Quantified Uncertainty Research Institute (QURI)

Stated Theory of Change

QURI's stated path to reducing AI risk is indirect and meta-level. The causal chain:

1. Build tools for probabilistic estimation at scale (Squiggle, Squiggle Hub, Squiggle AI)
2. These tools enable "super duper forecasters" -- small, hedge-fund-like teams -- to cheaply estimate the value of everything
3. Better estimates lead to better coordination among EA orgs, governments, and other decision-makers
4. Better coordination reduces the risk of catastrophic outcomes, including from AI

The recent "Epistemic AI" pivot adds a parallel chain:

1. Use LLMs to automate evaluation and estimation work (RoastMyPost, opinion fuzzing, AI question resolution)
2. "AI Intellectuals" gradually become trusted for strategic analysis
3. Better AI-powered epistemics reduce the risk of epistemic lock-in (where different groups use different AI reasoning systems that reinforce existing beliefs)
4. Correct beliefs win out, leading to better decisions about AI governance

The mechanism connecting these tools to AI risk reduction specifically is diffuse. Ozzie's ~15% P(AI net bad) and support for faster AI development mean he is less focused on preventing AI catastrophe than on building general epistemic infrastructure that could help with many problems, including AI.

Revealed Theory of Change

What QURI actually does, judged by allocation of time and resources:

Primary activity: Build and iterate on Squiggle (the programming language). This has consumed most of QURI's development effort since 2020. Five years, multiple developers (Slava Matyuhin, Sam Nolan, Quinn Dougherty), and the tool remains below v1.0 with ~30 documented external users.

Secondary activity: Publish blog posts and proposals about how estimation, forecasting, and epistemic AI could work. These are conceptual/visionary rather than empirical. They propose frameworks (scorable functions, opinion fuzzing, estimation consent, AI question resolution) that require adoption by platforms QURI does not control.

Tertiary activity: Build adjacent tools (RoastMyPost, Longterm Wiki, Metaforecast) that demonstrate the concept but haven't achieved wide adoption.

Notable divergence from stated theory: The stated theory requires that people USE the tools and that usage CHANGES DECISIONS. Neither condition has been demonstrated. The revealed theory of change is closer to: "build intellectually interesting tools and publish thoughtful proposals, hoping the ideas propagate." This is a research agenda, not an impact pathway.

Key Assumptions

Assumption 1: Coordination failures are primarily estimation failures.

• Evidence for: Some coordination failures clearly are (e.g., voters lacking information about policy impacts). Forecasting literature shows systematic overconfidence in expert judgment.
• Evidence against: Many coordination failures are driven by power, incentives, and identity, not information gaps. Climate change is well-estimated but poorly addressed. Ozzie's own example -- FTX -- was a fraud, not an estimation error.
• Testable? Partially. If QURI's tools were widely adopted and decisions didn't improve, this would be falsified. But adoption hasn't happened, so the assumption remains untested.
• What changes if wrong: The entire theory of change collapses. If better estimation doesn't lead to better coordination, QURI's tools are intellectually interesting but practically irrelevant.

Assumption 2: People will adopt estimation tools if the tools are good enough.

• Evidence for: Spreadsheets are ubiquitous. Some EA orgs use Guesstimate. LLMs are being rapidly adopted for various tasks.
• Evidence against: Seven years of QURI products with minimal adoption. Squiggle competitions attracted 3 entries from ~1K views. Squiggle AI had 30 external users. Professional forecasters still prefer simpler formats. Ozzie himself acknowledges "few people care about high-quality intellectual work."
• Testable? Yes, and the current evidence is negative. The LLM-based pivot (Squiggle AI, RoastMyPost) is a bet that AI can do the work humans won't.
• What changes if wrong: The shift to LLM-based tools is the backup plan. If people won't use manual estimation tools, maybe AI can do the estimation for them. This is a reasonable pivot.

Assumption 3: Epistemic AI is neglected, important, and tractable.

• Evidence for: Metaculus AI forecasting competitions show LLM ensembles can match human forecasters. Ozzie's case that strategic thinking is automatable is plausible given coding/medical progress.
• Evidence against: Major AI labs (Anthropic, OpenAI, Google) are all working on truthful and helpful AI. Perplexity, Elicit, and many startups are building AI research tools. The "neglected" claim is weakened by this activity.
• Testable? Yes, by whether QURI's specific tools outperform or complement general-purpose AI assistants.
• What changes if wrong: If Epistemic AI is not neglected (i.e., big labs do it better), QURI's niche disappears.

Strengths

Intellectual honesty. Ozzie is remarkably transparent about QURI's weaknesses, his own constraints (can't criticize funders), and the limitations of his tools. His "Downsides of Small Organizations" post is a genuine self-critique that few org leaders would publish.

Long-term vision coherence. From Guesstimate (2016) through Squiggle (2020) to Epistemic AI (2024), the core thesis has been consistent: better estimation at scale matters. The specific tools change, but the direction is stable.

Early positioning in Epistemic AI. If LLMs do enable cheap, reliable evaluation and estimation, QURI has spent years thinking about the frameworks, interfaces, and trust mechanisms this requires. The conceptual work (scorable functions, opinion fuzzing, epistemic selection protocols) could become relevant.

The Longterm Wiki. 736 pages tracking 527 organizations and 696 people in the AI safety ecosystem is genuinely useful infrastructure, regardless of QURI's other tools.

