Ben Horowitz: Why Open Source AI Will Determine America's Future

Horowitz explains how Biden administration's anti-open-source policy backfired spectacularly. While US companies went closed-source, China's DeepSeek became the dominant open-source model used by US companies and universities. The weights encode cultural values on everything from Tiananmen Square to free speech, giving China influence over global AI culture.

• •DeepSeek (Chinese) is now the dominant open-source model used by US companies and academia, not US models
• •AI model weights encode cultural values and interpretations of history, human rights, and political issues
• •Biden's Manhattan Project approach was delusional—Google and OpenAI employ many Chinese nationals, making secrets impossible
• •US strength is open society where everyone contributes, not top-down secrecy
• •OpenAI's recent open-source pivot lacks 'vibes'—open source is community-driven from inception

Drawing parallels to spreadsheets creating private equity and computers' gradual 40-year impact, Horowitz predicts AI will transform jobs unpredictably rather than eliminate them wholesale. He provides concrete examples like Hollywood using AI to generate 17 of 20 takes, improving economics without replacing positions.

• •AI cycle just started 4 years ago—expect 25-year arc similar to previous tech revolutions
• •Jobs from 1750s agriculture era would find today's roles (marketing exec, graphic design) incomprehensible
• •Spreadsheets enabled private equity by automating manual calculations—AI will have similar unexpected side effects
• •Hollywood example: shoot 3 takes, AI generates 17 more—improves economics without eliminating jobs
• •Writers already use AI for dialogue they don't want to write, enabling faster work rather than replacement

Horowitz dismantles arguments for regulating sentient AI and 'takeoff' scenarios, comparing them to regulating time travel based on physics thought experiments. He argues AI models are just math (200 billion variable equations doing matrix multiplication), and regulations should focus on applications like bomb-making instructions, not the models themselves.

• •AI 'takeoff' and sentience are thought experiments nobody knows how to build—like regulating time travel
• •Anesthesiologists know most about consciousness (how to turn it off)—we definitely can't build it
• •Regulations become 'you can do math, but not too much math'—absurd and competitively harmful
• •Bomb-making instructions are already illegal—AI shouldn't get a pass, regulate applications not models
• •Europe's precautionary principle would have prevented automobiles and nuclear power (safer than oil)
• •Copyright: Can't reproduce Michael Jackson, but training on copyrighted material for statistical models is like humans reading—China allows it, creating competitive disadvantage

Using autonomous vehicles as a benchmark (19 years from 2006 DARPA challenge to 2025 Waymo deployment), Horowitz explains why humanoid robots face harder problems. The critical bottleneck is lack of training data for 3D manipulation tasks and fat-tailed edge cases, plus the entire robot supply chain is China-based.

• •Autonomous cars took 19 years (2006-2025) despite easier 2D problem with existing map data
• •Robot manipulation is multi-dimensional—picking up glass vs shirt vs egg requires vastly different approaches
• •Fat-tailed human behavior (driving 75 in 25 zone) took years to solve—robots have zero training data
• •Video models can't move camera angles because they don't understand 3D space—not suitable for robots
• •Entire robot supply chain is China-based (one German company founded by Chinese nationals, bought by China)
• •US faces strategic vulnerability similar to rare earth minerals and chips if supply chain isn't developed

Horowitz positions crypto as the network layer for AI, analogous to how networks (LANs, Internet) paired with computers (PCs, smartphones) throughout history. AI agents need economic capabilities but can't get credit cards or bank accounts—crypto provides Internet-native money, identity verification, provenance tracking, and security architecture.

• •Computing evolution always has two pillars: computers and networks (mainframes/SNA, PCs/LANs, Internet/smartphones)
• •AI agents need to be economic actors but can't have credit cards or bank accounts (must be human)
• •Crypto is bearer instrument Internet-native money—AI banks already exist with KYC for AI agents
• •Crypto solves bot problem: proving you're human, not AI
• •Provenance solution: Is this deepfake or real? Registry of truth shouldn't be US government or Google—use game-theoretic blockchain
• •AI excels at breaking into systems and social engineering—current centralized architectures are 'honey pots'
• •Zero-knowledge proofs: prove creditworthiness without revealing bank account data

Horowitz traces how post-dot-com regulations (Sarbanes-Oxley, Reg FD, decimalization, order handling rules) made going public onerous and dangerous. Amazon went public at $300M in the 90s; now companies raise $30B privately. The result: venture firms must expand capabilities to replace investment banks' role.

• •Amazon IPO'd at $300M valuation, Netscape with $10M quarterly revenue—impossible today
• •Regulatory cascade post-dot-com: Sarbanes-Oxley, Reg FD, decimalization, order handling rules made public markets onerous
• •Cost of being public (lawyers, accounting, D&O insurance) prohibitive for small companies
• •Reg FD created asymmetry: short sellers gained power, companies lost ability to defend themselves
• •OpenAI can raise $30B privately—what's the value of going public? More lawsuits and quarterly earnings calls
• •Venture capital firms must now handle late-stage capabilities previously done by investment banks
• •AI valuations justified by product quality: OpenAI hit $10B revenue in 4 years (unprecedented), ChatGPT and Cursor 'work awesome'

Drawing from Bushido samurai philosophy, Horowitz argues culture is defined by daily behaviors under stress, not abstract values. He shares a16z's specific rules: $10/minute fine for being late to entrepreneur meetings (programs respect), instant firing for publicly criticizing startups on X (dream builders not killers).

• •Samurai Bushido principle: culture is set of actions, not beliefs—everyone claims integrity until tested
• •Virtues only matter under stress: honest until it costs your job or marriage
• •Don't define ideals, define exact daily behaviors—culture isn't quarterly review item
• •$10/minute fine for lateness to entrepreneur meetings (no exceptions)—programs that nothing is more important
• •Post criticism of entrepreneur ideas on X = instant firing, regardless of position
• •VC power dynamic (I have money, you need it) naturally creates disrespect—need strong behavioral rules to counter
• •Worst approach: off-site to discuss values and write flowery language—creates hypocrisy culture

Horowitz shares counterintuitive lesson: managers can develop people because they know the job, but CEOs hire CFOs, heads of HR, and marketing leaders for roles they've never done. Attempting to develop them wastes time and distracts from CEO-only responsibilities like setting direction and making critical decisions.

• •Managers can develop reports because they know how to do the job (engineering manager training engineers)
• •CEOs hire CFO, head of HR, head of marketing—roles they've never done and can't teach
• •Attempting to develop executives you can't improve wastes critical CEO time
• •CEO-only responsibilities: set company direction, articulate it, organize properly, ensure best people in place, make decisions only they can make
• •If CEO doesn't make critical decisions, entire company slows down—developing executives is distraction
• •Lesson titled 'The Sad Truth About Developing Executives'—used Lil Wayne quote: 'truth is hard to swallow and hard to say too'