.

Key People

Caleb Parikh -- Head of EA Funds and ARM's institutional leader. Former MIRI researcher, evaluated $34M+ in grant applications. Also LTFF interim chair. Controls both the EA-branded and "non-EA"-branded funding channels.

Oliver Habryka -- Cofounder of Lightcone Infrastructure (ARM's organizational home) and LessWrong. Also a permanent LTFF fund manager. Central node connecting ARM, LTFF, LessWrong, and Lighthaven.

Thomas Larsen -- ED of Center for AI Policy (largely defunct due to funding), former MIRI alignment researcher. Assigns >20% P(catastrophe by 2030). Zvi Mowshowitz specifically praised "Larsen's ability to judge projects."

Team also includes Lawrence Chan (METR/ARC Evals) and Lauro Langosco (Cambridge PhD under David Krueger). Advisors: Aviv Ovadya (AI & democracy), Sam Bowman (Anthropic alignment), Adam Gleave (FAR AI CEO).

Money and Incentives

ARM has no independent financial reporting. It operates as a program of Lightcone Infrastructure, previously under Effective Ventures Foundation USA (EIN 47-1988398). No information exists on: total funds raised, total funds distributed, overhead costs, or donor identity.

The funding landscape gap is real. Coefficient Giving estimates AI safety receives ~20x less philanthropic funding than climate. Open Phil/Good Ventures provides >50% of philanthropic AI safety funding (some estimates ~80%). CG itself argues donors outside their network can achieve "2-5x as cost-effective" grants. Open Phil has pulled back from funding several categories (rationality community, Republican think tanks, post-alignment causes), creating gaps ARM could theoretically fill.

Organizational overhead is hidden. ARM's 100% pass-through model means Lightcone absorbs operational costs. How Lightcone funds this is unclear, particularly given Habryka's candid descriptions of Lightcone's fundraising struggles post-FTX: the dominant question for donors is "does it look good to give money to your organization?" and Lightcone faces acute status-seeking dynamics.

LTFF financial context: $6.67M distributed in 2023, targeting $700K/month ($8.4M/year). In Sept 2023, LTFF described itself as "unusually funding-constrained." Open Phil offered $3.5M in matching funds as part of "distancing." ARM was born during this financial stress.

Donor identity unknown. Whether ARM's donors are genuinely "non-EA" (the stated goal) or EA-adjacent people who prefer the AI-risk branding cannot be verified. If the latter, ARM is a zero-sum rebranding, not additive funding.

What Others Say

Zvi Mowshowitz (SFF recommender, 2025): Rates ARM "High confidence" -- "Seems very straightforwardly exactly what it is, a regranter usually in the low six figure range. Fellow recommenders were high on Larsen's ability to judge projects." By contrast, rates LTFF "Low confidence," worried about whether it "favors insiders or extracts a time or psychic tax on participants, favors legibility, or rewards 'being in the EA ecosystem.'" Suggests LTFF should only be considered if you "want to empower and strengthen the existing EA formal structures."

Benjamin Todd (80K Hours): Mentions ARM alongside Manifund as worth "further topping up" for "smaller, often individual grants" -- a minor endorsement in a list dominated by specific organizations like METR, Apollo, MATS, and Lightcone.

Asya Bergal (former LTFF chair, now Coefficient Giving): On LTFF: "skeptical of a funding dynamic that moved money primarily within existing social circles with relatively little supervision." Noted the fund was "recruiting from too narrow of a pool."

No direct criticism of ARM specifically exists. All substantive criticism targets the LTFF, from which ARM inherits personnel and institutional culture. ARM is too new and too small to have attracted independent scrutiny.

What's Absent

No post-launch grants published -- over 2 years after launch, the single most important data point for evaluating a funder does not exist.

No ARM-specific financials -- total raised, total distributed, number of donors, overhead costs all unknown.

No conflict of interest policy -- despite a documented COI (Langosco's PhD advisor Krueger received a $200K grant from the same team). The LTFF's draft COI policy explicitly argued against formal recusal: "a conservative COI policy would reliably fail to make the most valuable grants."

No documented separation from LTFF -- the promised "greater practical separation" hasn't visibly materialized. Same 5 people sit on both funds.

No public communication from ARM's leader -- Caleb Parikh has not publicly articulated ARM's strategy or differentiation from LTFF. The founding vision was articulated by Linch Zhang, whose current involvement is unclear.

No annual report, no grant evaluation, no application process, no timeline.

