Center for Applied Rationality (CFAR)

Stated Theory of Change

CFAR has cycled through three stated theories of change:

Original (2012-2018): Teach Bayesian/Sequences-derived rationality techniques through immersive workshops. Better thinkers would naturally gravitate toward the world's most important problems, including AI safety. CFAR would be the community hub where this pipeline operated.

Pivot (2018-2020): Explicitly focused on AI safety talent pipeline. Run programs jointly with MIRI (MSFP, AIRCS, CML). Measure impact by alumni entering AI safety careers.

Relaunch (2025+): A community of people geeking out about rationality, with no AI safety recruitment goal, no claimed EA payoff, and no deference to any outside authority. "If we can't make a financial go of things under our new ethos, my plan is not to revert to our past ethos, it's to fold."

Revealed Theory of Change

The gap between stated and revealed theory of change is where CFAR is most interesting:

2012-2020: The revealed theory was: Create an intense community environment around Eliezer Yudkowsky's intellectual framework. Attract bright, searching young people. Give them a peer group, a sense of purpose (AI risk), and social validation. Some would move to Berkeley and work at MIRI or EA orgs. This worked -- CFAR alumni are at Anthropic, OpenAI, DeepMind, CEA, 80K Hours, and dozens of other organizations. But the mechanism was community formation and social influence, not rationality techniques per se. The strongest predictors of impact were "moved to Bay Area" and "attended alumni events," not "learned Double Crux."

Salamon's own self-critique is devastating on this point: the techniques she championed worked more through persuasion than through truth-seeking. "Internal double crux" was elevated because it convinced people to take AI risk seriously, not because it improved their reasoning. The workshops optimized for "does this cause people to look like people who will help with AI risk" rather than genuine epistemic improvement.

2025+: The revealed theory is much harder to read. CFAR has 7 part-time staff, $129K in cash, a $4.9M lawsuit hanging over it, and no institutional funding. The honest revealed theory might be: Anna Salamon found something she lost -- a way to run workshops that feels honest to her -- and she is trying to see if that can sustain itself. This is less a theory of change and more a bet on personal integrity.

Key Assumptions

Assumption 1: Rationality can be taught in workshops.

• Evidence for: Workshop reviews are consistently positive. The 2015 longitudinal study showed effects on well-being and personality. Many alumni credit CFAR with career changes.
• Evidence against: No RCTs. Self-selection bias is severe (people who pay $3,900 to attend a rationality workshop are not a random sample). Salamon's own analysis suggests the curriculum development got "stuck" for reasons common to all self-help movements. No external validation after 13 years.
• Testable: Yes, but CFAR has not tested it rigorously.
• If wrong: The entire workshop model is social value plus community formation, not technique transfer. Not necessarily zero-value, but a different thing.

Assumption 2: CFAR's community pipeline produced counterfactual impact.

• Evidence for: 159 alumni with "increase in expected impact" by 2017. Multiple MIRI hires from MSFP. Alumni at major AI labs.
• Evidence against: The strongest predictors of impact were prior EA/rationality involvement and moving to Bay Area -- both of which could reflect selection effects rather than CFAR's causal impact. Most of these people were already on the path.
• If wrong: CFAR was a social hub that attracted already-interested people, not a converter of new people. It had community value but limited counterfactual value.

Assumption 3: aCFAR can sustain itself without AI safety or EA funding.

• Evidence for: Some workshop alumni value the experience highly. The relaunch has genuine enthusiasm from supporters.
• Evidence against: Workshop revenue has never covered costs. The old donor base was EA/AI-focused. Salamon is explicitly rejecting EA efficiency framing in her fundraiser. No institutional funder has stepped forward.
• If wrong: aCFAR folds within 1-2 years, or reverts to EA messaging to survive.

Assumption 4: The community safety problems are behind CFAR.

• Evidence for: Part-time instructors reduce power dynamics. No "power over" dynamics in 2025 workshop. Salamon's personal growth and honesty.
• Evidence against: No formal safety policies published. Minimal board oversight. The FTX lawsuit is unresolved. The Zizian situation demonstrated that community-level harms can emerge from dynamics CFAR doesn't directly control.
• If wrong: Another community safety crisis could be fatal for the relaunch.

Strengths

1. Intellectual honesty of leadership: Salamon's self-critiques are genuinely remarkable. Few org leaders would publicly analyze how their own fear of AI risk corrupted their organization's curriculum, or admit they "were creating conditions for a cult." This level of transparency is rare and valuable.

2. Genuine community value: Workshop reviews are consistently positive. People make friends, have meaningful conversations, and sometimes solve real problems. The social value of bringing analytically-minded people together for intense shared experiences is real, even if the techniques are not uniquely effective.

