AI Safety Support

Stated Theory of Change

AISS had two stated theories of change, though it didn't clearly distinguish them:

1. People support: Provide career coaching, health coaching, community infrastructure (Slack, events, newsletter), and mentorship matching to early-career, independent, and transitioning AI safety researchers. By removing barriers to entry and reducing isolation, more people successfully enter and persist in the field, increasing total research capacity.

2. Institutional infrastructure: Serve as fiscal sponsor for AI safety programs that lack their own legal entity, enabling them to receive grants and hire staff. By providing administrative infrastructure, programs can focus on their mission rather than on nonprofit compliance.

The first theory of change is about individuals. The second is about institutions. JJ Hepburn articulated the first more clearly -- his forum posts are about lowering barriers for individual researchers, not about fiscal sponsorship mechanics. But the second is what attracted serious funding and continues today.

Revealed Theory of Change

AISS's actions reveal a pattern: the organization was excellent at building infrastructure for other people's work but could not build institutional support for its own. The $7.85M flowing through AISS for MATS, Astra, etc. demonstrates that JJ and team were trusted stewards of significant funds. The $425K received for AISS's own mission demonstrates that funders did not value what AISS itself was doing -- or at least did not value it enough to sustain it.

The revealed theory of change of the successor entity (Ashgro) is cleaner: be the best possible administrative infrastructure for AI safety projects. Ashgro's $4.2M revenue and positive client testimonials suggest this is working. The theory of change for the support services was never validated because it was never adequately funded or measured.

What actually happened is that JJ pivoted from a theory of change he cared about (supporting individual researchers) to one the market rewarded (fiscal sponsorship). Ashgro is JJ's answer to the question "how do I contribute to AI safety in a way that is both useful and sustainable?" -- and the answer turned out to be "operations, not coaching."

Key Assumptions

For the support services (defunct):

1. Individual career coaching meaningfully increases the probability someone enters/persists in AI safety. Evidence: 5 testimonials, all positive. Counter-evidence: no quantitative data. This assumption is plausible but untested. Career coaching in other domains has mixed evidence on long-term outcomes.

2. The people AISS served would not have found their way without help. Evidence: testimonials from people with non-traditional backgrounds who felt excluded. Counter-evidence: other pathways exist (AISC, MATS, 80K Hours). AISS may have been serving people who would have found other paths, or it may have been the only path for some. Unknowable without data.

3. Community infrastructure (Slack, events) increases research productivity of independent researchers. Evidence: one participant quote saying "I'm tired of reading all the time. I'm also hungry for discussions." Counter-evidence: none, but also no measurement of productivity gains. Plausible based on research on isolated workers in other fields.

For the fiscal sponsorship (active):

1. Without fiscal sponsors, promising AI safety projects would fail to start or would waste significant time on legal/administrative setup. Evidence: strong. Ashgro client testimonials confirm this. The AISC page states "Ashgro handles our financial admin." Fiscal sponsorship is a well-established model with clear benefits.

2. Ashgro specifically is better at this than alternatives (BERI, Effective Ventures, Rethink Priorities). Evidence: moderate. Ashgro's focus on AI safety and responsive service are cited by clients. But the evidence is self-selected -- we don't hear from projects that chose other sponsors or didn't need one.

Strengths

Ashgro is a genuine competitive advantage for AI safety. A well-run fiscal sponsor that specializes in AI safety and charges 5-10% is valuable infrastructure. The client testimonials, the project diversity (Timaeus, Apart, Robert Miles, AISC, AFFINE, CORAL), and the growth trajectory all suggest real product-market fit. JJ's near-zero salary signals mission alignment.

AISS identified a real gap. The existence of AI Safety Quest (its volunteer successor) and the continued demand for career navigation services confirms that the need AISS addressed was genuine. The gap between "interested in AI safety" and "working in AI safety" is real, and institutional programs like MATS only serve a fraction of the interested population.

Low overhead, high trust. JJ's approach -- unpaid work, minimal staff, functional focus -- generates trust. Ashgro's growth from $71K to $4.2M in two years, with projects actively seeking it out, suggests word-of-mouth reputation building.

JJ's philosophical approach was sound. "Just do a project. Drop the constraints." This anti-gatekeeping stance is the right corrective to an ecosystem that can feel exclusionary to non-traditional entrants. The coaching testimonials specifically cite this as valuable.

Weaknesses and Risks

AISS failed to measure its own impact, and this likely contributed to its death. The absence of any quantitative data on coaching sessions, users served, or outcomes meant AISS could not make a compelling case to funders. In a funding ecosystem that rewards legible metrics, "we had some good conversations" is not enough. This is a cautionary tale for any community infrastructure org.

