Scott Alexander / Astral Codex Ten

Stated Theory of Change

Scott Alexander does not formally state a theory of change, but it can be reconstructed: accessible, rigorous, entertaining long-form writing about important topics draws intelligent people into an intellectual community where they update their views toward taking AI safety (and other cause areas) seriously, eventually working on these problems or funding them. The blog is the engine; the community infrastructure (meetups, contests, grants) is the flywheel.

His most direct statement of the case for AI safety work is from 2015: "If humanity creates human-level AI, technological progress will continue and eventually reach far-above-human-level AI. If far-above-human-level AI comes into existence, eventually it will so overpower humanity that our existence will depend on its goals being aligned with ours." He placed >95% confidence on these premises.

The implicit logic: if this is the most important problem, then anything that increases the number of intelligent people who take it seriously — including writing a popular blog — is among the highest-leverage activities available.

Revealed Theory of Change

Scott's actions are remarkably consistent with his stated views, which is itself notable in a landscape where many AI safety actors show significant gaps between words and deeds.

What he actually does: writes prolifically about AI risk (10+ years), runs a grantmaking program (~$3M distributed), maintains global community infrastructure (~180 meetup cities), collaborates with credible AI safety researchers (Kokotajlo, Lifland), and publicly defends the EA movement that supports much AI safety work.

Where actions add nuance to stated views: Scott's ACX Grants span well beyond AI safety — genetically modified corn, screwworm eradication, kidney donation policy, lead-acid battery recycling. This reveals either (a) genuine uncertainty about cause prioritization, or (b) a belief that building a broad intellectual community is itself the theory of change, not narrowly optimizing for AI safety output. Given his statements, (b) seems more likely.

One notable divergence: Scott expresses "complicated conflicts of interest" about funding AI safety directly and has funded relatively few explicitly AI safety projects through ACX Grants. This is surprising for someone with a ~33% p(doom) and suggests either genuine caution about his epistemic authority in technical areas, or a judgment that the AI safety funding landscape is already well-served by larger funders.

His recent shift from pure commentary to active participation (AI 2027, AI Futures Project blog) represents a meaningful evolution. He went from "I write about AI safety" to "I collaborate with serious researchers on concrete forecasts," which is a higher-risk, higher-reward position.

Key Assumptions

Assumption 1: Accessible writing about AI risk actually moves people into safety careers.

• Evidence for: SSC ranked 4th in EA entry survey (2017). ~100 GWWC pledges from one post. The Dwarkesh podcast on AI 2027 got 166K views. Multiple ACX grantees have gone on to work in AI safety.
• Evidence against: No systematic tracking of the pipeline. The 2017 survey is 9 years old and measured "discovering EA" not "doing AI safety work." The community may attract people who enjoy thinking about AI risk more than people who do something about it.
• Testable? Yes, with a survey of current AI safety workers asking what influenced them to enter the field.
• What changes if wrong? The theory of change collapses to "entertaining writing that has no downstream impact on AI risk."

Assumption 2: The rationalist epistemic methodology produces genuinely better thinking.

• Evidence for: Scott's COVID mask analysis (March 2020) was prescient. The rationalist community's early attention to AI risk has been largely vindicated by mainstream adoption of these concerns. Prediction market enthusiasm has been partially validated.
• Evidence against: The "Why Don't I Trust" critique argues Scott demonstrates motivated reasoning on sensitive topics while claiming Bayesian objectivity. The HBD email reveals strategic ambiguity rather than genuine uncertainty. The AI 2027 timeline model had significant methodological issues (Titotal critique). The community's demographic narrowness (~10% female) suggests something is systematically filtering who can benefit from the methodology.
• Testable? Partially — through prediction calibration, replication of claims, etc.
• What changes if wrong? If rationalist epistemics are a sophisticated form of confirmation bias, the pipeline may produce articulate AI safety advocates who share blind spots rather than genuinely well-calibrated thinkers.

Assumption 3: Community infrastructure (meetups, grants, contests) meaningfully amplifies the writing's impact.

• Evidence for: 180 cities with meetups is remarkable. ACX Grants have seeded at least one nationally significant AI safety org (Good Ancestors Australia). The Book Review Contest has cultivated new writers.
• Evidence against: No data on meetup attendance, retention, or downstream effects. Grant retrospectives show mixed results.
• Testable? Yes, with systematic meetup surveys and grant outcome tracking.
• What changes if wrong? Scott's impact reduces to his direct readership influence, which is still substantial but less unique.

Assumption 4: Financial independence preserves intellectual integrity.

• Evidence for: Scott has no AI lab funding, no institutional constraints, and a diversified income (Substack + psychiatry + investments). He can write whatever he wants.
• Evidence against: Subscriber incentives may subtly push toward engaging controversy and maintaining audience interest. The private email reveals that controversial topics generate "5x more hits," creating a financial incentive for engagement with incendiary material.
• Testable? Through analysis of post topic choices and engagement metrics.
• What changes if wrong? Financial independence becomes a fig leaf for incentive structures that reward controversy over truth-seeking.

Strengths

Unique reach and accessibility. No other writer in the AI safety ecosystem reaches a comparable audience with comparable intellectual depth. Scott bridges the gap between technical alignment researchers and the general public in a way that academic papers, podcasts, and Twitter threads do not. The blog is genuinely enjoyable to read, which is an underappreciated competitive advantage for a cause area that often sounds like science fiction to outsiders.

Financial independence. Scott's Substack income and investment gains make him genuinely independent. He is not beholden to any funder, employer, or institution. This is rare in the AI safety ecosystem, where most actors depend on a small number of funders (principally Open Philanthropy/Coefficient Giving).

