Altman: I may not be in charge of the next stage of OpenAI; either find someone or build AI

According to the Motion Research from Beating, Stripe CEO Patrick Collison asked Sam Altman at Stripe Sessions: If you set OpenAI's headcount today to 100, what is it in 5 years? Altman's answer: "Hopefully 200."

He said OpenAI is entering its third phase. The first phase was a pure research lab, talking about building AGI sounded like nonsense. The second phase added a product company. The third phase aims to become a large-scale "token factory," turning intelligence into a public utility, as cheap and easy to use as possible, requiring deep full-stack integration and a large amount of infrastructure. Each transition in management style has to be significant, he said the third phase is probably "not my style," either find someone to run it or "build an AI to run it."

Collison asked for OpenAI's craziest untold story. Altman said there was an 8-month gap between training GPT-4 and releasing it, and the whole company used it internally, knowing it was completely different from before, but no one outside knew. Sometimes they would ask each other in the hallway: "Have we collectively fallen into delusion?" There was no external feedback to calibrate their judgment, and it was an extremely strange time.

Talking about how to manage a group of people who all think they are the smartest, Altman relayed a summary from an author writing an OpenAI biography: "Your most powerful point is making a group of people who all think they are the most capable work together until a breakthrough is made." He said the cost is "a lot of pain," and the key is consensus: everyone bets on scale, concentrating computing power in one direction. When training GPT-3, almost the entire company's computing power was focused on this line, the DeepMind researchers they were in contact with at the time thought this approach was crazy, and computing power had to be evenly distributed among the teams, otherwise internal vicious resource competition would arise. OpenAI did not listen and continued to concentrate its bets.