Greg Brockman, co-founder and former CTO of OpenAI, discusses the company's journey from startup idea to AGI leader. Brockman left Stripe in 2015 after a conversation with Sam Altman about starting an AI company, leading to OpenAI's founding with the mission of ensuring artificial general intelligence benefits all humanity.

The conversation covers OpenAI's evolution from nonprofit to for-profit structure, the technical breakthroughs from Dota to GPT-4, and the dramatic board crisis that temporarily removed Sam Altman. Brockman explains the company's iterative deployment philosophy, massive compute investments, and vision for personal AGI that will transform how humans work and live.

From Stripe to OpenAI: The Mission That Mattered

Brockman left Stripe because he wanted to work on a problem that was uniquely his: 'It wasn't the problem I'd grown up thinking about... I felt like it was going to succeed with or without me.'

Patrick Collison connected Brockman with Sam Altman, who was also thinking about AI: 'I'm also thinking about doing something in AI. We should keep in touch.'

The July 2015 dinner addressed whether it was too late to compete with DeepMind's dominance in AI research and talent acquisition.

The founding team crystallized during a Napa off-site where they developed OpenAI's three-part technical plan that remains unchanged after 10 years.

The Nonprofit-to-Profit Transformation

By 2017, compute requirements made the nonprofit structure unsustainable: 'There is essentially a cap to what is possible' in nonprofit fundraising - Greg

The discovery of Cerebris computing hardware revealed the massive compute advantage possible with exclusive access to advanced systems.

Elon Musk, Sam Altman, Ilya Sutskever, and Brockman unanimously agreed that creating a for-profit entity was 'the only path to achieve the mission.'

Technical Breakthroughs and Scaling Insights

The 2017 unsupervised sentiment neuron paper marked the first time semantics emerged from language modeling 'We are building machines that can learn semantics, not just where the commas are.'

Dota proved that massive compute with simple algorithms could exceed human performance in complex, unprogrammable environments using just an 'insect brain' sized neural network.

GPT-4 challenged AGI definitions: 'Someone asked, why is this thing not an AGI? It's like actually really hard to put your finger on it.'

Current AI systems write virtually all code at OpenAI, with humans focusing on architecture and interfaces rather than implementation.

The Board Crisis and Loyalty Test

Brockman learned of Sam Altman's firing via video call with the board, receiving the same public messaging with no additional context or reasoning.

Brockman was simultaneously removed from the board but told he was 'very critical to the company and the mission' and should stay.

His immediate response was decisive: 'Right after I hung up the call, I talked to my wife and I said, gotta quit. She said, I agree.'

The employee petition crashed Google Docs from simultaneous editing attempts, demonstrating unprecedented solidarity in tech history.

Despite competitor offers during the crisis weekend, OpenAI 'did not lose a single person' to competing companies.

The Compute-Constrained Future

Global compute demand vastly exceeds supply: 'You wanted one GPU for every person in the world, you're talking like 8 billion GPUs' versus current hundreds of thousands.

OpenAI's early data center investments, once mocked by competitors, now provide significant competitive advantage in the compute-constrained environment.

Data centers will become problem-specific: 'Having this giant machine... dedicated towards a problem' like cancer research could happen 'this year.'

Society must decide compute allocation priorities: 'Where does the compute go? What problems are worthy?' becomes the critical question.

Personal AGI and Human Empowerment

Personal AGI will operate proactively: 'If your favorite musician is in town, it just goes and proactively purchases tickets' based on deep personal knowledge.

The vision extends to 8 billion people having personal AI agents that know them well and can take trusted actions on their behalf 24/7.

Future workers will become 'managers of agents' and potentially 'CEO of an autonomous AI corporation' with 100,000-person equivalent workforce.

The technology enables universal building capability: 'Anyone can be a builder... everyone now can be a software engineer' through tools like Codex.

Safety, Regulation, and Societal Adaptation

Safety is fundamentally a product feature: 'No one wants a model that is not aligned with them' - making safety investment essential for success.

Iterative deployment prevents the risk of first-time deployment of powerful systems: 'Do you want to be sitting in a room... you've never deployed anything ever before?'

AI conversations need legal privilege protection similar to doctor-patient or attorney-client relationships to encourage beneficial use.

Regulation should ensure broad benefit distribution and maintain American AI leadership while supporting societal adaptation to compute-powered economy.

Resources Mentioned

Talent Acquisition Excellence Using Digital Capabilities and Analytics to Improve Recruitment

The July 2015 dinner addressed whether it was too late to compete with DeepMind's dominance in AI research and talent acquisition. The founding team crystallized during a Napa off-site where they d

Welcome Cuntz

ill become problem-specific: 'Having this giant machine... dedicated towards a problem' like cancer research could happen 'this year.' Society must decide compute allocation priorities: 'Where does t

Homeschool Bravely How to Squash Doubt, Trust God, and Teach Your Child with Confidence

I trained language models on. That's when you learned how to do it, right? Like you did the self-study thing right on your blog? Well, no, so I actually had done that. I actually had done that thro

ideas and test those out

computers. And we're going to be hitting a phase soon where the AI will also come up with its own research ideas and test those out, run experiments. And so I think that the speed of iteration and i