Stripe CEO Patrick Collison asked Sam Altman at Stripe Sessions: if OpenAI's headcount today is 100, what will it be in five years? Altman's answer: "Hopefully 200."
He said OpenAI is entering its third phase. The first was a pure research lab where talking about building AGI sounded crazy. The second added a product company. The third is about becoming a massive token factory — turning intelligence into a public utility that's as cheap and easy to use as possible. That requires deep full-stack integration and massive infrastructure. Every transition demands a complete overhaul of management style, and Altman says the third phase is probably "not a style I'm good at." His solution: either find someone to run it, or "build an AI to run it."
Collison asked for OpenAI's craziest untold story. Altman said the eight months between finishing GPT-4 and releasing it were bizarre. The whole company used the model internally and knew it was fundamentally different from anything before, but nobody outside knew. They'd ask each other in the hallway: "Are we all delusional?" With no external feedback to calibrate against, it felt incredibly strange.
On managing a team full of people who all think they're the smartest in the room, Altman recounted a biographer's summary: "Your greatest skill is getting a bunch of people who each think they're the only competent one to work together and produce a breakthrough." The cost, he said, is "a lot of pain." The key was consensus — everyone bet on scale, pooling compute in one direction. When training GPT-3, almost all of OpenAI's compute was riding on that single line. A DeepMind researcher they consulted thought it was insane — compute had to be evenly split across groups or you'd create a toxic competition for resources. OpenAI ignored that advice and kept placing concentrated bets.