Quantified Uncertainty Research Institute (QURI)

Theory of Change

QURI's theory of change, in Ozzie Gooen's own words: "Our bigger agenda is... can we make these very advanced estimation infrastructure systems and then apply them to many things?" The causal chain runs: build tools for probabilistic estimation at scale --> enable people to cheaply estimate the value of everything --> better estimates lead to better coordination --> reduced global failures including AI risk.

Ozzie articulated the end state as an "estimation utopia" -- a world where "a lot of key items are parameterized and estimated, and the estimates are, we have justified trust in those estimates, they're good, and people know that they're good." In this world, you could evaluate every sentence of a government bill, score every intellectual's track record, and render warfare unprofitable through coordinated public action.

This is a "mistake theory" worldview: coordination failures are primarily estimation failures, not power struggles. Ozzie: "If you imagine replacing the 8 billion people that we have with people who are a thousand IQ... do you really think that the world would be so messy?"

Since 2023-2024, the theory of change has evolved toward "Epistemic AI" -- using LLMs to automate the evaluation and estimation work that humans have proven reluctant to do with manual tools.

What They Do

QURI is primarily a software organization. Active projects (as of early 2026):

Squiggle -- A domain-specific programming language for probabilistic estimation, running on JavaScript. In development since 2020 (Ozzie's Guesstimate preceded it from 2016). Version 0.10.0 released January 2025. Open source, MIT licensed. 99 GitHub stars, 20 forks.

Squiggle Hub -- Platform for sharing Squiggle code. Launched August 2023. Squiggle AI adds LLM-powered model generation (Claude integration). Documented adoption: 30 external users, 168 workflows over an unspecified period.

RoastMyPost -- LLM-powered blog post evaluation tool, launched December 2025. Runs narrow automated checks (fact, spell, fallacy, math, link, forecast). Optimized for EA Forum and LessWrong content. Open source, free, explicitly experimental.

Longterm Wiki -- Strategic intelligence platform for AI safety. ~736 pages tracking 527 organizations and 696 people. Quality scoring system, causal diagrams, entity cross-linking.

Metaforecast -- Aggregates forecasts from multiple prediction platforms. Now in maintenance/archived mode.

Research output has been consistently conceptual: proposals for "scorable functions" (code-as-forecast), "opinion fuzzing" (ensemble LLM sampling), "evaluation consent policies," and AI-driven question resolution protocols. A collaboration with Arb Research produced the Shallow Review of Technical AI Safety 2025 (800+ papers/posts catalogued).

The 2022 Squiggle competitions offer the hardest adoption data: 3 long entries for a $1,000 prize, 5 entries (mostly short) for a $5,000 prize, from ~1,000-1,700 EA Forum views.

Key People

Ozzie Gooen -- Executive Director, sole full-time employee. Harvey Mudd engineering. EA since ~2008 (co-founded .impact 2013, which became Rethink Charity). Founded Guesstimate (2016), FHI Research Scholar, then QURI (2019). Core worldview: utilitarian, mistake theorist, estimation maximalist. P(AI net bad) ~15% -- unusually low for EA AI safety. Supports faster AI development.

Board (3 members, all unpaid): Andrew Critch (Secretary) -- CEO of Encultured AI, Research Scientist at CHAI, volunteers for SFF. Ben Goldhaber (Treasurer) -- Director at FAR Labs. Abigail Olvera -- Research Director at Golden Gate Institute for AI.

Team has shrunk over time. Key collaborators Nuno Sempere (Metaforecast builder) left for Sentinel. Eli Lifland moved to AI safety research. Slava Matyuhin listed as "previous contractor" despite releasing Squiggle 0.10.0 in January 2025.

Money and Incentives

Financial picture is deliberately incomplete. Since 2023, QURI operations flow through Rethink Priorities fiscal sponsorship. The 501(c)(3) holds assets but understates actual spending. Ozzie's 2024 compensation from the 501(c)(3) was $0, meaning he is likely paid through RP.

