Scott Alexander / Astral Codex Ten

Theory of Change

Scott Alexander's theory of change is implicit rather than formally stated, but reconstructible from his writing and actions:

Write accessible, rigorous long-form analysis about important topics, especially AI risk
Build a large, engaged readership of intelligent people who might not otherwise encounter these ideas
Shift their priors toward taking AI safety seriously through the accumulated weight of well-argued posts
Provide community infrastructure (meetups, comments, contests) that turns isolated readers into a networked community
Fund promising projects across cause areas through ACX Grants, seeding new organizations and research programs

The quantitative evidence for this pipeline: Slate Star Codex ranked 4th in the 2017 EA entry survey for how people first heard about effective altruism. Scott's "Nobody Is Perfect" post led approximately 100 people to take the Giving What We Can pledge. ACX meetups now operate in ~180 cities worldwide.

His most famous articulation of why AI safety matters is "Meditations on Moloch" (2014), which frames misaligned AI as the ultimate coordination failure: "The implicit question is -- if everyone hates the current system, who perpetuates it? And Ginsberg answers: 'Moloch.'" The essay argues that aligned superintelligent AI could be the "gardener" that solves these failures, but misaligned AI would be the final, inescapable Moloch.

What They Do

Writing. Scott has published AI safety content since at least 2015. The blog spans 9,000+ PDF pages. Key AI safety posts include "No Time Like The Present For AI Safety Work" (2015), "AI As Profoundly Abnormal Technology" (2025, on AI Futures Project blog), and the AI 2027 scenario co-authored with Daniel Kokotajlo. His personal p(doom) is approximately 33%.

AI 2027. Co-wrote a month-by-month scenario of AI progress through 2028 with the AI Futures Project team (Daniel Kokotajlo, Eli Lifland, Thomas Larsen, Jonas Vollmer). The project went viral: almost 1 million website visitors, 166K Dwarkesh podcast views, talks at Harvard, the Federation of American Scientists, and OpenAI.

ACX Grants. Three rounds (2021, 2024, 2025), totaling approximately $3M distributed. The 2025 round gave $1.5M to 42 projects from 654 applications. AI-related grants include introspection benchmarks ($15K), LLM bias/truth-seeking research with Philip Tetlock ($50K), and Australian AI safety policy ($65K). Scott personally contributes $200-250K per round.

Community infrastructure. ACX Meetups Everywhere operates in ~180 cities worldwide, twice per year. Annual Book Review Contest cultivates new writers. ACX Survey (5,975 respondents in 2025) provides community demographics data.

EA defense. Scott is the most prominent public defender of effective altruism. Post-FTX, he published "In Continued Defense of Effective Altruism," arguing EA had saved approximately 200,000 lives through malaria prevention alone.

Key People

Scott Alexander Siskind. Born 1984. Psychiatrist (MD, University College Cork, 2012). Practices at Lorien Psychiatry in Oakland, CA, specializing in treatment-resistant depression. Former LessWrong writer (username "Yvain"). Extreme introvert; writes at 4:30 AM. The Dwarkesh podcast (2025) was his first-ever podcast appearance. On AI 2027: "I have always wanted to get more involved in the actual attempt to make AI go well. Right now, I just write about it."

Key collaborator: Daniel Kokotajlo. Executive Director of AI Futures Project. Resigned from OpenAI in 2024, rejecting a non-disparagement clause and risking millions in equity to speak out about AI safety. Co-author of AI 2027. Scott describes working with him as a turning point in his own AI views.

Intellectual network. Readers and defenders include Sam Altman ("essential reading among the people inventing the future"), Patrick Collison (Stripe CEO), Steven Pinker, Scott Aaronson, and Peter Singer. Debate partners include Tyler Cowen, Gary Marcus, Freddie deBoer, and Eliezer Yudkowsky.

Money and Incentives

Substack income. "Mid six figures" annually (plausibly $400-600K/year). Substack Pro deal in 2021; organic revenue exceeded the advance. Matt Yglesias, with a comparable deal, earned $775K/year. Scott subscribes to ~20K new unpaid readers/year but lost ~500 paid subscribers (as of 2024). Subscription: $10/month.

