Theory of Change
Atlas Fellowship aimed to identify exceptionally talented high school students globally (ages 13-19), expose them to rationality and AI safety thinking through an 11-day intensive program, provide scholarships and community, and hope they would pursue ambitious work on important problems.
In their own words: "We empower exceptional students to understand and change our future. Like an atlas, we help you navigate the world: in our programs, fellows practice careful reasoning that enables them to make sense of confusing evidence, mitigate cognitive biases, and develop accurate beliefs that map onto reality."
The program explicitly drew on CFAR/rationality traditions ("We've adapted ideas from CFAR, though we are not affiliated") and effective altruism, but positioned itself as "an experiment in going beyond these communities." The FAQ listed inspirations including "Emergent Ventures, quantitative trading, progress studies, RadicalXChange, entrepreneurship" alongside the usual EA references.
The causal chain: select exceptional teens -> 11-day immersion in rationality + AI safety + agency -> $50K scholarship (later $10K) -> community + mentorship -> fellows pursue ambitious work on global challenges -> reduced AI risk. This chain is long and speculative, depending on multiple assumptions about counterfactual career trajectories.
Status: DEFUNCT as of November 2023. Co-founder Jonas Vollmer's wind-down letter cited three reasons: (1) changed funding landscape won't support scaling, (2) "our program succeeded... to a lesser degree than I initially hoped," (3) staff concluded they'd have more impact doing direct AI safety work. He linked to a Manifold prediction market on outcomes as an accountability measure.
What They Do
Ran from January 2022 to November 2023. Two cohorts:
- 2022: 111 fellows from 18 countries selected from 7,000+ applicants (1.8% acceptance rate). Five 11-day programs in Berkeley and Oxford. $50,000 scholarships.
- 2023: ~120 additional fellows. Programs in Berkeley only. Scholarships reduced to $10,000 per fellow ($1,000 for finalists).
Total: ~230 fellows. Program covered all costs (flights, visa, housing, food) in addition to scholarships.
Curriculum included: Fermi estimates, interpretability of small transformers, poverty analysis, AGI risk, epistemic rationality, forecasting, hacker ethics, and fun activities (juggling, chess trading markets). Application process tested analytical thinking with creative questions, not polished essays.
Also ran Atlas India program (Oxford, UK, Rs 10 Lakh scholarship) and Atlas Fund (fast grants + Metaplanet VC partnership for fellow investments in startups).
Outcomes: Jonas Vollmer's Manifold market on whether 20+ fellows would be "actively working on ambitious projects" by October 2024 resolved at 50%. His estimate: "around 20" qualifying fellows (range 16-28) out of ~230 total, or roughly 9%. In a Coefficient Giving survey of 329 AI safety workers, Atlas Fellowship was listed as a top-4 career influence by 2% of respondents (6 people), comparable to ESPR.
Key People
Jonas Vollmer -- Co-founder, President until April 2023. Previously ran EA Funds (2020-2022), co-founded Center on Long-Term Risk (2013). Now COO at AI Futures Project (the "AI 2027" scenario forecast read by JD Vance). Board of Macroscopic Ventures (Swiss nonprofit VC that invested in Anthropic Series A & B, Apollo Research, Gray Swan). Sits at an unusual number of nodes in the AI safety financial ecosystem.
Leopold Aschenbrenner -- President since April 2023 (caretaker of residual entity). Columbia valedictorian at 19. Was on FTX Future Fund team that granted $5M to Atlas, then became Atlas president. Later at OpenAI Superalignment (fired 2024), published "Situational Awareness," founded a $1.5B hedge fund. Engaged to Avital Balwit (Anthropic chief of staff). Receives $0 from Atlas.
Sydney Von Arx (now at METR, adversarial stress-testing) and Ashley Lin (now at CSET Georgetown, AI policy) -- co-founders who both moved to direct AI safety work after closure.
Money and Incentives
Total funding: ~$13.8M from exactly two sources:
- Open Philanthropy: $8.8M across three grants (March 2022, December 2022, April 2023)
- FTX Future Fund: $5M committed June 2022
Cost per fellow: ~$60,000 ($13.8M / 230 fellows). Critics estimated $80-90K including overhead. SPARC, a comparable rationality program for high schoolers, operates entirely for free.