501(c)(3) financials (EIN 84-3847921):

Identified funding sources:

• SFF (Jaan Tallinn): ~$733K total (2020-2024). Comprises ~79% of all identified funding.
• Future Fund (FTX): ~$200K (2022, defunct).
• Long-Term Future Fund: Amount unknown.
• Open Philanthropy: $0. Zero grants in the org's entire history.
• Individual donors: Unknown amounts.
• Product revenue: $0. All tools are free.

Estimated actual operating budget: $150K-$250K/year (inference from 1 FTE + 1 contractor + API costs + RP overhead).

Key financial observations:

• Extreme funder concentration: one effective source (Jaan Tallinn/SFF) provides the vast majority of funding.
• Revenue is extremely lumpy, driven by irregular SFF grants.
• No path to financial sustainability. Zero earned revenue. No business model beyond grants.
• Ozzie in February 2025: "Funding has been a highly significant bottleneck recently."
• Open Philanthropy's complete absence across 7 years is a strong negative signal about how the largest EA funder views QURI's approach.

Governance overlap: Andrew Critch sits on QURI's 3-person board AND volunteers for SFF, QURI's primary funder. This is a textbook conflict of interest for a small nonprofit.

What Others Say

No external critical assessment of QURI exists. Despite extensive searching, no independent voice has published a substantive critique of QURI's approach. The org is too small and niche to attract critical attention.

Ozzie's self-criticism is the strongest available critique. In June 2023: QURI has "2 people," he feels "very isolated," and expects his "professional development to be slower." He explicitly identifies that EA has "gone too far along the axis of having lots of tiny organizations with poor ability for coordination" -- a description that applies directly to QURI itself.

On funding constraints and criticism: Ozzie acknowledges he cannot fully criticize "groups with much power over QURI" -- i.e., his funders. The person building evaluation tools is constrained from honest evaluation of his own funder network.

Broad forecasting skepticism: The general critique of QURI's space is that better forecasting does not demonstrably change decisions. Ozzie himself acknowledges "few people care about high-quality intellectual work" and forecasting organizations have "struggled to produce decision-relevant estimates." He notes that even 50% more accurate forecasts "would take time for many people to notice."

Adoption data as implicit criticism: 30 external Squiggle AI users. 3 competition entries from 1,000 views. 10 people staying for office hours at EA Global. These numbers are the community's revealed preference.

What's Absent

• No documented instance of any QURI tool changing a real-world decision in 7 years of operation.
• No public usage metrics for Squiggle Hub, Longterm Wiki, or RoastMyPost.
• No independent evaluation of QURI's effectiveness by any charity evaluator, funder report, or external researcher.
• No published plan for scaling beyond 1-2 people, despite Ozzie's own diagnosis of small-org problems.
• No peer-reviewed publications for a self-described "research institute."
• Actual operating budget obscured by RP fiscal sponsorship structure.

Recommended Reading

1. Ozzie Gooen on Mutual Understanding podcast (May 2023) -- 89 minutes of the founder speaking candidly about estimation utopia, justified trust, utilitarianism as power tool, FTX, and EA governance. The most unfiltered source for understanding how QURI's architect thinks. https://mutualunderstanding.substack.com/p/ozzie-gooen

2. "Downsides of Small Organizations in EA" (June 2023) -- Ozzie's self-aware diagnosis of exactly the structural problems his own org exemplifies. The strongest counterargument to QURI's organizational form, written by QURI's founder. https://quri.substack.com/p/downsides-of-small-organizations

3. "6 (Potential) Misconceptions about AI Intellectuals" (Feb 2025) -- The current strategic vision. Reveals Ozzie's P(AI net bad) ~15% and makes the case for "Epistemic AI" as a neglected cause area. https://quantifieduncertainty.org/posts/6-potential-misconceptions-about-ai-intellectuals/

4. Winners of the Squiggle Experimentation and 80,000 Hours Challenges (March 2023) -- Hard adoption data that says more about product-market fit than any vision statement. https://quantifieduncertainty.org/posts/winners-of-the-squiggle-experimentation/