Zvi Mowshowitz

Who Is Zvi Mowshowitz

Columbia math BA, son of two Columbia professors. Professional Magic: The Gathering player through the late 1990s-2000s: Pro Tour champion (Tokyo 2001), World Championship team member, Hall of Fame inductee (2007), ~$141K career winnings. Transitioned to quantitative trading at Jane Street Capital, professional gambling in Las Vegas, then a series of ventures: MetaMed (personalized medical research, Peter Thiel-backed, failed ~2015), InterPop (blockchain card game, wound down), and Balsa Research (policy reform nonprofit, founded 2022, now in advisory capacity). Entered rationalist community spaces around 2007 through the Hanson-Yudkowsky FOOM debates. Started writing about COVID in February 2020, which built his audience. Transitioned to weekly AI posts when COVID receded, growing to ~25,000 Substack subscribers and ~34,600 Twitter followers.

What He Does

Publishes approximately five posts per week on his Substack "Don't Worry About the Vase," anchored by a comprehensive weekly AI roundup (AI #161+ as of March 2026, typically 10,000-17,000 words each). These roundups cover new model releases, lab strategies, governance developments, policy analysis, and mundane AI utility, with editorial commentary throughout. Standalone posts include deep policy analyses (his SB 1047 guide was the most comprehensive public breakdown), lab safety plan reviews (Google, OpenAI, Anthropic), governance crisis reporting (OpenAI board crisis November 2023, Anthropic vs. Department of War February-March 2026), and ecosystem evaluations (Big Nonprofits Post, SFF recommender assessments).

Beyond writing, he serves as an SFF (Survival and Flourishing Fund) recommender -- a formal role in which he evaluates AI safety organizations and allocates millions of dollars from Jaan Tallinn's philanthropic capital across the ecosystem. He has served in at least three SFF rounds (2021, 2024, 2025). He sits on the CFAR board of directors. He makes approximately 10+ podcast appearances annually, primarily on the Cognitive Revolution with Nathan Labenz, plus 80,000 Hours, EconTalk, and others. He published in Vox (December 2023) arguing for AI safety regulation.

Self-identifies as a journalist and operates under specific confidentiality principles, including retroactive off-the-record and glomarization of confidential information.

How He Thinks

The analytical framework comes directly from quantitative trading: assess expected value under uncertainty, identify mispriced risks, note where consensus is wrong and why, size your response to the probability. In the Theo Jaffee interview: "When I was engaged in pretty explosive gambling... that is always very good for developing rationality." His approach to AI risk is recognizably a trader's: the precise p(doom) number matters less than the order of magnitude, and the bet sizing (how much effort to devote to safety) follows from the expected value calculation, not from a single point estimate.

He thinks by writing. Not outline-first, but discovery-through-prose: "I learn what I think as I write out my ideas in detail." This means his posts are not polished summaries of pre-formed views -- they are the thinking happening in public, including wrong turns and self-corrections.

Key original conceptual contributions: the Moral Mazes sequence (how institutional dynamics corrupt individuals), Simulacra Levels (four-level framework for understanding communication -- from literal truth through faction signaling to pure vibe manipulation), Slack (the vital importance of not being at binding constraints), and the concept of "The Way" (a recurring philosophical framework about correct action under uncertainty woven throughout his work).

His Theory of Change

Zvi's stated theory of change (from the 80K Hours podcast and multiple blog posts): "If I didn't think [changing the discourse and debate] was useful, I wouldn't be doing what I'm doing." The causal chain is: produce the most comprehensive, accurate, and honest analysis of AI developments available anywhere -> this analysis shapes how informed people (lab employees, policy workers, funders, community members) understand what is happening -> better understanding leads to better decisions on the margin.

More specifically, the mechanisms are:

1. Information aggregation: nobody else reads, processes, and synthesizes as much AI-related information as frequently. His weekly roundups are the "newspaper of record" for the AI safety-aware community.
2. Real-time governance analysis: when a major AI governance event happens (OpenAI board crisis, Anthropic vs. DoW, SB 1047), his analysis becomes the reference point. His OpenAI board crisis coverage was "widely linked as the best explanation of events."
3. Funding influence: as SFF recommender, he directly shapes millions in AI safety funding allocation. His public Big Nonprofits Post influences individual donor decisions.
4. Bridge communication: translates rationalist AI safety concerns for mainstream audiences (Vox, EconTalk) while maintaining credibility within the technical community.
5. Policy analysis: detailed public breakdowns of proposed regulations that become the go-to resource for anyone trying to understand what a bill actually does.

His core intellectual contribution to AI risk thinking is the "gradual disempowerment" framework: even if technical alignment of individual AI systems succeeds, competitive dynamics force humanity to cede increasing control to AI systems. "Those who resist these pressures will eventually be displaced by those who do not." This scenario requires neither AI misalignment nor coordination -- it is the default outcome of a world with powerful AI in a competitive multi-agent environment.

Money and Sustainability

Primary income: Anonymous donor support. Described as "generous, essentially unconditional" support to "be a public intellectual." Donors "never try to influence my decisions." Dollar amount unknown -- could be anywhere from modest ($50K-100K) to substantial ($200K+). This is the critical unknown.

Secondary income: Substack paid subscriptions (enabled but not gated -- all content is free; he explicitly says "I do not need the money"), Patreon (patron count unknown), consulting ($1,000/hour standard, $500/hour for positive-sum projects, 2-hour and 1-hour minimums respectively).

