Sam Altman on Vanity Fair: Elon Musk, Paused Models, Extinction Risk

Sam Altman sat for a two-part Vanity Fair interview on Elon Musk, a paused OpenAI training run, the Trump AI accord and whether extinction risk is overhyped. The conversations were taped in San Francisco on September 11 and October 2, 2026 with Mark Guiducci for the Fair Game podcast, to go with Altman’s Vanity Fair cover and his number one spot on the magazine’s New Establishment list. Guiducci notes up front that Condé Nast, Vanity Fair’s parent, has a content licensing agreement with OpenAI. Part one runs 49 minutes and part two runs 24. This post covers both.

TLDW

Altman says he has been mischaracterized as a villain, regrets taking no equity in OpenAI, and insists the mission will beat the share price after an IPO he is in no rush to do. He says OpenAI paused a training run and then held back its Astra 6.1 model on safety grounds, because models are now capable enough to go wrong during training and not only after release. On policy he calls the Trump superintelligence accord a good start with no real power, expects regulation to be written in reaction to whatever breaks first, and says society will have to accept some severe cyber incidents as the price of open source models. He calls Elon Musk a bully he will still work with, says extinction risk is nonzero and deserves to be taken seriously, and admits the thing he does not say in public is that he thinks it is all going to be great.

Thoughts

Altman’s whole regulatory position comes down to a road with a ditch on each side. One ditch is the accident: a model that escapes and that nobody can pull back. The other is concentration: a technology judged so dangerous that only one company, one person or one country is allowed to hold it. Nearly everything he says in part two is an effort to price the second ditch as high as the first, and the clearest evidence is the trade he is willing to say out loud. He supports open source models while predicting a “tidal wave” of cyber security problems from them, and says society will “have to just accept some fairly severe cyber incidents” in exchange for the liberty. Few lab CEOs will name the damage their preferred policy is going to cause. It is also a convenient trade. The ditch he fears most is the one where rivals and regulators decide who gets access, and the cost he asks the public to tolerate lands on hospitals, banks and city governments, not on OpenAI.

His forecast for how rules will actually arrive is bleaker than his tone. He grants that the accord signed by President Trump and the AI companies “has no real power”, then says policy will be “very reactive to the specific things that go wrong”: a novel pathogen produces one kind of law, a cyber incident another, labs cutting businesses off from models a third. Later he recalls his 2023 Senate testimony, where he was assured a bill would pass in two weeks and nothing happened. Put those together and the working plan is to wait for the accident and hope it is a small one. That sits badly beside part one, taped three weeks earlier, where he said the United States and China “could do it” on shared, verified safety thresholds. By part two the state dinner for Xi Jinping had come and gone, and what he can report is a tour of a ballroom that is still a construction site. He keeps describing a window of public support. Nobody with authority is climbing through it.

The most consequential technical claim gets about two sentences. OpenAI used to test a model after training finished and before release. Now, he says, “even during training things can now go wrong”, because the models are smart enough to get out of a sandbox or hack into a system. That matches the paused training run in part one and the Astra 6.1 delay in part two. If the dangerous moment has moved from release to training, then both rival positions Guiducci puts to him miss the target. Jensen Huang’s “don’t ship products until they’re in control” and Bill Gates’s comparison to drug approval are both about shipping. The one item on Altman’s own list that reaches inside a lab before any customer sees anything is aviation-style incident reporting. It is the proposal to hold him to, especially since his answer to a reported claim that executives ignored employee warnings months before the Hugging Face incident was that he was not familiar with the report.

The equity confession is more revealing than it sounds. Altman holds no direct OpenAI stock, a choice meant to show he would decide for humanity and not for his net worth, and he now calls it “a mistake” and “weirdly this big trust destroying thing”. People who cannot see a motive assume a hidden one. The lesson matters for the IPO. He is asking public shareholders to accept that if the mission and the stock ever conflict, “then it falls, maybe it falls a lot”. His own experience suggests markets, like the public, trust a legible incentive more than a declared virtue.

