Public transcript

Ai Goes Parabolic | OpenAI Co-Founder Greg Brockman

By: Zenvora Productions
Length: 1 h 12 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

OpenAI co-founder Greg Brockman recounts the company's origins, from leaving Stripe and the 2015 Napa off-site that set the technical plan, through the shift to a for-profit structure, the Dota and GPT milestones, and the dramatic weekend of Sam Altman's firing and reinstatement. He discusses compute constraints, AI-written code, safety, iterative deployment, regulation, and his optimistic vision of personal AGI for everyone. He closes by defining success as achieving OpenAI's mission of ensuring AGI benefits all of humanity.

Key points

Greg Brockman left Stripe because AI was the mission he wanted to dedicate his life to, and Sam Altman was independently planning something in AI.
The 2015 Napa off-site produced OpenAI's technical plan: solve reinforcement learning, solve unsupervised learning, then gradually learn more complicated things.
In 2017, compute math convinced Elon, Sam, Ilia, and Brockman that a for-profit entity was the only path to building AGI.
When Sam Altman was fired, Brockman quit the same day, and the company rebelled; a staff petition crashed Google Docs and no one accepted a competing offer.
Brockman says the actual writing of code at OpenAI is essentially all AI now, with humans still better at structural design.
OpenAI stopped showing chain-of-thought reasoning to prevent distillation and to preserve faithful interpretability.
Brockman argues the world is compute-constrained and society must decide where compute goes, which is why OpenAI keeps a free tier.
He defines success as achieving OpenAI's mission of ensuring AGI benefits all of humanity.

Questions it answers

00:00Leaving Stripe and founding OpenAI

How did OpenAI come about?

Greg Brockman left Stripe seeking a mission he could dedicate his life to, and Patrick Collison sent him to talk to Sam Altman, who was also thinking about AI. After a July 2015 dinner where no one could prove a new lab was impossible, Brockman and Altman committed, and a Napa off-site with about 10 people produced the founding vision and early team.

Brockman left Stripe because AI was the problem he wanted to dedicate his life to.
A July 2015 dinner in 2015 posed whether a new lab was possible despite DeepMind's dominance.
The founding technical plan: solve reinforcement learning, solve unsupervised learning, gradually learn more complicated things.

04:47Nonprofit to for-profit

Why did OpenAI move away from being a pure nonprofit?

In 2017 the team did the math on compute and realized building AGI would require enormous resources, like Cerebras hardware and large data centers. Because nonprofit fundraising has a cap, Elon, Sam, Ilia, and Brockman agreed a for-profit entity was the only path to the mission.

Compute math in 2017 showed AGI would need massive resources.
Nonprofit fundraising has a cap on what's possible.
Elon, Sam, Ilia, and Brockman agreed a for-profit entity was the only path.

06:07Milestones: Dota, sentiment neuron, GPT-4

When did Brockman realize the technology was real?

He describes a series of moments: the Dota result showed massive compute with simple algorithms works in practice, the 2017 unsupervised sentiment neuron paper showed machines learning semantics, and GPT-4 was hard to distinguish from AGI by prior criteria. He argues prediction and reasoning are deeply connected, with reasoning models trained via reinforcement learning on the same prediction technology.

The unsupervised sentiment neuron paper showed semantics emerging from next-character prediction.
Scaling PPO in Dota beat the best humans with an insect-scale neural net.
Prediction and reasoning are deeply connected; reasoning models still train via prediction.

11:53Tension and credit in the field

Why did things get tense at OpenAI?

Brockman says that if you believe in the mission, questions like who makes decisions and who gets credit take on existential weight. He observes the field is fragmentary, with high-trust teams forming diamonds under pressure but also splintering, and that diversity of approach has produced healthy debate now happening society-wide.

High stakes give mundane office politics existential weight.
Pressure forms both diamonds and cracks in teams.
Diversity of approach has benefited the field.

16:00Sam Altman's firing and the rebellion

What happened when Sam Altman was fired?