Ashgro has key-person risk. JJ Hepburn is the CEO, founder, and primary relationship holder. The organization is hiring its first operations associate. If JJ left, it is unclear whether Ashgro would maintain its client relationships and reputation. The $0 salary also creates a sustainability question -- what happens when JJ needs to earn a living?

Governance gaps are growing. Ashgro processes $4.2M with no published conflict-of-interest policies, no recusal procedures, and a board member (Gaidashenko) who works at a sponsored project (FAR.AI). At current scale this is manageable; at 2x scale it becomes a real risk.

The support services theory of change was never validated. We don't know if AISS's coaching moved anyone into AI safety who wouldn't have gotten there otherwise. Five testimonials are better than zero but far from evidence of impact at scale. The theory of change died without being either confirmed or refuted.

Ashgro's revenue depends on a small number of large funders. CG/Open Phil grants to MATS alone account for the majority of Ashgro's revenue (as flow-through). If MATS moved to a different fiscal sponsor or set up its own entity, Ashgro's revenue would crater. Concentration risk is high.

Cross-References

MATS: AISS's largest fiscal sponsorship client. MATS is now an independent 501(c)(3) (EIN: 99-0648563) as of the MATS website, yet CG grants still go to "AI Safety Support" for MATS as recently as October 2024. The relationship between MATS, AISS Ltd, and Ashgro is unclear and likely transitional.

AI Safety Camp (AISC): AISS's sibling organization -- both were co-founded by Linda Linsefors, both use Ashgro for fiscal administration. AISC is $200K in SFF 2025. The organizations share DNA but serve different populations (AISC: project teams; AISS: individuals).

BERI / Effective Ventures: Other fiscal sponsors in the EA/AI safety space. Ashgro's niche is AI-safety-specific focus and lower overhead.

80,000 Hours: Provides career advice to a much larger audience but is less personalized. AISS served people who 80K's content didn't reach or wasn't specific enough for.

AI Safety Quest: Direct successor to AISS's support services. Volunteer-based, unfunded, lower quality but nonzero. Evidence that the need persists.

What Would Change This Assessment

• Impact data from AISS. If someone surfaced records showing AISS coached 500+ people and 50 entered AI safety roles, the "support services" theory of change would look much stronger, and the shutdown would look like a bigger loss.
• Ashgro losing a major client. If MATS fully transitioned away from Ashgro, the fiscal sponsorship model's sustainability would be tested.
• JJ Hepburn leaving Ashgro. Would reveal whether the org is JJ-dependent or has institutional resilience.
• An ACNC filing showing AISS Ltd's actual finances. If AISS was earning 5-10% fees on $7.85M flow-through ($390K-$785K), the "$425K in own funding" narrative would need revision -- AISS may have been less underfunded than it appeared.
• A well-argued critique of fiscal sponsorship as a model. If someone demonstrated that fiscal sponsors create dependency and prevent projects from building institutional capacity, Ashgro's theory of change would need re-evaluation.

Self-Critique

What sources should I have checked but didn't?

• ACNC annual reporting for AISS Ltd (fetch failed due to timeout)
• The full EA Forum comment thread on the shutdown post (behind bot detection -- I relied on search snippets)
• JJ Hepburn's LinkedIn profile (inaccessible)
• Ashgro's actual 990 forms (only ProPublica extracts available)

Where is this analysis potentially biased?

• I may be overly sympathetic to AISS/JJ because the evidence is almost entirely self-reported or from supporters. No critics exist to balance the narrative.
• I may be overstating the systemic funding critique. Perhaps AISS was genuinely low-impact and funders were right not to prioritize it. The absence of impact data cuts both ways.

What would a thoughtful person who disagrees say? "AISS was a small, unmeasured community project that ran for a few years and shut down when funding dried up. This is normal nonprofit lifecycle, not a systemic failure. The fiscal sponsorship role was always more valuable and appropriately attracted more funding. JJ recognized this and pivoted successfully to Ashgro. The 'tragedy of underfunding' narrative is sentimental -- funders make hard choices and they chose to fund MATS over AISS's coaching calls because MATS has 446 alumni, 170 papers, and an h-index of 44."

What's my single weakest claim? That the LTFF donor-signaling failure "likely killed AISS." This is based on search snippets of comment-thread discussions, not direct reading of the comments. The causal chain (LTFF indecision -> donor holds back -> AISS runs out of money) is plausible but I'm relying on indirect evidence.

What information would most change my view? AISS's actual financial statements. If AISS was earning fiscal sponsorship fees on the $7.85M flow-through, the narrative of "underfunded org dies while processing millions" would need significant revision. AISS might have had adequate funding but chose to spend it on fiscal sponsorship operations rather than its own support services -- a very different story.