Recommended Reading

  1. LTFF Inaugural Newsletter (Dec 2023) -- The most candid source for understanding ARM's creation rationale, written by Linch Zhang. Explains why ARM was spun out, the shared personnel with LTFF, and plans for future separation. Newsletter link

  2. Zvi's Big Nonprofits Post 2025 -- The sharpest external assessment of ARM and LTFF. ARM gets "High confidence" while LTFF gets "Low confidence" -- the contrast reveals how insiders perceive the two funds differently. Substack link

  3. ARM's "The case for independent AI safety funding" -- ARM's own theory of change: 4 arguments for why non-lab funding matters. Short (448 words) but dense. airiskfund.com/why-ai

  4. Thomas Larsen on AI Measurement and Evaluation -- Podcast showing Larsen's worldview: >20% catastrophe probability, focus on evals and government preparedness. The policy expertise he brings to ARM. Podcast transcript

  5. Coefficient Giving: AI Safety Needs More Funders -- CG's structural case for funder diversity that validates ARM's niche. Key stat: AI safety gets ~20x less philanthropic funding than climate. coefficientgiving.org

Show Claude’s analysis
An opinionated read. Read the brief first to form your own view.

Stated Theory of Change

ARM's stated theory: AI labs have misaligned incentives and cannot be the sole guardians of safety. Independent philanthropic funding is needed to support (1) foundational technical safety research that labs neglect, (2) external policy researchers who can challenge lab orthodoxy, and (3) field-building to expand the small pool of safety researchers. ARM specifically positions itself as a donation channel for people who care about AI catastrophic risk but "might not be EA or longtermist" -- removing the ideological packaging to access a broader donor base.

The mechanism: ARM takes donations from non-EA-identified donors, pools them, and makes "hits-based" grants in the $2,500-$300,000 range to early-career researchers, seed-stage organizations, and policy projects. By doing this outside the EA Funds brand, ARM theoretically expands total philanthropic capital flowing to AI safety rather than simply redirecting EA donors.

Revealed Theory of Change

ARM's actions reveal something simpler than its stated theory: ARM is an LTFF brand extension designed to capture a different donor demographic.

Evidence:

  • All 5 ARM program associates are current or former LTFF fund managers
  • No documented differentiation in grant process, evaluation criteria, or funding bars
  • No post-launch grants published, so we cannot verify any grantmaking divergence
  • ARM's organizational home (Lightcone) is run by an ARM program associate who is also a permanent LTFF fund manager
  • The LTFF newsletter explicitly described ARM as sharing "many personnel and infrastructure with LTFF for now"
  • The promised "greater practical separation" has not visibly occurred in 2+ years

The revealed theory is that the same people who make LTFF grants can raise more money by offering a simpler, non-EA-branded donation option. This is not inherently bad -- if ARM truly captures new donors who wouldn't have given to LTFF, it's additively valuable regardless of whether the grantmaking is differentiated. But the value proposition hinges entirely on donor-base expansion, which ARM has not demonstrated.

There is one important wrinkle: Zvi's divergent ratings (ARM "High confidence" vs. LTFF "Low confidence") suggest insiders may perceive ARM as having a different grantmaking culture despite the personnel overlap. Specifically, Zvi praises Larsen's judgment and ARM's straightforward approach, while worrying that LTFF "favors insiders" and "rewards being in the EA ecosystem." It's possible that ARM's smaller scope and clearer focus on AI risk produces a different funding bar in practice, even with the same people. But this is speculation in the absence of published grants.

Key Assumptions

Assumption 1: ARM can attract genuinely new donors who wouldn't have given to LTFF or other EA-affiliated funds.

  • Evidence for: The AI safety funding gap is real (20x less than climate per CG). Many people care about AI risk but don't identify as EA. ARM's clean branding could reach them.
  • Evidence against: ARM's team, organizational home, and community connections are deeply EA. The AISafety.com donation guide (EA-ecosystem product) is ARM's most visible referral source. No evidence of outreach beyond EA-adjacent channels.
  • Testable: Compare donor demographics to LTFF's. ARM has not disclosed this.
  • If wrong: ARM is a zero-sum rebranding that fragments EA donors across two funds rather than expanding total funding.

Assumption 2: Small, judgment-based grantmaking by well-connected individuals produces better outcomes than larger, more structured grantmaking.

  • Evidence for: LTFF's hits include SERI MATS (became major program), Lennart Heim (became compute governance expert), Palisade Research (started from $98K). The 30-40% success rate on 2018-2019 grants is defensible for hits-based giving.
  • Evidence against: Asya Bergal's critique that LTFF "moved money primarily within existing social circles." LTFF's own admission of being "unable to provide strong assurances" that work isn't net-negative. Concerns about anonymous vetoes biasing against policy work.
  • Testable: Compare LTFF/ARM grantee outcomes to CG or SFF grantee outcomes. Not yet possible for ARM.
  • If wrong: ARM's grantees may be drawn disproportionately from the team's social network rather than the best available applicants.

Assumption 3: ARM's team can maintain good judgment at ARM's scale while also managing LTFF, Lightcone, METR, CAIP, and their other commitments.