Consistency over time. Scott has been writing about AI safety since 2015, consistently, at high volume, while maintaining readability. This is not a bandwagon effect — he was writing about these issues when they were deeply niche.

Intellectual honesty about uncertainty. His p(doom) of ~33% is neither the "nothing to see here" of AI optimists nor the "we're all going to die" of Yudkowsky-adjacent doomers. He models uncertainty in a way that is more credible to mainstream audiences than either extreme. On the Dwarkesh podcast, he is genuinely candid about what he does and does not know.

The Moloch framework. "Meditations on Moloch" is genuinely one of the most important intellectual contributions to the AI safety discourse. It provides a philosophical foundation for why alignment matters that goes beyond narrow technical arguments and connects to deep intuitions about civilizational dysfunction.

Weaknesses and Risks

The epistemic trustworthiness problem. The HBD email and the "Why Don't I Trust" critique raise genuine questions about whether Scott's apparent balance on sensitive topics is strategic rather than sincere. If the most prominent rationalist writer demonstrates the failure mode of motivated reasoning dressed in rationalist language, this poisons the well for the entire epistemic methodology he champions. This is not about whether his views on any particular topic are correct — it is about whether his readers can trust his method.

Demographic narrowness of the pipeline. If the community systematically excludes or marginalizes certain demographics (women, racial minorities), then the talent pipeline is leaving value on the table. AI safety needs diverse perspectives, and a pipeline that selects for a particular demographic profile may produce blind spots in the field.

No accountability structure. ~$3M in grants, enormous influence on a critical cause area, and zero formal governance. If Scott makes a bad judgment call — about whom to fund, what to write, or how to handle controversy — there is no mechanism to catch or correct it except public criticism and subscriber churn.

AI timeline views may be systematically biased toward drama. The Titotal critique reveals significant methodological problems with AI 2027's forecasting. Scott's engagement incentives (controversial topics -> more hits) may pull him toward more dramatic AI scenarios. Freddie deBoer's bet challenge represents a real risk: if AI progress is slower than predicted, the urgency narrative that drives the pipeline weakens.

EA entanglement. Scott is so identified with the EA movement that EA's reputational problems become his problems. Post-FTX, post-OpenAI boardroom drama, any future EA scandal will splash on ACX. This makes his influence fragile to events he cannot control.

Cross-References

Complementary to: MIRI (Yudkowsky's technical alignment work — Scott makes these ideas accessible), 80,000 Hours (career guidance — Scott provides the intellectual foundation), Open Philanthropy/Coefficient Giving (institutional funding — Scott is informal/individual funding), LessWrong (community platform — Scott is the most prominent individual contributor).

In tension with: PauseAI (Scott argues against quitting AI companies and is skeptical of coordinated pause strategies), academic AI ethics (different framing, different politics, different methodology).

Overlapping with: AI Futures Project (active collaboration via AI 2027), Manifund/Manifold Markets (grants infrastructure), Zvi Mowshowitz's Substack (similar audience, complementary analysis style).

Unique position: No other individual occupies Scott's exact niche — a genuinely popular writer who bridges rationalist philosophy, EA cause prioritization, AI safety argumentation, and practical grantmaking. The closest analogues (Zvi, Kelsey Piper at Vox) reach smaller audiences or operate within institutional constraints.

What Would Change This Assessment

• Evidence that the talent pipeline works. A systematic survey showing that a significant fraction of current AI safety workers cite SSC/ACX as a major influence would substantially upgrade this assessment.
• Evidence that the talent pipeline fails. Data showing that ACX readers mostly stay passive consumers, or that ACX-pipeline researchers systematically underperform on AI safety work, would substantially downgrade it.
• Scott addresses the HBD/epistemic trustworthiness critique directly. A candid, detailed response would either resolve the concern or deepen it, but either way would be more informative than the current silence.
• AI 2027 predictions prove accurate or inaccurate. If the scenario's 2025-2026 predictions track reality, Scott's credibility as a forecaster rises substantially. If they do not, the "research-backed" framing of his AI views weakens.
• Major expansion or formalization of ACX Grants. If the program develops formal governance, external evaluation, and systematic outcome tracking, the grantmaking theory of change becomes more credible.
• Continued demographic narrowness of the community. If ACX meetups and readership remain heavily skewed (white, male, tech-adjacent) despite the community's growth, this constrains the theory of change's ceiling.

Self-Critique

What's weakest: The assessment of the talent pipeline's effectiveness relies heavily on one 9-year-old survey data point and qualitative impressions. I cannot quantify the most important claim — how many people entered AI safety careers because of Scott's writing.

Potential bias: I may be giving too much weight to the HBD/epistemic trustworthiness critique because it is the most intellectually interesting criticism. Many readers may find it irrelevant to Scott's AI safety contributions.

What a thoughtful disagreer would say: "You're treating Scott as if he's an organization with formal obligations. He's a blogger. The fact that he writes well about AI safety, maintains a community, and gives away millions in grants is already far more than most people do. Demanding formal accountability and systematic measurement from an individual is unreasonable."

Single weakest claim: My suggestion that Scott's engagement incentives (more hits from controversy) may bias his AI timeline views toward drama. This is speculative and not well-evidenced.

What information would most change my view: Rigorous data on whether SSC/ACX readers actually enter AI safety careers at higher rates than comparable populations. If the pipeline demonstrably works, Scott's theory of change becomes one of the most cost-effective interventions in the entire AI safety landscape — one person's blog output generating hundreds of safety-focused careers.