Insider details from OpenAI's core members come to light: Dario Amodei left over a crisis of trust to found Anthropic, revealing the palace-intrigue details of Silicon Valley's most expensive "breakup" and the battle over the path to AGI.
The details of Silicon Valley's most expensive "breakup" have finally been laid bare.
Recently, The Atlantic published the first excerpt from The AGI Chronicles , the new book by Kevin Roose, a well-known American bestselling author and podcast host. For the book, Roose interviewed more than 150 insiders at OpenAI, Anthropic, and Google DeepMind, and drew on a large volume of undisclosed company documents and internal memos to piece together this most sensational falling-out in Silicon Valley history.
If all you know is that Dario Amodei left OpenAI with a group of people at the end of 2020 to found Anthropic, that's only the tip of the iceberg. What's truly jaw-dropping is the "palace intrigue" that unfolded before he left.
Amodei once had high hopes for OpenAI. In 2016, he left Google to join this young nonprofit research institution. According to Roose's account, he was a key figure in steering OpenAI toward the large language model path, and the team he led took part in developing GPT-2 and GPT-3.
As the research progressed, Amodei became increasingly convinced that larger-scale models could lead to AGI—artificial general intelligence. But at the same time, his confidence in the company's leadership kept eroding.
One of the earliest things that stuck in his craw happened in 2017.
At the time, Musk's financial support for OpenAI was beginning to waver, and the lab urgently needed to find a new way forward. Co-founder Greg Brockman and chief scientist Ilya Sutskever proposed an internal plan that later came to be called the "countries plan."
The plan started out respectable enough, envisioning an international coalition in which everyone would jointly manage the risks of AGI and share in the benefits.
But by the time Brockman and Sutskever formally presented it to employees at an internal meeting, it had gone completely off the rails: it had turned into a fundraising campaign to "auction off" the future rights to use AGI technology to various governments, with the countries mentioned including the United States, China, and Russia.
The two men's logic went like this: AGI involves national security, so sooner or later governments would seize it from whichever company built it. Since it would have to be handed over anyway, they might as well sell it for a good price while the company was short on cash.
One person familiar with the presentation complained: "It sounded like a bunch of tech people pitching a crazy startup, and at the end they just said—let the government foot the bill."
Dario Amodei was OpenAI's vice president of research at the time. After hearing it, he fired off an email to Brockman, Sutskever, and other executives, in rather blunt terms. In the email, he wrote that the plan "would exacerbate rather than reduce an arms race," could "lead to human rights atrocities," and would let them do "unspeakable things" to their own people—and that it "would undermine OpenAI's moral legitimacy."
In plain terms: you want to sell humanity's most powerful technology to them? Have you lost your minds?
Fortunately, Microsoft's subsequent large investment saved the day, and the plan never materialized. OpenAI's response was that the company "briefly considered whether to turn OpenAI into an international cooperative endeavor," but "did not consider selling AGI to China or Russia."
The plan ultimately never came to pass, but Amodei did not forget that management had seriously discussed an idea he found so repugnant.
After that, a series of seemingly smaller matters kept deepening his distrust.
During GPT-3's private beta, Amodei asked Altman for the list of customers approved to use the model. Altman gave him one list, but Amodei claimed he later discovered another, longer list. He began to suspect that the information he was shown had been filtered.
And a dispute over the Microsoft investment deepened his misgivings.
OpenAI's public charter contains a rather unusual provision: if a value-aligned, safety-conscious project comes close to achieving AGI first, OpenAI will stop competing with it and instead assist it. This means that, under certain circumstances, the company's mission should take priority over winning the race.
According to Roose's new book, when negotiating a $1 billion investment with Microsoft, Amodei wanted to ensure that Microsoft would accept the OpenAI acquisition deal that might be used to carry out this commitment. Altman told him Microsoft would not accept the relevant terms.
Amodei did not believe it was that simple. He suspected that Altman, eager to close the deal quickly, might have given up on pushing for it—or never even raised the demand—while shifting the blame onto Microsoft.
In Amodei's eyes, the problem was becoming harder and harder to resolve through business discussions: he had begun to doubt whether the Altman sitting across from him was willing to give him a full account of the negotiations.
By 2020, Kanye West—the rapper who would later publicly express admiration for Hitler—unexpectedly appeared on this growing list of grievances.
Citing two people familiar with the matter, Roose reports that after meeting Kanye, Altman proposed letting him join GPT-3's private beta.
At the time, GPT-3 had not yet been publicly released and was OpenAI's crown-jewel core asset. Jack Clark, then OpenAI's policy director, stepped in to block the proposal, noting that Kanye had recently posted a video that appeared to show him urinating on a Grammy award, and that his public behavior was concerning. Then, not long after, Kanye began publicly expressing admiration for Hitler.
The invitation never went through, but it further deepened Amodei's dissatisfaction with Altman's judgment. Roose writes that, taken individually, most such incidents would not have been enough to make him resign, but cumulatively they had made it hard for him to believe Altman was fit to wield a technology as powerful as AGI.
Roose discloses in the book that by 2020, Amodei had become thoroughly disillusioned with OpenAI's leadership. He described Altman as a "two-faced manipulator" and considered Brockman "power-hungry and reckless."
In fact, the core conflict between them came down to two words: trust.
Amodei reportedly kept a private little notebook throughout his time at OpenAI, jotting down the moments when he felt Altman had misled him.
In one of those notes, he wrote: "OpenAI's problem is Sam himself."
On the other side, people inside OpenAI were also growing increasingly fed up with Amodei.
According to Roose's reporting, Amodei and his sister Daniela gradually drew in a group of like-minded colleagues. They set up private chat channels, held separate meetings, and frequently discussed safety issues. To some employees outside the circle, this group became increasingly difficult to work with.
One former colleague complained to Roose that Amodei would often declare a particular choice to be the morally correct one, as if merely invoking safety should end the debate. This infuriated other teams.
Amodei felt that commercial goals would always override safety commitments; some colleagues felt he had turned safety into an unanswerable justification. Their mutual dissatisfaction thus spread from specific decisions to working styles and judgments of character.
Ultimately, the conflict converged on one concrete power: could the safety team veto new product launches?
According to Roose, before leaving, Amodei presented Altman with a set of conditions for staying, including giving the safety team such veto power. He hoped that the pressure of core researchers potentially leaving en masse would push Altman to accept changes.
Meanwhile, even as negotiations were underway, an exit route was already being prepared.
Amodei and six colleagues set up a private Slack channel named "BATNA"—the best alternative to a negotiated agreement. Members included his sister Daniela, as well as Jack Clark, Sam McCandlish, Tom Brown, Chris Olah, a