Sam Altman on Building OpenAI & Betting on the Impossible

• Altman killed Sora and Atlas browser despite being good products—in a world of limited compute and talent, you sacrifice good ideas for great ones (the platform over products)
• The first day of OpenAI: 12 people showed up at Greg Brockman's apartment, got a whiteboard, looked at each other and realized "we have no idea what to do next"—spent 4.5 years without launching a product, inverting all YC advice
• Even Altman can't break his own computer habits: he has Codex but still clicks around, pastes between apps, and manages email the old way—force of habit is stronger than technology
• He learned more from studying successes than failures as an investor—when something works, the lessons are specific and applicable; failures teach generic grit
• The two biggest AI risks aren't technical: loss of human control, and power concentration in a few companies/people—the "dear peasants, we'll give you cures and entertainment, just give up your autonomy" pitch is anti-human

Altman's core thesis is that OpenAI should function as a platform—one interface to powerful AI, one API for building on top—rather than trying to build every product category. This requires brutal focus: they killed Sora (video generation) and Atlas (their web browser) despite both being excellent products, because compute and talent are finite. The decision framework came from his investing background: power laws mean your best bet outperforms everything else combined, so kill the good to double down on the great.

The messy origin story reveals how different OpenAI was from typical startups. On day one in January 2016, a dozen people gathered in Greg Brockman's apartment, got a whiteboard, and quickly realized nobody knew what to do next. They spent four and a half years without launching a product—the opposite of YC's "ship when you're embarrassed" advice. Managing research without customer feedback required inventing new systems: leaderboards for tracking progress, external demos for eminent researchers, learning what didn't work (fake deadlines). The breakthrough came from "chaotic stumbling" that eventually produced the research path to GPT.

On AI adoption, Altman admits being wrong about timelines. He expected massive disruption after GPT-4 in 2023, but the economy has "so much inertia"—people keep doing things the same way. His own behavior proves it: despite building Codex, he still uses his computer the old way (clicking, pasting, managing email manually) because 20 years of habits are encoded as "what it means to work." The missing piece isn't technology but product design—we're in the "smartphones before iPhone" phase, straddling two worlds without the product ideas that make the transition seamless.

The intellectual influences are specific: Paul Graham for the "launch early, iterate" philosophy and YC's emphasis on technical founders making non-consensus bets. Peter Thiel for non-linear thinking—when Altman brought him a list of alternatives to ChatGPT two months after launch (when growth seemed unstable), Thiel said it was "obvious" to do nothing but double down on the empty text box, the first thing since Google with that kind of flexible power. From his investing years, Altman learned that lessons from success are more valuable than lessons from failure—failures teach generic grit, successes teach specific applicable patterns.

On AI risks, he's most worried about two things in tension: loss of control (AI becoming too powerful to guarantee human control) and power concentration (one company/model/person with too much power). Both are anti-human. He rejects the "dear peasants" sales pitch: trading human autonomy and agency for cures and cheap stuff. Instead, AI should expand human agency, enable small business formation, and keep people deeply in control of the future. Iterative deployment—putting models in the world, learning from real use, studying failures like the FAA—is how you make AI safe, not disconnecting from reality in a lab.