OpenAI Co-Founder: “I just had to Quit”

KEYPOINTS

• 01

  OpenAI started from a 2015 dinner question: is it too late to compete with DeepMind?

  DeepMind had the researchers, capital and data, and AlphaGo was months away. At Sam Altman's July dinner, people listed reasons it would be hard but nobody could name one that made it impossible. Driving back, Brockman and Altman agreed 'we got to do this', and Brockman went full-time the next day. Dario Amodei and Chris Olah turned down the early team and went to Google Brain.

• 02

  A Napa offsite turned interested researchers into a team and produced a 10-year technical plan.

  About ten people were each waiting to see who else would join. At Altman's suggestion, Brockman (who even made t-shirts) took them to Napa before any offers or structure existed. The plan they came up with was to solve reinforcement learning, solve unsupervised learning, then gradually learn more complicated things. Offers went out right after, with a two-to-three-week start date.

• 03

  The pure nonprofit was dropped because doing the compute math in 2017 showed charity money couldn't fund AGI.

  Projections of the compute needed, plus the prospect of exclusive access to Cerebras hardware or very large data centers, showed that the scale required was far beyond what nonprofit fundraising could reach. Raising $100–500 million was possible; a billion was 'pretty hard'. Elon Musk, Altman, Ilya Sutskever and Brockman agreed that a for-profit entity of some form was the only route to the mission.

• 04

  Dota showed that simple algorithms plus massive compute work in messy real-world settings.

  OpenAI chose Dota to force the invention of new methods, because PPO plans at every timestep with no hierarchy and seemed doomed not to scale. They kept scaling PPO anyway and beat the best humans, with a network about the size of an insect brain. The finding was that pushing baselines until they hit the wall can show there's no wall.

• 05

  Prediction isn't a shallow task: predicting Einstein's next word means being at least as smart as Einstein.

  The 2017 'unsupervised sentiment neuron', trained only to predict the next character, learned whether text was positive or negative, the first sign of semantics emerging. Reasoning models add reinforcement learning, in which the model generates its own data from actions, but training is still prediction, just on differently structured data.

• 06

  When stakes feel existential, ordinary office politics like credit and decision rights take on enormous weight.

  Brockman says the field is 'fragmentary': under pressure you get diamonds, meaning high-trust teams, or cracks, with groups splintering off. He frames the resulting diversity of labs and approaches, including on safety, as healthy debate that has moved from inside OpenAI into the wider world.

• 07

  His first-person account of the firing: no reasons given, and he quit the same day.

  On a video call with the board minus Altman, he was told Altman was out and that he himself had been removed from the board but would stay as 'critical to the mission'. When he asked for reasons, he got none. He told his wife 'got to quit' and she agreed. He put the chance of getting the company back at 10% that first day. When the board replaced interim CEO Mira Murati that Sunday, the company rebelled.

• 08

  Most of his regrets are about waiting too long, not about acting too fast.

  Asked to write his lessons on one page, he lists persistence through 'it's over' and 'so back' swings, steadiness as a leader, and decisiveness. When opinions contradict each other, you have to pick, knowing someone will be upset or quit. His regrets are mostly cases where they knew someone was the wrong person for a role, or a technical direction was wrong, and dragged their feet. Lesson learned more than once: 'Make the hard decision. Have the hard conversation.'

• 09

  OpenAI hides chain-of-thought both to resist distillation and to keep reasoning traces honest.

  Reading a model's reasoning turned out to be an unexpected way to interpret it. But if you train the chain of thought to look presentable, it stops faithfully reflecting how the answer was reached, so OpenAI decided early never to optimize it for appearance. On distillation, he argues the real moat is 'the machine that makes the models', since they're always on to the next model.

• 10

  The world is heading toward a compute shortage, and who gets compute becomes a societal question.

  One GPU per person would mean 8 billion GPUs, while fleets are in the hundreds of thousands to millions. He expects dedicated data centers, for example one aimed at cancer, possibly this year. OpenAI keeps a free tier because it believes everyone needs access, and it treats safety as a product feature, since nobody wants a model that isn't aligned with them.