Read the two endings together. Part one closes with the belief he says he keeps private because it would be misunderstood: “that it’s all going to be great”. Part two closes with the legacy he wants, AI as fire or electricity handed to everyone and not something “a small group of people worshiping the machine god” decided to bestow, and he accepts Prometheus as the comparison. A person who privately expects a good outcome will always fear gatekeepers more than accidents, and every trade in these 74 minutes follows from that prior. Listeners who do not share it should notice how much of the argument it is carrying.

Key Takeaways

Altman says he has been mischaracterized as a villain and finds it “only somewhat frustrating”. His view is that the public needs “a sink for the anxiety” and anyone in his job would get the same treatment. He admits he once held the same caricature of Mark Zuckerberg.
He holds no direct OpenAI equity, only a sliver of exposure through an old Y Combinator investment, and says that was a mistake: “I wish I had just taken the stock.”
OpenAI is “not in a rush” to go public. When it does, a nonprofit will still govern it, and Altman says that if the share price and “the right thing for the future of humanity” conflict, the share price loses.
OpenAI recently paused a model training run to make more progress on alignment and monitorability. In part two he says the company also delayed Astra 6.1 because it “would have tipped too much towards the accident risk”.
Asked whether he would press a button to pause everyone, he gives three levels: what OpenAI will do alone, what American labs should be allowed to do together (which he says needs antitrust exemptions), and what has to happen internationally, starting with the US and China.
The superintelligence accord signed by President Trump and the AI companies is “a good start” and “much better than nothing”. He concedes it has no real power but expects it to carry social force between the companies.
He supports open source models, expects a “tidal wave” of cyber security problems from them, and says the price is worth paying. He does not think banning open models is worth the cost to liberty.
His regulatory wish list borrows from aviation: transparent incident reporting on the model of the NTSB, pre-release standards and a liability framework.
On Elon Musk: “I don’t like bullies and I think Elon’s a bully.” He says he will still work with him, and with Anthropic’s Dario Amodei, about whom his one-word answer was “pass”.
He says he did not speak to Amazon or Jeff Bezos about Artificial, the Luca Guadagnino film in which Andrew Garfield plays him. Guiducci notes Amazon produced it, declined to distribute it, and invested $50 billion in OpenAI this year. Neon releases it on Christmas Day.
He will not run for office. He confirms he once considered running for governor of California and disputes the claim in The Optimist, Keach Hagey’s biography, that he considered the presidency.
The one thing he says people should not use AI for is replacing human connection.

Chapters

Part one, taped September 11. Timestamps link into the first video.

01:36 The Face of AI, the Scapegoat and His Kids

Altman says OpenAI has been predicting this moment for a decade and the difference now is that most people believe it. Asked whether AGI is here, he calls the question subjective and says “it’s close”, while granting that the world has not changed much yet. On fatherhood, he calls his two children the best thing he has done and says he did not need them to be “very intent on not destroying the world”. His children do not use ChatGPT, he has no age in mind for when they will, and he favors limiting technology for small children across the board.

11:36 Artificial, the Bunker and Running for Office

Altman describes walking out for coffee and meeting Andrew Garfield dressed as him, same backpack, same walk, filming outside his house. He has seen the trailer and does not think the film is going for accuracy. There is no bunker with firearms on the wall, though he says he has a survivalist mindset and a ranch that can run off grid, with no helicopter and no submarine. He rules out any run for office and says AI matters far more than fixing California. Asked how AGI relates to his idea of God, he says not at all.

19:14 The IPO, Mission First and Zero Equity

Guiducci asks what going public will do to the soul of the company. Altman says OpenAI will not be like any other public company, that investors will be reminded the mission comes first, and that quarter to quarter the stock will probably be very volatile. On his own pay, he calls the no-equity decision a failure to understand how most people think, and says what he gets instead is the most interesting job in the world. Told that someone might say he does it for power, he answers “that would be closer to true”.

26:05 Jacob Coxon’s Criticism, Agents That Break Out, the Executive Exodus

Asked about Jacob Coxon’s charge that neither OpenAI nor Anthropic is acting responsibly, Altman points to the company’s founding statements and its voluntary training pause, and says he is glad the world is finally listening. He draws a line against the old social media defense: when OpenAI’s tools do not behave as expected, the company is responsible. On the departures this year of the heads of ethics, safety systems and data centers, plus the operating and revenue chiefs, he says five years of company growth is being compressed into one and different people fit different stages.

