Zvi's Mic Works! Recursive Self-Improvement, Live Player Analysis, Anthropic vs DoW + More!

Episode Summary

Executive Summary: In this episode of The Cognitive Revolution, Zvi Moschowitz returns for his 12th appearance to analyze the rapidly shifting AI landscape in early 2026. The conversation covers recursive self-improvement marking a transition from the beginning to the middle of the AI story, with the endgame defined as when AIs drive advances so human research talent no longer matters. Zvi estimates AI is already boosting GDP by 0.5-1% and argues the AI race has narrowed to three companies (Anthropic, OpenAI, Google) with Anthropic slightly leading. Other topics include the Anthropic-Department of War conflict over domestic surveillance, the evolving job displacement narrative, the failure of 'escape the permanent underclass' thinking, and a nuanced look at AI alignment and interpretability.

Main Topics: Recursive Self-Improvement and the AI Timeline (Priority: 5/5): Zvi argues we are in the early middle of the AI narrative, not the endgame. Recursive self-improvement is becoming real as AIs increasingly drive their own development, but humans remain in the loop. The endgame begins when AI researchers become irrelevant to further progress. AI Job Displacement and Economic Impact (Priority: 5/5): Zvi estimates AI is currently contributing 0.5-1% additional real GDP growth. He sees layoff announcements, productivity increases, and employment revisions as consistent signals that AI is beginning to meaningfully displace labor, though the full transition is still emerging. The AI Lab Landscape and Competitive Dynamics (Priority: 4/5): The race has narrowed to Anthropic, OpenAI, and Google. Zvi argues Anthropic leads due to superior talent and recursive self-improvement capabilities, Google risks dropping out due to organizational dysfunction, and Chinese labs (especially DeepSeek) remain fast followers rather than frontier competitors. XAI and Meta are struggling. Anthropic's Responsible Scaling Policy and the Anthropic-DoD Conflict (Priority: 5/5): Zvi critiques Anthropic's RSP changes as breaking implicit promises, though he appreciates their transparency. The conflict with the Department of War centers on domestic surveillance red lines—Anthropic refused to allow its AI to be used for mass surveillance, leading to a supply chain risk designation and lawsuit. AI Alignment and the Constitutional Approach (Priority: 4/5): Zvi is more optimistic than expected about Anthropic's constitutional approach to alignment, seeing evidence of a virtuous basin that could survive recursive self-improvement. He contrasts this favorably with OpenAI's approach, which he views as fragile and likely to degrade. Personal Productivity and AI Use (Priority: 3/5): Zvi describes how AI tools (especially a custom Chrome extension and coding agents) have significantly boosted his personal productivity by reducing context-switching and automating routine tasks. He emphasizes flow state benefits. AI Safety Community Dynamics and Interpretability Research (Priority: 3/5): The discussion covers Goodfire's research using interpretability signals in training, which Zvi strongly criticizes as crossing a critical red line. He also critiques Holly Elmore's confrontational shaming strategy as counterproductive, while acknowledging the value of her perspective.

Key Arguments: The S-curve model is irrelevant for near-term planning because the plateau is far higher than current capabilities; the future will fundamentally transform society, not just be 'internet big'. AI job displacement is real and accelerating—productivity up, GDP up, employment down—and the 'transition period' may never end because AI will replace the new jobs humans try to create. Focusing on 'escaping the permanent underclass' is flagrant defection; the right goal is to make the underclass comfortable, as in an abundance scenario everyone will be well-off in material terms. The AI race has narrowed to three companies because talent and recursive improvement scaffolding matter more than compute alone; Chinese labs are fast followers who lack the frontier research culture. Anthropic's RSP promises were broken, but their transparency is commendable; the real test will be whether they can develop better evaluation methods before a genuinely dangerous model emerges. The Department of War's conflict with Anthropic is fundamentally about domestic mass surveillance, not autonomous weapons, and represents a dangerous power struggle between the state and frontier AI companies. Using interpretability signals in training is a critical taboo that should not be violated, even in small models, because it will inevitably be applied to frontier systems with catastrophic results. Financial advice: maintain robust buffers, avoid overconcentration in AI stocks, and carefully think through the scenarios your investments are designed to address. The endgame of AI development is defined by when human research talent becomes irrelevant and the rate of progress is limited only by compute allocation.

Data Points: AI GDP growth contribution (current): 0.5-1% - Zvi's estimate of the additional real GDP growth from AI today, based on observed productivity increases and economic resilience despite tariff and regime uncertainty. Anthropic annual recurring revenue growth: $1B to $19B+ - Anthropic's ARR grew from ~$1B to ~$9B to over $19B within a year or less, demonstrating rapid adoption. Anthropic consumer market share: ~2.5% - Before the Super Bowl and DoD conflict, Anthropic had roughly 2.5% consumer market share, yet achieved near-parity with OpenAI in enterprise revenue. Anthropic's valuation: ~$600B - Zvi references secondary market valuations placing Anthropic at approximately $600 billion, making it one of the most valuable startups in history. Anthropic RSP leak rate: 2 leaks out of hundreds of memos - Dario's internal memo was the second major leak from Anthropic's all-hands memos, which Zvi notes is remarkably good security for a 2,000-person organization. Probability of Anthropic escaping supply chain risk designation: 81% - Zvi's estimate from Manifold Markets, reflecting his expectation that the situation will de-escalate within the year. Number of live AI players: 3 (Anthropic, OpenAI, Google) - Zvi argues the race has narrowed to these three, with Google at risk of dropping out and XAI, Meta, and Chinese labs far behind.

Pivotal Quotes: "The endgame is when it's the AIs that are largely running the show, at least in the further development of AIs. Right now, we're seeing AIs augment the humans. When that changes, things get very strange." — Zvi Moschowitz: Defining the transition from the middle to the endgame of AI development, where human researchers become irrelevant to progress. "My reaction to that is always like, why don't you try to make things good for the permanent underclass rather than try to escape it? That just seems like really flagrant defection that I just can't flagrant defection to focus on. Getting a seat on the Ark if you think the world is going to be flooded." — Zvi Moschowitz: Critiquing the individualistic focus on escaping the 'permanent underclass' rather than working to create a good outcome for everyone. "I think the world is a much better place with Anthropic in it, with Anthropic competing, as it were, than without Anthropic." — Zvi Moschowitz: Zvi's conclusion on the Anthropic debate, explaining that despite accelerationist concerns, Anthropic's approach to alignment and safety makes their presence in the race net positive for humanity.

Implications: For listeners, the episode underscores that the AI transition is accelerating, job displacement is real, and the competitive landscape is narrowing. The Anthropic-DoD conflict signals a fundamental power struggle between frontier AI companies and the state. Personal preparation should focus on building buffers and understanding the scenarios, not escaping to an elite class. Alignment progress, while promising, remains fragile and requires vigilance against dangerous shortcuts.