Public transcript

The OpenAI Founders On Their Plan To Battle Elon, Compute And Everything Else

By: Core Memory Podcast
Length: 1 h 23 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

OpenAI co-founders Sam Altman and Greg Brockman join Ashley Vance and Kylie Robison on the Core Memory podcast for their first joint podcast appearance in years. They reflect on their decade-long partnership, OpenAI's evolving approach to safety and product strategy, and the recent refocusing of the company around agents, Codex and a personal AGI, including the decision to deprioritize Sora. They also discuss inequality, US hardware competitiveness, government relations, the recent attack on Altman, and the upcoming Elon Musk lawsuit, which they frame as a chance to finally tell their side of the story.

Key points

Sam Altman and Greg Brockman say their constant contact and complementary approaches have been one of the nicest things about OpenAI through ten years of drama.
They argue OpenAI's greatest contribution has been a different way of talking about safety, centered on iterative deployment rather than fear.
Altman lays out two futures: one with massive prosperity but trillionaires and worse inequality, and one with less of both, and says people should want the first but emotionally often don't.
Brockman has taken over product and research focus, prioritizing an agentic platform, Codex for everyone doing computer work, and a personal AGI that knows your context.
Sora was deprioritized because it sits on a different branch of the tech tree and its use case doesn't fit the unified product suite.
Altman says compute is a profit center, not a cost center, and denies OpenAI is pulling back on infrastructure or its chip effort.
He criticizes fear-based marketing around models too dangerous to release, and says he doesn't think Anthropic was treated well in the Department of War episode.
On the Elon Musk lawsuit, Brockman says the breaking point was Elon's demand for absolute control, and Altman sees the trial as a chance to finally tell their story.

Questions it answers

02:00Founding story and co-founder bond

How did Sam Altman and Greg Brockman's relationship form and evolve?

They met around 2010 through the Collisons, and after a July 2015 dinner they decided to start the lab, staying in near-constant contact since. Brockman says hard work in stressful times forged a relationship he hasn't seen formed any other way.

They decided to found OpenAI driving back from a July 2015 dinner.
They remain in constant sync, roughly five short calls a day.
They credit complementary approaches: Brockman's focus, Altman's ambition and connections.

08:00Safety framing and iterative deployment

How did they learn to talk about AI safety?

They never disagreed on safety's importance but did on framing; Brockman resisted the field's traditional fear-based frame, and Altman says iterative deployment became one of OpenAI's greatest contributions, against extreme pressure.

Iterative deployment means shipping less powerful technology first to help society adapt.
AI resilience is a pillar of the OpenAI Foundation, which holds roughly 25-30% of OpenAI equity.
Altman says the field hasn't filled in the dot dot dot of why superintelligence benefits people.

12:00ChatGPT moment and intuitive technology

Why did ChatGPT change public perception of AI?

Altman says pre-ChatGPT feats like the Rubik's Cube robot got no real attention, while ChatGPT, far less impressive technically, changed the world's view because people could feel it and get value. Brockman adds AI is the most intuitive technology, adapting the machine to you.

Coding models were the second big moment the world updated on AI.
A new image model is coming soon that Altman says shows how much further image generation can go.
Models still lack awareness of your life; a model that knows all your context is coming.

23:00Inequality and compute access

What happens to inequality in an AI world?

Altman sees one world with a dramatically raised floor but trillionaires and worse inequality, and another with less total prosperity but less inequality; he says people should prefer the first but emotionally many don't. Brockman argues universal cheap compute access enables the most extreme version of the American dream.

Altman says everyone should want much more compute and cheapest possible AI access.
A doctor in your pocket, free and better than today's best medical team, is coming.
The OpenAI Foundation's equity, over 150 billion dollars, is meant to benefit the world.

28:00US hardware competitiveness and robotics

Can the US compete in hardware and robotics?

Both agree the US is far behind in robotics and componentry, and Altman says AI plus general-purpose robots is the only credible plan to catch up fast enough; OpenAI is trying to be very successful at robotics. They recall the 2018 robotic hand project, cancelled after tendon strings kept snapping.