Cost per "active outcome": ~$500K-$690K ($13.8M / 20-28 fellows working on ambitious projects as of October 2024). This is a rough calculation that ignores long-term effects and the fact that most fellows are still in college.
990 financial trajectory:
| Year | Revenue | Expenses | Net Assets |
|---|---|---|---|
| 2022 | $6.6M | $3.5M | $3.1M |
| 2023 | $4.9M | $4.5M | $801K |
| 2024 | ~$640K | ~$835K | ~$538K |
Compensation was lean: total executive compensation peaked at $155K in 2023 (Jonas $94K, Ashley $41K, Sydney $19K). Zero compensation in 2024.
The residual entity still holds $2.18M in total assets and $1.64M in liabilities (2024). It generates $261K/year in rental income from Berkeley real property and $123K in investment income. No programming. No public plan for asset disposition. $300K in contributions received in 2024 despite no active programs.
FTX clawback status unknown. Whether Atlas received the full $5M from FTX, and whether any was returned to the FTX estate, is not publicly documented. The 2022 990 shows $4.6M in liabilities, which may include FTX-related obligations.
Conflict of interest: Leopold Aschenbrenner was on the FTX Future Fund team that committed $5M to Atlas, then became Atlas's president. Regardless of intent, this is a funder-to-grantee leadership pipeline that raises governance questions.
What Others Say
Core criticism (EA Forum): "Why do 100 high-schoolers need $50K each from Open Philanthropy?" The argument: SPARC identifies similar talent for free, $50K scholarships are wasteful, the money could be better deployed. The program's subsequent reduction to $10K partially validates this critique.
Gaming concern (CitizenTen, EA Forum): "The Vultures Are Circling" reported Discord servers organized around gaming Atlas admissions. "Having a reward larger than most adults' yearly income removes hesitation around dishonesty." The $50K scholarship incentivized applicants to simulate rationalist values rather than hold them.
Counterfactuality question: If Atlas selects from the global top 0.01% of curious, ambitious teenagers, would many of them find AI safety and ambitious projects anyway? The program's marginal contribution vs. selection effect is unknowable but central to evaluating its theory of change.
In Atlas's defense: Jonas Vollmer's willingness to link to an ambiguous Manifold market in his closure announcement is well above the transparency norm. The founders' subsequent moves to METR, CSET, and AI Futures Project validate their stated priorities. The program was genuinely innovative in its application design and curriculum.
What's Absent
- No comprehensive outcomes tracking beyond one Manifold market. For $13.8M in talent pipeline spending, the absence of systematic impact evaluation is the most significant gap.
- No post-mortem or detailed reflection from founders on what worked, what didn't, and lessons for future field-building programs.
- FTX clawback resolution undisclosed.
- No public plan for the residual entity's $2M+ in assets.
- No public curriculum materials despite claiming pedagogical innovation.
- No explanation for why Leopold Aschenbrenner specifically became president given the FTX conflict.
- No formal comparison against SPARC/ESPR outcomes.
Recommended Reading
- Jonas Vollmer's wind-down letter (https://www.atlasfellowship.org) -- The most candid primary source. Acknowledges partial failure, links to prediction market on outcomes. Start here.
- "[Atlas Fellowship] Why do 100 high-schoolers need $50K each from Open Philanthropy?" (https://forum.effectivealtruism.org/posts/fMrtoKBFK7p6oRHpu/) -- The core cost-effectiveness critique, well-argued.
- "The Vultures Are Circling" by CitizenTen (https://forum.effectivealtruism.org/posts/W8ii8DyTa5jn8By7H/) -- Early warning about gaming dynamics around large scholarships.
- Manifold market on Atlas Fellow outcomes (https://manifold.markets/JonasVollmer/will-20-atlas-fellows-be-actively-w) -- Jonas's own accountability mechanism, resolved at 50%.
- ProPublica 990 data (https://projects.propublica.org/nonprofits/organizations/881678673) -- Three years of financials showing the transition from active program to holding entity.