Ben Horowitz: RSI, Crypto as AI Money, & Classified Physics

What It Covers

Ben Horowitz of a16z joins Peter Diamandis' Moonshots podcast to argue that recursive self-improvement in AI has already begun, crypto is the natural currency for AI agents, US regulatory overreach poses a greater threat than AI itself, and Apple holds an underutilized hardware advantage that could redefine its position in the AI era.

Key Questions Answered

• •Recursive Self-Improvement Timeline: RSI is not a future event — it is already underway. Every Frontier Lab currently uses its own models to develop next-generation models, which is the functional definition of recursive self-improvement. The distinction between human-in-the-loop and fully autonomous RSI is blurring rapidly, as engineers increasingly rubber-stamp AI decisions rather than genuinely directing them. Expect 2026 to reflect compounding acceleration already in motion, not a discrete future trigger.
• •AI Regulation = Regulating Math: Horowitz directly told Biden administration officials that restricting AI models is equivalent to outlawing mathematics. Their response cited the 1940s classification of nuclear physics — some of which remains classified today — as precedent. Horowitz argues this approach failed then (the USSR replicated the atomic bomb trigger mechanism exactly) and would fail again, while handing China decisive influence over how AI reshapes global society.
• •Crypto as AI-Native Money: AI agents cannot open bank accounts, obtain credit cards, or hold fiat currency without human Social Security numbers. Crypto, being Internet-native, borderless, and permissionless, is the only viable financial infrastructure for autonomous AI economic actors. Horowitz predicts a new category of AI-focused crypto banks will emerge, and that stable coin legalization in the US significantly accelerates this transition. Crypto and AI form a compounding economic system, not parallel trends.
• •Apple's $1T+ Hardware Opportunity: Mac Mini and Mac Studio units are selling out with two-month wait times because their unified memory architecture — combining CPU and GPU RAM into a single pool — allows users to run large open-source models like OpenClaw locally. Horowitz states that if Apple formally adopted a strategy of owning local AI hardware and agent hosting, it would represent the single best product strategy available to the company, leveraging infrastructure already built without requiring new foundational R&D.
• •US AI Chip Export Controls as Structural Risk: The Biden administration's final executive order required US government approval before selling a single GPU to most of the world. Horowitz frames this not as a pause on AI globally, but as a mechanism that slows US progress enough for China to lead AI's societal reshaping. With 150,000 people dying daily worldwide, he argues that delaying AI development carries a concrete human cost that regulators consistently fail to weigh against theoretical risks.
• •AI Scientific Discovery Horizon: Horowitz and co-hosts predict AI will independently produce a discovery equivalent in significance to relativity within approximately two years. AlphaFold-style breakthroughs in structural biology are cited as early evidence that AI can collapse entire scientific disciplines overnight. Portfolio company Physical Superintelligence is explicitly working on this problem. The practical implication: companies and investors should position now for AI that does not assist scientists but replaces entire research verticals autonomously.
• •Labor vs. Capital Shift Accelerating: Since 2019, average wages grew 3% while corporate profits rose 43%. Nvidia is now 20x more valuable and 5x more profitable than IBM was in the 1980s, with one-tenth the staff. Horowitz advises new graduates to orient toward directing AI agents entrepreneurially rather than competing as labor. Funding rounds of $500M at $4B valuations are now accessible to two- or three-person technical teams, a scenario that was structurally impossible before 2023.