A Codex-equivalent for robots that could configure factories and mines would change the playing field.
The 2018 robotics hand team went on to work on what became GitHub Copilot.
Gigawatts of power, chip fabrication and racks are areas where the US is extremely behind.

35:00Agent transition and product cuts

What did OpenAI prioritize and cut in its recent refocus?

Brockman, now leading product focus, prioritizes an agentic platform, Codex for everyone doing computer work, and a personal AGI that knows your context. Sora was cut because it's a different branch of the tech tree with a less unified use case.

Models have shifted from being the product to being a part of the product with a fat software layer.
Codex will expand beyond software engineers to all computer work, with updates coming quickly.
A new model solved a long-standing Erdős problem, which Brockman says no one is pricing in.

41:00Compute, chips and the Anthropic rivalry

Is OpenAI pulling back on compute, and how do they see Anthropic?

Altman says compute is a profit center, not a cost center, since the business rents or buys compute and resells it at a margin, and denies pulling back on infrastructure or the in-house chip. On Anthropic, Brockman concedes they were late to apply coding models to messy real-world repos, but says Codex now gets favorable head-to-head results versus Claude.

Altman says stories of pulling back are wrong and OpenAI will keep building as much compute as possible.
He's very excited about the chip team but won't give timelines.
Robotics is not the current thing but will be clearly important in the future.

46:00Dangerous models and government relations

What do they think about models too dangerous to release?

Altman calls fear-based marketing the most effective way to justify keeping AI in a small group's hands, referencing the "we built a bomb, buy our bomb shelter" pitch, but says some safety concerns are legitimate. He says he doesn't think Anthropic was treated well in the Department of War episode and that leading AI labs must assist the US government.

OpenAI has had cybersecurity in its preparedness framework with mitigations and trusted access programs.
Altman criticizes using DPA threats and supply chain risk designations against labs.
He sees no good future where leading AI efforts don't assist the US government.

Notes

The co-founder relationship

Sam Altman and Greg Brockman describe staying in constant sync since OpenAI's founding after a July 2015 dinner, often with five short calls a day.
Brockman credits Altman with seeing connections between ideas and constantly raising ambition, especially on compute; Altman credits Brockman with pushing focus on the single most important thing.
Their biggest past disagreement was how to talk about safety; Brockman held a line against the field's traditional fear-based framing, which Altman says shaped OpenAI's strategy of iterative deployment.

Safety and society

The OpenAI Foundation, the nonprofit governing OpenAI, holds roughly 25-30% of OpenAI equity, over 150 billion dollars, with AI resilience as a pillar.
Altman says the field has not filled in the "dot dot dot" of why superintelligence is good for people; what people want is prosperity, agency, meaningful work, and even a "right to adversity."

Products and the agent transition

Brockman, recently taking a foreground product role, says models have shifted from being the product to being a part of the product, with a fat software layer of skills, connectors, context and memory acting as the body around the model brain.
Priorities: an agentic platform, Codex for everyone doing "computer work" (not "knowledge work"), and a personal AGI that knows your context and can act on your behalf.
Sora was cut because it's on a different branch of the tech tree with a different use case, though the team's work will live on.
New image models and writing improvements are coming soon; Altman says a model that knows all your context will completely change what using a computer feels like.

Inequality, compute and hardware

Altman sketches two futures: one with a dramatically raised floor but trillionaires and worse inequality, another with less prosperity but less inequality, and says the choice is the real crux.
Brockman argues cheap, universal compute access enables the most extreme version of the American dream for kids who lean in.
Both agree the US is far behind in hardware and robotics; Altman says AI plus general-purpose robots is the only credible plan to catch up, and OpenAI is trying to be successful at robotics.

Rivals, government and the lawsuit

Altman calls compute a profit center and denies pulling back on Stargate or the in-house chip effort.
He criticizes fear-based marketing of "too dangerous to release" models and says he doesn't think Anthropic was treated well in the Department of War dispute, while insisting leading AI labs must assist the US government.
On the Musk lawsuit, Brockman says the breaking point was Elon's demand for absolute control of OpenAI; Altman calls the trial a real opportunity to tell their story and jokes his fear is Elon drops the case first.

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