Sam Altman on Building OpenAI & Betting on the Impossible

• Sam can't break his own 20-year computer habits despite having Codex—the empty text box and rote tasks feel like "what it means to work," revealing the deep psychological resistance to AI adoption even among builders
• Managing AI research follows the same power laws as startup investing: non-consensus bets, outlier talent, and exponential thinking—but required breaking YC's core advice by spending 4.5 years without shipping a product
• OpenAI's first day: 12 people in Greg Brockman's apartment staring at a whiteboard with no idea what to do, eventually finding their way through "chaotic stumbling" to the research path that led to GPT
• Peter Thiel's advice after ChatGPT launched: ignore Silicon Valley wisdom about feeds and network effects—the empty text box is the first thing since Google that works this way, so double down on it
• The two biggest AI risks aren't technical—they're loss of human control and centralized power, both fundamentally anti-human positions that trade autonomy for safety or convenience

Sam Altman reveals a striking contradiction: despite building Codex, he still clicks through emails, pastes between apps, and maintains to-do lists the same way he has for 20 years. He describes this as a psychological encoding where "doing this kind of stuff is what it means to work." This admission captures why AI adoption will be slower than technologists expect—human habits and institutional inertia create friction that actually makes the transition smoother. He compares it to watching people still go to Blockbuster after Netflix launched, a force of habit that tech nerds consistently underestimate.

The core of OpenAI's strategy emerged from startup investing principles applied to research. Sam learned that both domains follow power laws: your best bet outperforms everything else combined. This means backing non-consensus ideas and outlier talent—the researchers who think fundamentally differently, not those with a thin veneer on the same idea everyone else has. When OpenAI started in 2015, pursuing AGI was so non-consensus they got "hammered by intellectual giants of the field." But Sam understood from his investing background that high-risk bets are fine if they're super valuable when they work. The key difference from startups: they spent 4.5 years without shipping a product, completely counter to YC advice, forcing them to invent new ways to measure progress like leaderboards and external demos for eminent researchers.

OpenAI's platform strategy came from hard choices about killing good ideas for great ones. They killed Sora (video generation) and Atlas (web browser) to focus compute on general intelligence for knowledge work. Sam argues OpenAI should offer one direct interface to AI and one API—that's it. Not competing with customers, not subsuming the economy, but enabling 100 million new businesses. His two biggest concerns aren't technical: they're loss of human control (AI becoming too powerful) and centralized power (a few companies controlling access). Both are "fundamentally anti-human positions" that either hand control to AI models or trade human autonomy for safety. The right approach keeps people deeply in control and deeply empowered—people are the whole point.