Low cost. At an estimated $150K-$250K/year, QURI is extremely cheap compared to most AI safety orgs. Even modest contributions to the ecosystem (Shallow Review, Longterm Wiki, Metaforecast) may justify this cost.

Weaknesses and Risks

Seven years without demonstrated impact. The most serious weakness. No documented instance of any QURI tool changing a decision. No usage metrics published. No external evaluation. When adoption data does exist (30 users, 3 competition entries), it's discouraging.

Single point of failure. QURI is Ozzie. If he loses motivation, gets sick, or pivots to something else, QURI ends. No succession plan, no institutional resilience, no second person who deeply understands the codebase and vision.

Chronic financial fragility. ~79% funding from one source (Jaan Tallinn via SFF). Zero from Open Philanthropy. No earned revenue. Ozzie describes funding as a "highly significant bottleneck." An org that has been fundraising for 7 years without securing stable multi-year funding may never.

The adoption gap. QURI builds tools for a problem (poor estimation) that decision-makers don't perceive they have. Forecasting platforms exist, prediction markets exist, LLM assistants exist -- and all of these have more users than Squiggle. The question "why will someone use this?" has not been convincingly answered.

Funder-worldview mismatch. QURI's primary funder (SFF/Jaan Tallinn) presumably funds AI safety work because they believe AI risk is severe. Ozzie puts P(AI net bad) at ~15% and supports faster AI development. This is a tension that could lead to funding withdrawal if funders realize the mismatch.

Project proliferation without completion. Guesstimate, Squiggle, Squiggle Hub, Squiggle AI, Metaforecast, relative values, RoastMyPost, opinion fuzzing, Longterm Wiki, scorable functions, evaluation consent, Shallow Review collaboration -- for a 1-2 person org over 7 years, this is too many projects. Many remain half-built or in maintenance mode.

Cross-References

Metaculus -- The closest comparison in forecasting infrastructure. Metaculus has actual users (thousands), prediction tournaments, institutional clients, and demonstrated forecasting accuracy. QURI's Squiggle is theoretically more expressive (functions vs. point estimates) but has orders of magnitude fewer users. QURI's scorable functions proposal would need a platform like Metaculus to adopt it, and Metaculus has not.

Rethink Priorities -- QURI's fiscal sponsor. RP is a much larger research org (~40 staff) that does quantitative cause prioritization research. The fiscal sponsorship relationship suggests RP sees some value in QURI's work, but the details are opaque.

MIRI -- Opposite end of the AI safety theory-of-change spectrum. MIRI works on fundamental alignment theory; QURI works on meta-level epistemic infrastructure. MIRI has also struggled with demonstrated impact, but has published influential technical work. QURI has not produced work of comparable influence.

Elicit/Ought -- AI-powered research tools with similar "better epistemics" goals. Elicit has actual users and VC funding. QURI's approach is more research-oriented and less product-oriented.

What Would Change This Assessment

• Usage metrics showing significant adoption. If Squiggle Hub had 1,000+ active users, or RoastMyPost was routinely used by EA Forum writers, the "no adoption" critique dissolves.
• A documented case where QURI tools changed a real decision. One compelling case study -- a funder using Squiggle models to allocate grants, an org using relative values to prioritize -- would dramatically strengthen the theory of change.
• Stable multi-year funding from a second major source. If Open Philanthropy funded QURI, or if a new funder provided multi-year support, the financial fragility concern diminishes.
• Hiring a second full-time person who stays. This would indicate the org is growing, not just sustaining one person's research agenda.
• External validation of Epistemic AI thesis. If a major AI lab or research group began citing QURI's frameworks, or if RoastMyPost evaluations became a standard in EA writing, the "too niche" concern weakens.

Self-Critique

What sources should I have checked but didn't:

• Rethink Priorities' 990s, which might reveal QURI's actual operating budget through the fiscal sponsorship.
• The Doom Debates podcast full transcript (April 2025), which would show Ozzie's evolved AI safety views.
• Ozzie's EA Forum shortform posts (quick takes), which contain extensive writing about AI policy and P(doom).
• Squiggle Discord member count and activity level.

Where this analysis is potentially biased:

• I may be too harsh on QURI because the adoption metrics are so low. It's possible that the tools are highly valuable to a small number of users in ways not captured by public metrics.
• I weight "demonstrated impact" heavily. QURI could argue that intellectual infrastructure takes decades to pay off, and 7 years is too early to judge.
• My framework favors legible, measurable impact. QURI's influence on how EA thinks about estimation and forecasting may be real but unmeasurable.

What would a thoughtful person who disagrees say: "You're judging a research institute by product metrics. Squiggle doesn't need a million users to be valuable -- it needs to enable a few key models that change how EA allocates resources. The Longterm Wiki, the Shallow Review, the conceptual frameworks -- these shape how the community thinks. That's QURI's real theory of change, and it's working."

Single weakest claim: My claim that Open Philanthropy's absence is a strong negative signal. OP's grant decisions reflect their priorities and capacity, not necessarily quality judgments about every applicant. Some excellent orgs simply haven't applied or don't fit OP's current categories.

What information would most change my view: Evidence that Squiggle models have been used in actual EA funding decisions, or that the Longterm Wiki is routinely consulted by major funders. This would show that QURI's tools matter even with low user counts -- quality of users matters more than quantity.