3. Historical pipeline contribution: CFAR genuinely helped build the AI safety field during a period when the community was tiny. Many current AI safety researchers, EA leaders, and AI lab employees passed through CFAR workshops. This historical impact is significant even if the counterfactual is debatable.

4. Lean operational model: The part-time instructor structure is financially efficient and psychologically healthier. CFAR is running on a budget that most nonprofits would consider impossibly small. This frugality means donations go far.

5. SPARC: The high school program continues to receive institutional funding and is broadly respected as a talent identification and development program.

Weaknesses and Risks

1. No rigorous evidence base after 13 years: A rationality organization that cannot demonstrate its workshops improve rationality is failing its own test. The absence of external validation is not just an academic concern -- it means CFAR cannot distinguish between "our techniques work" and "people enjoy our workshops."

2. Financial existential threat: $129K in cash, a $4.9M lawsuit, workshops that lose money, and a deliberate rejection of the org's historical funding base. The financial position is precarious.

3. Reputational damage: The Zizian deaths, Brent Dill scandal, FTX connections, and Guardian investigations have permanently altered how CFAR is perceived outside the rationalist community. This may not affect current workshop demand but severely limits institutional funding and mainstream credibility.

4. No clear theory of how rationality workshops reduce AI risk in 2026: The original pipeline theory is obsolete -- MATS, BlueDot, ARENA, and other programs now serve this function with more structure and clearer impact models. If aCFAR is not about AI safety, what is its theory of marginal impact?

5. Leadership concentration risk: Everything depends on Anna Salamon. No succession plan, no other organizational leaders, no institutional memory beyond her.

Cross-References

• MIRI: CFAR was created to support MIRI's mission and shared office space, personnel, and worldview. The relationship is foundational but its current status is undiscussed. MIRI's own dramatic pivots (abandoning agent foundations, questioning whether alignment research helps) create uncertainty about the intellectual framework CFAR was built on.

• MATS/BlueDot/ARENA: These modern field-building programs now serve the AI safety talent pipeline function CFAR pioneered. They have more structured curricula, clearer mentorship models, and stronger institutional support. aCFAR is not competing in this space -- it has explicitly opted out of the AI safety recruitment function.

• Lightcone Infrastructure: Shared DNA (both emerged from rationalist community), former fiscal sponsorship, and overlapping networks. Lightcone runs LessWrong, Lighthaven, and AI Lab Watch. The separation (2024) was necessary for both organizations.

• SFF/BERI: Andrew Critch cofounded CFAR, BERI, and SFF. This creates a tight funding circle where one person's organizations fund each other. Not unusual in the AI safety ecosystem but notable for governance analysis.

• Leverage Research: CFAR and Leverage represent parallel experiments in applying rationalist principles to organization-building, both of which failed in ways their founders now discuss publicly. The Asterisk article treats them as sharing common failure modes despite surface differences.

What Would Change This Assessment

• Positive: (1) A rigorous external evaluation showing workshops produce measurable cognitive improvements. (2) A sustainable funding model that doesn't depend on EA/AI donors or Salamon's personal network. (3) aCFAR alumni producing clearly valuable work in domains unrelated to AI safety, demonstrating the "rationality, not recruitment" theory of change. (4) Resolution of the FTX lawsuit without financial devastation.

• Negative: (1) Another community safety incident. (2) aCFAR folding within a year due to financial failure. (3) Evidence that the relaunch quietly reverts to the AI safety recruitment model. (4) The FTX lawsuit resulting in an adverse judgment that bankrupts the organization.

Self-Critique

Weakest claim: My assessment that CFAR's workshops had limited counterfactual impact (vs. being a social hub for already-interested people) is based on the 2017 self-selection data and Salamon's own critique. But it's possible the workshops genuinely changed some people's trajectories in ways the data cannot capture.

Potential bias: I may be weighting the dramatic failures (Zizians, Brent Dill, mental health crises) too heavily relative to the mundane positive impact of hundreds of workshops where nothing went wrong. The availability heuristic would predict exactly this pattern. However, the failures are not random events -- they emerge from the same community dynamics that produced the positive outcomes.

What a thoughtful defender would say: CFAR, at its best, created a space where brilliant, weird people could be themselves and think deeply about hard problems. The world has very few such spaces. The harms were real but were caused by specific individuals exploiting general community dynamics, not by CFAR's workshop curriculum. And the new CFAR, with its part-time model and radical transparency, has genuinely addressed the worst failure modes.

Missing sources: I could not access the NYT "Happiness Code" profile (2016), Bloomberg abuse article (2023), TIME 30-person interview article (2023), or Harper's "Come with us if you want to live" (2015). These might provide important context. Duncan Sabien's specific objections to the relaunch are referenced but not fully documented.

Information that would most change my view: A well-designed external study showing CFAR workshops produce measurable improvements in reasoning ability (not just self-reported satisfaction). This would fundamentally alter the analysis of whether the techniques have value beyond community formation.