  • Evidence for: The team members are reportedly effective part-time grantmakers. ARM's grant volume is probably small enough not to overwhelm.
  • Evidence against: Larsen's CAIP ran out of funding. Habryka is managing Lightcone's infrastructure + fundraising + LessWrong + ARM + LTFF. Parikh runs all of EA Funds. Chan works at METR full-time. Langosco is a PhD student. These are busy people with many competing demands.
  • Testable: ARM's grant volume and turnaround time, if disclosed, would reveal capacity.
  • If wrong: ARM becomes a low-priority side project that makes occasional grants with less scrutiny than LTFF.

Assumption 4: The 100% pass-through model is sustainable.

  • Evidence for: Common for small funds. Lightcone can absorb overhead for a small operation.
  • Evidence against: Lightcone itself has severe fundraising challenges (Habryka's "17 donors of last resort"). If Lightcone's funding becomes more constrained, ARM overhead may become a burden.
  • If wrong: ARM may need to charge overhead or find independent funding, undermining the pass-through commitment.

Strengths

Genuine niche in a real funding gap. Coefficient Giving's own analysis confirms that funder diversity in AI safety is critically needed. ARM's simple, non-EA-branded presentation could reach donors who bounce off EA/longtermist framing. Even CG recommends that non-GV donors can achieve 2-5x cost-effectiveness.

Strong team expertise. The program associates collectively have deep knowledge of the AI safety funding landscape. Larsen brings policy expertise (Zvi specifically praises his judgment). Chan brings technical evals experience from METR. Habryka brings deep community knowledge. This is a well-connected team that knows where the funding gaps are.

Track record of identifying talent early. The LTFF grants now claimed by ARM include genuine wins: SERI MATS, Lennart Heim, Palisade Research, AXRP. Funding early-career researchers before they're legible to larger funders is a valuable and underserved function.

Zvi's "High confidence" endorsement. In a post where he rates dozens of nonprofits with notable candor, Zvi's unconditional positive assessment of ARM stands out, especially contrasted with his concerns about LTFF. This suggests ARM has earned trust from sophisticated evaluators.

Low overhead, simple model. 100% pass-through is donor-friendly and structurally transparent about costs.

Weaknesses and Risks

Zero published post-launch grants is a critical red flag. A funder that publishes no grants for 2+ years cannot be evaluated on its actual grantmaking. The website's "announcements are being prepared" language has been static since Dec 2023. This could mean ARM has made grants but hasn't published them (bad transparency), or ARM has made few/no grants (bad execution). Either interpretation is concerning.

"Non-EA" branding is misleading. ARM is run by the head of EA Funds, housed at the LessWrong cofounder's organization, staffed entirely by LTFF fund managers, and was announced in an LTFF newsletter. Calling this "non-EA" stretches credibility. A more honest framing would be "EA grantmaking, minus the longtermist packaging."

Complete governance vacuum. No board, no COI policy, no documented decision process, no external oversight. For a grantmaking organization -- even a small one -- this is structurally inadequate. The documented Krueger-Langosco COI is unaddressed.

No evidence of donor-base expansion. The central value proposition -- reaching non-EA donors -- has not been demonstrated. ARM's most visible referral sources (AISafety.com, EA Forum, 80K Hours) are all EA-ecosystem products.

Inherited LTFF concerns. Bergal's "social circles" critique, Zvi's "insider bias" worry, the anonymous policy vetoes, the "unable to provide strong assurances" admission -- these all apply to ARM given identical personnel. ARM was supposed to be the cleaner version, but without documented reforms, it may simply replicate LTFF's failure modes with less oversight.

Key person risk is extreme. Habryka is the organizational infrastructure. Parikh is the institutional connection to EA Funds. Larsen is the praised grantmaking judgment. If any one leaves, ARM's value proposition changes dramatically.

Cross-References

vs. LTFF: ARM is functionally LTFF's AI-risk-focused, non-EA-branded sibling. Same team, same grantmaking philosophy, different audience. If you trust the team and want AI-risk-focused giving without EA branding, ARM is the channel. If you want the broader longtermist portfolio (biosecurity, forecasting), LTFF is the channel.

vs. Manifund: Both serve small, often individual grants. Manifund operates a regranting model with many independent regranters; ARM operates a centralized panel model. Manifund is more transparent (public project pages, donation tracking). ARM is more curated.

vs. SFF: SFF operates at much larger scale ($19M in a recent round) and uses the S-process (multiple recommenders, democratic allocation). ARM is vastly smaller and uses centralized judgment. SFF funded many of the same types of organizations but at higher scale.

vs. Longview Philanthropy: Longview advises larger donors ($250K+) and runs pooled funds. ARM targets smaller donors. They serve different donor segments.

vs. Coefficient Giving: CG is the 800-pound gorilla ($100M+ annual in AI safety). ARM is explicitly positioned as a non-CG funding channel that fills gaps CG creates. CG itself validates this positioning.

