Greg Brockman on Building OpenAI, the Board Crisis, and the Race to AGI

OpenAI’s co-founder and President Greg Brockman traces the lab from a 2015 Napa offsite with no offers and no structure, through the scaling bet and the November 2023 board crisis that fired and reinstated Sam Altman, to a compute-constrained race in which the durable advantage is ‘the machine that makes the models’, not any single model.

Key ideas

1. No one could prove it impossible. In 2015, Google DeepMind was the ‘10,000-lb gorilla’ — all the researchers, capital, and data. Brockman and Altman found people who could list why a new lab was hard but none who could say it was impossible, and started anyway. The founding team crystallised at an offsite in Napa, before any offers, around a three-line technical plan — solve reinforcement learning, solve unsupervised learning, gradually learn harder things — that OpenAI then pursued for a decade.
2. Massive compute with simple algorithms works in practice. OpenAI built its Dota bot to prove that the era’s reinforcement-learning method (PPO) would not scale — then kept scaling it until it beat the best humans. The lesson was the finding itself: crude algorithms plus enormous compute succeed in messy, unprogrammable environments, and the same approach scaled towards brain-scale networks becomes the whole bet.
3. Prediction is intelligence. Predicting the next word sounds pedestrian, but predicting the next word out of Einstein’s mouth requires being as smart as Einstein — prediction on situations never seen before is deeply connected to intelligence. Unsupervised pre-training and reinforcement learning use the identical predictive technology; only the structure of the data changes.
4. Pressure makes diamonds or cracks. The board fired Altman and removed Brockman with no reasons given; Brockman quit within the hour. Nearly the whole company signed a petition, cancelled Thanksgiving flights, and not a single person accepted a competing offer over the weekend — ‘a diamond moment’ — and the team came back intact.
5. The world is heading compute-constrained. As models move from answering questions to spending vast tokens solving hard problems, demand outruns supply; OpenAI’s early, mocked data-centre bet becomes an advantage. The moat is not one model but the machine that produces the next one, which iterative deployment refines against reality rather than a single high-stakes launch.

Summary

Founding OpenAI: from Stripe to a Napa offsite

Brockman left Stripe — a startup solving a problem he judged would succeed with or without him — to work on the one problem he wanted to spend his life on: how AI plays out in the world. Patrick Collison sent him to Sam Altman hoping Altman would talk him out of leaving; instead the two discovered a shared intent, and a July 2015 dinner asked whether it was still possible to start a lab with the best researchers when DeepMind held the field. Driving back to the city that night, they decided they had to. The team did not cohere easily — early candidates included Ilya Sutskever, Dario Amodei, and Chris Olah; Amodei and Olah went to Google Brain — so Brockman broke the symmetry with a Napa offsite (he made the t-shirts), where the flowing ideas produced the technical plan the lab pursued for ten years.

The scaling bet and the nature of intelligence

Escalating compute maths in 2017 — and the promise of Cerebras hardware — convinced Brockman, Altman, Sutskever, and Elon Musk that a nonprofit could not raise the dollars an AGI would need, forcing the for-profit structure. Progress arrived as a series of moments that felt ‘real now’: the 2017 sentiment-neuron paper, where semantics first arose from pure next-character prediction; the Dota bot, built to expose PPO’s limits but which scaled past the best humans and proved massive compute plus simple algorithms works in practice. Brockman frames reasoning and prediction as connected in a deep way — reinforcement learning has the model generate its own data and predict the consequences of its actions, but the underlying technology is the same predictive machinery as pre-training, only the data structure differs.

The board crisis

Told on a video call — the board minus Altman — that Altman would be removed and that he too was off the board, with no reasons offered, Brockman felt only that ‘it just wasn’t right’, and told his wife he had to quit; she agreed. The outpouring surprised him: messages of support, close collaborators quitting the same day, a petition that crashed Google Docs, employees cancelling Thanksgiving flights to fill the office. The turning point was seeing Sutskever sign the petition. Not a single person accepted a competing offer over that weekend of circling rivals — what Parrish’s guest Coach Belichick would call playing for the person beside you, and what Brockman calls a diamond moment. The hardest personal moment came later, when Sutskever left; Brockman took time off, trained language models on DNA sequences for the Arc Institute, and found his way back to the mission.

Building for AGI: suffering, resilience, and decisiveness

Brockman leads from the front — ‘I don’t always look back to see if everyone’s following’ — and prizes building environments where others do the individual-contributor work, a type-two fun that is painful in the moment but worthwhile. He invokes Sutskever’s dictum that you have to suffer to build value: OpenAI’s method has been to encounter reality as it is rather than paper over hard truths with Silicon Valley reality-distortion. The lesson he relearns is to make the hard decision and have the hard conversation; his regrets are almost always dragging feet on a call he already knew — the wrong person in a role, the wrong technical direction — held too long.

The race to AGI: compute, safety, and iterative deployment

Brockman calls the present a ‘global AI renaissance’ with country dynamics not yet settled, and argues American leadership matters for preserving democratic values while bringing the world along. Distillation and model theft miss the point: on an exponential, any stolen model is already superseded, and the real strength is the machine that makes the models — one reason OpenAI stopped showing chain-of-thought, both to hinder distillation and to keep reasoning traces faithful rather than trained to look presentable. He sees a compute-constrained world (8 billion GPUs would be needed for one each; the world builds hundreds of thousands), where society must decide which problems deserve compute, and frames safety as a core product feature plus societal resilience — seat belts and roads for the engine, not just the engine. Iterative deployment — shipping intermediate systems so the lab and the world adapt together — is his answer to the alternative of one secret, untested, world-changing launch.

Speakers

• Greg Brockman — co-founder and President of OpenAI and its first CTO; previously CTO of Stripe; a hands-on, lead-from-the-front engineer-leader.
• Shane Parrish — founder of Farnam Street; host of The Knowledge Project.

See also

• Sam Altman — OpenAI co-founder and CEO; his firing and reinstatement is the episode’s pivot
• Ilya Sutskever — OpenAI co-founder; the ‘you have to suffer’ dictum and the petition signature that turned the crisis
• Dario Amodei — an early prospective team member who went to Google Brain before founding Anthropic
• Scaling Laws — the scaling bet: scale the compute, scale the result
• Bitter Lesson — engaged directly by the Dota result: massive compute with simple algorithms beats hand-crafted method
• The Road to AGI — the founding mission and the compute-constrained race this episode narrates
• Shane Parrish — host