Grants budget source. ACX Grants personal contributions (~$200-250K per round) come from crypto and AI stock investment gains, not Substack income. "Some generous readers sent me crypto during the crypto boom... Some of the crypto went up. Then I reinvested it into AI stocks, and those went up too." He estimates ~5 more rounds are affordable.

Co-funders and institutional connections. ACX Grants co-funders include Craig Falls, Calvin French-Owen, Shauna Kravec, Anton Makiievskyi, Geoff Price, Adam Winkel, and anonymous donors. Retroactive prize funders include SFF, Long-Term Future Fund, Animal Welfare Fund, and EA Infrastructure Fund.

Financial independence. No evidence of direct financial ties to AI labs. Scott maintains a separate psychiatry practice. He is not employed by any AI company or funded by any AI lab. The AI Futures Project, where he contributes blog posts, is a separate nonprofit.

Potential conflicts. In 2025, ACX Grants experimented with SAFEs (equity stakes in startup grantees). Scott acknowledged "complicated conflicts of interest" regarding AI safety grants. Some evaluators are also grantees, with conflicts noted and those grants covered by outside funders. No formal conflict-of-interest policy exists for the program.

What Others Say

The strongest critique -- epistemic trustworthiness. "Why don't I trust Scott Alexander?" (muireall.space) argues that in a 2014 email, Scott stated "HBD [human biodiversity] is probably partially correct" and described a strategic approach to engaging with neoreactionaries, including that his blog gets "5x more hits" on controversial topics. The critic contrasts this with Scott's later public presentation as agonized and uncertain about the same questions. Key claim: "Alexander is by various means spreading parts of their thought that he agrees with while denying the extent of his agreement, where his motivation for advocating these ideas resides." Scott has not publicly responded to this critique.

The NYT controversy. In 2020, Scott deleted Slate Star Codex after learning a NYT reporter planned to publish his real name. 6,000+ petition signatures defended him. The resulting NYT article (Feb 2021) described the blog as "Silicon Valley's Safe Space" and noted that while SSC attracted "an unusually wide range of voices," the only people who "struggled to be heard" were "social justice warriors." The article quoted Elizabeth Sandifer: "The contrarian nature of these ideas makes them appealing to people who maybe don't think enough about the consequences."

AI 2027 critique. Titotal's detailed technical critique of the AI 2027 timeline models argues the "superexponential" curve has no empirical validation, produces mathematical infinities, and "the particular curve in their model does not match empirically with data." The critique also notes that a graph shared by Scott as "AI 2027's prediction" was "not the curve from any version of their actual model." The AI 2027 team acknowledged some criticisms and agreed to make changes.

Inclusivity critique. Ben Kuhn (2014) argued LessWrong's rationalist community drives away people from underrepresented groups, citing "rampant evo-bio-determinism" and a community that was ~10% female despite drawing from fields that are 20-30% female.

Defenders. Matt Yglesias: "the role of 'blogger who will present a contrarian take on a range of different kinds of subjects' has a lot of inherent value." Scott Aaronson: "It must have taken incredible guts for Scott to express his thoughts, misgivings and questions about some major ideological pillars of the modern world so openly." Sam Altman: described a specific essay as "an inflection point for Silicon Valley."

What's Absent

No formal accountability mechanism -- no board, no advisory committee, no external audit -- despite distributing ~$3M in grants and wielding substantial influence on the AI safety talent pipeline.

No systematic measurement of the theory of change's effectiveness. The 2017 EA entry survey (ranking 4th) is the best data, but it is 9 years old and measures "discovering EA" rather than entering AI safety careers.

No response to the most substantive epistemic critique. The "Why Don't I Trust" author emailed Scott about survey methodology issues and "never heard back."

No 80,000 Hours podcast appearance despite being a major influence on the AI safety talent pipeline. The Dwarkesh podcast (2025) was his first and only podcast.

No systematic cost-effectiveness analysis of the ACX Grants program, despite the community's emphasis on evidence-based giving.