OpenAI's co-founders bet everything on agents, personal AGI and endless compute, while conceding AI could worsen inequality.
Greg Brockman says Sora is a different branch of the tech tree: the models powering it are not unified with the core GPT series, and its creative-expression use case does not fit the focus goal. He calls it an incredible model from a great team, whose technology will live on in other applications.
First, an agentic platform, the body built around the model brain with skills, connectors, computer use, context and memory. Second, applying agents to computer work, making Codex for everyone, not just software engineers. Third, personal AGI: an AI that knows your context, has your trust and acts on your behalf.
Brockman says Sam, Ilya, Greg and Elon all agreed a for-profit was the only path. They conceded majority equity and even Elon as CEO, but Elon demanded absolute control over OpenAI. Believing no single person should control the future, whoever it is, that was the breaking point that made them say no.
Altman sketches one future where everyone gets about 10 times richer but some become trillionaires and inequality worsens, and another where people feel only twice as rich but inequality falls. He thinks the first is better, and says everyone should want much more compute and the cheapest possible AI access, since scarcity would exacerbate inequality.
Brockman concedes OpenAI came late to applying coding models to messy repos and real-world data, something Anthropic appreciated earlier, and gives them kudos. He says competition elevated OpenAI's execution, and head-to-head Codex versus Claude now gets very favorable results, though what matters is progress towards AGI, not the news cycle.
Altman says some people have long wanted AI held by a small group, and fear-based marketing is the most effective justification: build a bomb, then sell a hundred-million-dollar bomb shelter. He grants Mythos is probably a great cybersecurity model, but OpenAI uses its preparedness framework, trusted access program and mitigations instead.
Brockman explains models learn by predicting text, then reinforcement learning rewards good outputs, which is easy in math and science but hard to judge in open-ended fields. Altman adds the model serves about a billion people with different tastes; better personalization is the goal, and new models on this dimension are coming soon.
The hand was trained with the same reinforcement learning algorithm used for competitive video games, but its string tendons snapped after about 20 hours, requiring a mechanical engineer to fix it. You cannot do ML that way, so the project was cancelled and the team went on to work on what became GitHub Copilot.
No. Altman says individual sites get dropped for specifics, like only having air cooling, but OpenAI will keep building as much compute as possible. Brockman adds compute is a profit center, not a cost center: OpenAI rents or buys compute and resells it at a margin, and demand is unlimited.
After a dinner in July 2015, driving back to the city, Sam and Greg agreed they had to start a lab going after AGI; the conclusion was that it was not obviously impossible. Greg, unemployed, was full-time on it the next day. They had met around 2010 through the Collisons, when Greg joined Stripe.
“We are building something that contorts the machine to you.”
“But really, truly, no one cared.”
“It's a strange phrasing because I don't feel constrained by the business.”
“One thing that I think is also worth thinking about is that compute for us is not a cost center.”
“The OpenAI Founders On Their Plan To Battle Elon, Compute And Everything Else.” https://www.youtube.com/watch?v=NCKQL0op30E. Transcript summary by WhipScribe, https://whipscribe.com/transcript-pages/the-openai-founders-on-their-plan-to-battle-elon-compute-and.html.