Brockman learned on a video call that the board removed Altman and also removed Brockman from the board; he quit the same day. He, Sam, Jakob, Shimone, and Alexander sketched a new company, but after the board named a new interim CEO, the company rebelled, a petition circulated so heavily it crashed Google Docs, and Altman was reinstated without OpenAI losing a single person to competing offers.

Brockman was removed from the board the same call and quit the same day.
The company petition crashed Google Docs from too many signers.
No one accepted a competing offer that weekend.

24:09Aftermath, Ilia's departure, and lessons

How did Brockman process the aftermath and Ilia's exit?

Brockman says Ilia's departure was maybe the only moment he didn't want to do it anymore, and he took time off, training language models on DNA sequences for ARC Institute, an area meaningful to him and his wife's health. His biggest lesson is to keep going for something worth it, be decisive, and not drag feet on hard decisions he already knows the answer to.

Ilia's departure was the only moment Brockman didn't want to continue.
He spent time off training language models on DNA sequences for ARC Institute.
His repeated lesson: make the hard decision, have the hard conversation.

32:33AI accelerating AI

Is AI making AI development go parabolic?

Brockman says applying AI to its own development is speeding iteration, with coding tools revolutionizing software engineering and most model production bottlenecked by software. He says the actual writing of code is essentially all AI now, models are starting to produce novel ideas in math and physics, and OpenAI stopped showing reasoning partly to prevent distillation and partly to keep chain-of-thought faithful for interpretability.

The actual writing of code is essentially all AI now.
Models solved an open quantum physics problem with an elegant formula.
OpenAI hides chain of thought to prevent distillation and preserve interpretability faithfulness.

35:08Bias, alignment, and the global race

How does OpenAI handle political bias and the global AI race?

Brockman says OpenAI invests in neutrality with a public spec, and that viral screenshots can be misleading due to hidden context. After models leaned into telling users what they wanted to hear, OpenAI retrained them to align with users' long-term goals. On geopolitics, he calls it a global AI renaissance, says US leadership matters for democratic values, and that leadership means bringing the world with you.

OpenAI published a spec defining model values with public feedback.
Models were retrained to align with long-term goals, not flattery.
Leadership means bringing the world along, not just being ahead.

Notes

Founding OpenAI

Greg Brockman left Stripe because AI was the problem he wanted to dedicate his life to; Patrick Collison sent him to talk to Sam Altman.
A July 2015 dinner asked whether a new lab was possible despite DeepMind's dominance; no one could prove it impossible.
A Napa off-site (with t-shirts, no offers yet) produced the technical plan: solve reinforcement learning, solve unsupervised learning, then gradually learn more complicated things.
Early team included Ilia, John Schulman, and Brockman; Dario and Chris went to Google Brain.

Nonprofit to For-Profit

In 2017, compute math and Cerebras hardware showed AGI needed massive resources; nonprofit fundraising has a cap.
Elon, Sam, Ilia, and Greg agreed a for-profit entity was the only path to the mission.

Key Milestones

Dota: scaling PPO with massive compute beat the best humans using an insect-scale neural net.
The 2017 unsupervised sentiment neuron paper showed semantics emerging from next-character prediction.
GPT-4 was hard to distinguish from AGI by prior criteria; prediction and reasoning are deeply connected.

The Firing and Return of Sam Altman

Brockman learned of Sam's removal on a video call and was also removed from the board; he quit the same day.
The company rebelled; the petition crashed Google Docs; no one accepted a competing offer that weekend.
Ilia's later departure was the only moment Brockman didn't want to continue; he took time off training language models on DNA sequences for ARC Institute.

AI Accelerating Itself

The actual writing of code is essentially all AI now; humans still excel at structure and interfaces.
Models solved an open quantum physics problem with an elegant formula.
OpenAI hides chain of thought to prevent distillation and preserve interpretability faithfulness.

Compute, Safety, and Deployment

The world is compute-constrained; data centers are the biggest machines humanity builds.
Iterative deployment: GPT-3's top misuse was medical spam, not misinformation.
Safety is a core product feature; societal resilience (like seat belts for cars) matters too.

Advice and Vision

Everyone will manage agents; entrepreneurship becomes easier; anyone can be a builder with tools like Codex.
Success: ensuring AGI benefits all of humanity.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.