What Would Change This Assessment

Upward update: ARM publishes a list of post-launch grants showing (a) grants to applicants outside the team's social network, (b) grants in areas LTFF historically underfunded (e.g., policy), (c) evidence of new non-EA donors, (d) a documented COI policy.

Downward update: ARM continues publishing no grants for another year. Or: ARM's published grants turn out to be identical to what the team would have funded through LTFF. Or: ARM's donor base turns out to be entirely EA-community members.

Neutral information that would sharpen the picture: ARM's total funds raised and distributed. Donor demographics (EA vs. non-EA). Whether the separation from LTFF has progressed.

Self-Critique

Weakest claim: That ARM's "non-EA" framing is misleading. It's possible that the framing genuinely resonates with a distinct donor demographic even though the team is EA-embedded. I'm inferring deception from circumstantial evidence when the reality might be that the branding works as intended.

What I might be wrong about: ARM may have made substantial grants that it simply hasn't published yet, perhaps for legitimate confidentiality reasons (LTFF makes some anonymous grants). The 2+ year publication gap might reflect real privacy concerns rather than inactivity.

Sources I couldn't access: EA Forum and LessWrong posts about LTFF criticism, Bergal's reflections, and the COI policy draft were only accessible via search summaries, not full text (these sites block bot access). Full text of these posts would provide more nuance.

Potential bias: I may be holding ARM to too high a standard for a very small, new fund. Many grantmaking organizations take years to publish grant lists or develop governance infrastructure. ARM's team may be doing excellent work that simply hasn't been publicly documented yet.

What would most change my view: Seeing ARM's actual post-launch grants. This single piece of information would transform the analysis from speculation about a brand to evaluation of a grantmaker.

Connected to (8)

Effective Ventures Foundation USAspun off from
Lightcone Infrastructureboard overlap · Oliver Habryka
Anthropicadvisor at · Sam Bowman
Center for AI Policyboard overlap · Thomas Larsen
EA Fundsboard overlap · Caleb Parikh
FAR AIadvisor at · Adam Gleave
Long-Term Future Fundspun off from · Linch Zhang
METRstaff from · Lawrence Chan
Sources (27)
Every URL that was read during research.
  1. 1.AI Risk Mitigation Fundairiskfund.com
  2. 2.AI Risk Mitigation Fundairiskfund.com
  3. 3.It looks like there are some good funding opportunities in AI safety right now | 80,000 Hours80000hours.org
  4. 4.The Big Nonprofits Post 2025thezvi.substack.com
  5. 5.Long-Term Future Fund's Inaugural Newslettermailchi.mp
  6. 6.AI Safety and Security Need More Funderscoefficientgiving.org
  7. 7.Bits and bricks: Oliver Habryka on LessWrong, LightHaven, and community infrastructurecomplexsystemspodcast.com
  8. 8.Homecalebp98.github.io
  9. 9.#1: Thomas Larsen on AI Measurement and Evaluationaipolicypod.substack.com
  10. 10.AI Security Forum Substack | Caleb Parikh | Substackaisecurityforum.substack.com
  11. 11.Lawrence Chanchanlawrence.me
  12. 12.Episode 6 - Oliver Habryka on LessWrong and other projectsthefilancabinet.com
  13. 13.Funding – AISafety.comaisafety.com
  14. 14.Long-Term Future Fund | Effective Altruism Fundsfunds.effectivealtruism.org
  15. 15.Safety Cases: Justifying the Safety of Advanced AI Systems | Center for AI Policy | CAIPcenteraipolicy.org
  16. 16.Effective Ventures Foundation Usa - Nonprofit Explorer - ProPublicaprojects.propublica.org
  17. 17.Donation guide – AISafety.comaisafety.com
  18. 18.AI Risk Mitigation Fund Funding | Complete Analysisextruct.ai
  19. 19.Sam Bowmansleepinyourhat.github.io
  20. 20.aviv ovadyaaviv.me
  21. 21.Adam Gleavegleave.me
  22. 22.Who’s Funding AI Regulation and Safety?insidephilanthropy.com
  23. 23.The Center for AI Policy (CAIP)centeraipolicy.org
  24. 24.Ep 11 - Technical alignment overview w/ Thomas Larsen (Director of Strategy, Center for AI Policy) - Artificial General Intelligence (AGI) Show with Soroush Pourtheagishow.com
  25. 25.Lawrence Chan | FAR.AIfar.ai
  26. 26.Effective donations | (More) philanthropy is needed for safe AIeffektiv-spenden.org
  27. 27.This century is critical for humanity. We build tech, infrastructure, and community to navigate it.lightconeinfrastructure.com