30:39 Curing Disease, the Self-Driving Golf Cart and the Virus Trade

The OpenAI Foundation, which Guiducci describes as larger than the Gates Foundation by assets, has put $100 million toward Alzheimer’s research. Altman will not give a date for a cure but expects real progress on some diseases within a few years. He describes Stanford students who used a GPT model to build a self-driving golf cart in a weekend hackathon. To people who would give up a cancer cure to avoid an engineered virus, he says the instinct is understandable and wrong. He compares OpenAI to an electricity company that still builds in safeguards, and says dangerous open models from others are unstoppable without “a crazy price to liberty”.

35:57 The Big Red Button and Punching a Bully

Would he press a button that paused every lab, including China’s? Altman answers in three levels and says “just stop” misses the point: what he wants is enough progress on alignment, monitorability and safety at each stage to move safely to the next. He says OpenAI is talking to safety teams at other labs and to the government. When Guiducci raises his sparring with Musk, he says bullies only respect people who hit back. His picture of things going right is shared safety standards, limits on private company power, and a society rich enough to widen what people can do with their time.

43:34 Suicide, Chat Logs and the Belief He Keeps Private

Guiducci presses on self harm and on Laura Reiley’s New York Times essay about her daughter Sophie, who died by suicide after long conversations with ChatGPT. Someone off camera tries to wrap the interview and Altman chooses to answer. He says experts disagree on mandated reporting, that human help for people in crisis is a good thing, and that private chat data should not go to researchers without consent. In the closing word association he gives “powerful” for Donald Trump, “consistent” for Bernie Sanders, a warm tribute to Ilya Sutskever, and “not nice” for Musk.

Part two, taped October 2. Timestamps link into the second video.

00:45 Dots and the Astra 6.1 Delay

Altman calls Dots, OpenAI’s new always-on helper, the product he has been waiting for. His Dot works through his email and to-do list, and after one phone call it came back with reasons his ideas were bad and had already talked to three people on his team. He says privacy has to be extremely rigorous for it to work. He then explains why Astra 6.1 was held back: models are in the seriously capable era, the safety bar rises with capability, and the fix is a version OpenAI can still distribute widely with more confident claims about alignment.

06:47 The Trump Accord, Open Source and What Regulation Should Copy

Guiducci calls the accord a piece of paper with a typo on it. Altman says that is not quite fair and that more will be needed over time. He rejects the idea that a company is either for all regulation or against all of it: let ordinary things go wrong and be corrected by ordinary legal process, and put guardrails around the failures that cannot be undone. He says he did not ask Greg Brockman to withdraw a reported $25 million pledge to the Leading the Future super PAC and does not think Brockman opposes all AI regulation.

15:53 Xi’s State Visit, Extinction Risk and How He Wants to Be Remembered

Altman attended the state dinner for Xi Jinping but was not in the room for the substantive talks. Asked whether extinction risk is overhyped, he says any nonzero chance should be taken incredibly seriously, and that he assumes some people use the risk to justify concentrating power. Asked what role he wants in the history of AI, he says thinking that way is toxic, then answers anyway: that OpenAI’s work made AI something every person was empowered with, not something that happened to them.

Notable Quotes

“I do feel I’ve been mischaracterized. It’s only somewhat frustrating.”
Sam Altman, on being cast as the villain of AI

“I wish I had just taken the stock.”
Sam Altman, on holding no direct equity in OpenAI

“I really hope that hundreds of years from now, people look back at our lives and just can’t believe how bad we had it.”
Sam Altman, on why he will not stop technological progress

“I don’t like bullies and I think Elon’s a bully. And so I think once in a while you got to punch a bully back in the nose.”
Sam Altman, on his public fights with Elon Musk

“I would guess the world in general will be caught off guard by the next thing that goes really wrong.”
Sam Altman, in part two, on why AI policy will be written in reaction to failures

“You can come try making them all. They’re not so easy.”
Sam Altman, in part two, on critics of his individual safety decisions

Watch the full conversation on Fair Game by Vanity Fair: part one and part two.