Greg Brockman: Inside the 72 Hours That Almost Killed OpenAI

What It Covers

Greg Brockman, OpenAI co-founder, traces the company's origins from a 2015 dinner in San Francisco through the November 2023 board crisis that nearly destroyed it. He covers the technical roadmap that emerged from a Napa offsite, the shift from nonprofit to for-profit structure, and why massive compute investment became the defining strategic bet.

Key Questions Answered

• •Mission selection framework: When choosing what to dedicate your career to, Brockman applied one filter: would working on this problem for the rest of your life, even if you only moved it slightly forward, constitute a life well lived? For him, AI cleared that bar; Stripe did not, because he believed Stripe would succeed without him regardless of his contribution.
• •Breaking symmetry in team formation: To get 10 undecided researchers to commit to OpenAI before any formal structure existed, Brockman organized a Napa offsite with no offers, no company, and no contracts. The shared experience generated enough momentum that everyone committed within weeks. The technical roadmap produced that day — solve reinforcement learning, solve unsupervised learning, scale complexity — guided the next decade of work.
• •Compute as the decisive variable: In 2017, OpenAI ran the math on what AGI would require and concluded nonprofit fundraising had a hard ceiling around $500M–$1B. That gap forced the for-profit entity creation. The same logic later drove data center investment that competitors dismissed. Brockman frames this as "encountering reality as it is" rather than optimistic projection, and credits it as the core operational discipline at OpenAI.
• •Iterative deployment as a safety mechanism: Rather than building in secret and deploying once, OpenAI's strategy treats each release as the 100th deployment in a series of increasing capability. GPT-3 revealed the top misuse was medical spam advertising — something no internal threat model anticipated. Each deployment cycle builds institutional knowledge about real-world failure modes that no amount of internal testing can replicate.
• •Prediction and reasoning are the same process: Brockman argues that predicting the next token and reasoning from first principles are deeply connected. If a model can accurately predict what Einstein would say next in a genuinely novel situation, it is operating at Einstein's level. The reinforcement learning stage adds real-world feedback loops on top of the unsupervised prediction base, but both stages use identical underlying technology with different data structures.
• •AI code generation is near-total: At OpenAI, the fraction of code written by humans rather than AI has become vanishingly small. AI currently outperforms humans at writing code given correct context and structure. Human expertise remains valuable for module architecture, interface definitions, and system design decisions — but the actual code generation layer has effectively transferred to AI, with Codex positioned as a tool for non-engineers to build production software.