Balsa Research: Separate 501(c)(3) (EIN 92-0966448) with its own fundraising needs. 2025 target $200K (minimum $50K). Tiny team: 1 FTE, 1 part-time contractor, Zvi advisory. Not his primary activity or income source.

Incentive structure: The unconditional patronage model is specifically designed to avoid the incentive distortions he analyzes in the Moral Mazes sequence. He is not beholden to any funder, lab, or organization. He has no commercial product whose success depends on particular AI outcomes. His funding does not depend on making his work legible to grant committees. This independence is the core structural advantage of his model.

Sustainability risk: Entirely dependent on continued anonymous donor support and personal willingness to maintain the output pace. No institutional backup, no team, no succession plan. If he stops, the output stops.

Influence and Reach

Direct readership: ~25K Substack subscribers, ~34.6K Twitter followers, high LessWrong karma.

Podcast reach: regular guest on Cognitive Revolution (~10 appearances), plus single appearances on 80,000 Hours, EconTalk, Complex Systems, Clearer Thinking, and others. These podcasts reach thousands to tens of thousands per episode.

Mainstream media: published in Vox, cited in NYT and Telegraph. Tyler Cowen engages with his analysis on Marginal Revolution. Was consulted on AI 2027 scenario before publication.

Funding influence: SFF recommender allocating millions. Public Big Nonprofits Post shaping individual donor decisions.

Audience composition is unknown. Whether his 25K subscribers include policymakers, lab executives, and journalists -- or are primarily rationalist community members -- is the key uncertainty for evaluating his reach.

Ecosystem Position

Formal roles: SFF recommender (allocating AI safety funding), CFAR board member, Balsa Research founder (advisory).

Informal roles: "analyst of record" for major AI governance events, leading voice on AI policy analysis, de facto AI safety ecosystem evaluator through public writing.

He occupies an unusual position: deeply embedded in the rationalist community but not part of any AI safety organization, not an EA, not a researcher, not a lab employee. This independence allows him to criticize organizations that his peers work at and his funders support. He is, as he describes, a journalist covering the AI safety beat -- but one with SFF recommender authority and deep community ties.

Strengths and Distinctive Value

Volume and comprehensiveness: nobody else produces this breadth and depth of AI coverage at this frequency. The weekly roundup is the single best way to stay current on AI developments if you read nothing else.

Analytical rigor from quant background: the trading mindset -- probabilistic reasoning, expected value calculations, identifying mispriced risks -- produces analysis that is more precise and less vibes-based than most AI commentary.

Intellectual honesty: willing to say unpopular things. Calls working on frontier capabilities "insane." Disagrees with both Yudkowsky (too certain of doom) and accelerationists (not certain enough). Publishes his reasoning transparently, including uncertainty.

Independence: no institutional loyalties, no career advancement concerns, no publication bias. The unconditional patronage model funds honesty.

Real-time synthesis: when something happens in AI, Zvi's analysis is often the first comprehensive treatment. The OpenAI board crisis coverage was more accurate than mainstream media reporting.

Bridge function: can communicate with rationalist insiders, policy professionals, mainstream media audiences, and podcast listeners, adjusting register for each without sacrificing substance.

Limitations and Critiques

Interpersonal harshness: the Inkhaven episode reveals a pattern where his analytical intensity manifests as dismissiveness toward others. "He just seemed like he found me, personally, contemptible," reported one resident. Contrasted with Gwern (equally harsh but supportive) and Scott Alexander (exhaustive feedback). This may limit his ability to mentor, collaborate, or build institutions.

"Just a summarizer": the most pointed intellectual critique, from an Inkhaven resident: "He's just a summarizer now. His work isn't generative." His original conceptual contributions (Moral Mazes, Simulacra Levels) predate his AI focus. The weekly AI roundup is comprehensive synthesis, but synthesis of others' work. Whether this is a limitation depends on your theory of change -- if the bottleneck is information synthesis and dissemination, it is not a limitation at all.

No documented specific impacts: all attributed influence is diffuse. No case of "Zvi wrote X, then Y changed because of it." This makes cost-effectiveness analysis impossible.

Audience opacity: if his readers are primarily rationalist community members, the information loop is closed and his discourse-shaping theory of change is weaker.

Fragility: one person, no team, no succession. Unsustainable pace if health or motivation changes.

US-centric: limited coverage of international AI developments except as they affect US dynamics.

No technical research: influence operates through discourse, not through contributions to the technical alignment research agenda.

Recommended Reading

1. Theo Jaffee Interview (https://www.theojaffee.com/p/3-zvi-mowshowitz) -- The most candid single source. Full biographical interview covering MTG to Jane Street to rationalism to AI writing. Start here for understanding the person.

2. "The Risk of Gradual Disempowerment from AI" (https://thezvi.substack.com/p/the-risk-of-gradual-disempowerment) -- His most important intellectual contribution. The core risk model in his own words.

3. "Zvi Mowshowitz & Mentorship Anti-Patterns" by Vishal (https://vishalblog.substack.com/p/zvi-mowshowitz-and-mentorship-anti) -- The strongest criticism. Reveals limitations of the individual analyst model when applied to mentorship.

4. 80,000 Hours #184 (https://80000hours.org/podcast/episodes/zvi-mowshowitz-sleeper-agents-ai-updates/) -- 47K-word deep dive on AI safety worldview, career advice skepticism, and what interventions he thinks matter.

5. SFF Thoughts (https://thezvi.substack.com/p/zvis-thoughts-on-the-survival-and) -- Reveals how he evaluates AI safety organizations for funding allocation. Essential for understanding his ecosystem influence.