The details of Silicon Valley’s most expensive “divorce case” have finally been fully revealed.
A few days ago, The Atlantic released the first excerpt from Kevin Roose, the bestselling American author and podcast host, new book, *The AGI Chronicles*. Roose interviewed over 150 insiders from OpenAI, Anthropic, and Google DeepMind, along with a wealth of previously unpublished company documents and internal memos, to piece together this most dramatic fallout in Silicon Valley history.
If you only know that Dario Amodei left OpenAI at the end of 2020 with a group of people to found Anthropic, you’ve only seen the tip of the iceberg. What’s truly astonishing is the power struggle he was involved in before leaving.
"The National Project": A Crazy Proposal to Auction AGI
Amodei once had high expectations for OpenAI. In 2016, he left Google to join this young nonprofit research organization. According to Roose, he was a key figure in steering OpenAI toward the large language model path, leading the team that contributed to the development of GPT-2 and GPT-3.
As research progressed, Amodei became increasingly convinced that larger models could lead to AGI—artificial general intelligence. Yet, at the same time, his confidence in the company’s leadership was steadily eroding.
One of the earliest things that troubled him happened in 2017.
At the time, Musk’s financial support for OpenAI began to waver, and the lab urgently needed to find a new path forward. Co-founders Greg Brockman and chief scientist Ilya Sutskever proposed an internal plan, later called the “countries plan.”
This plan initially seemed respectable, envisioning an international alliance to jointly manage AGI risks and share its benefits.
But by the time Brockman and Sutskever formally presented it to employees at an internal meeting, the tone had completely shifted: it had become a fundraising event “auctioning” the future rights to AGI technology to governments, with countries including the United States, China, and Russia specifically named.
Their logic is this: AGI involves national security, and governments around the world will eventually take it away from the companies that create it. Since it will have to be handed over anyway, it’s better to sell it for a good price now while cash is needed.
A knowledgeable attendee who heard the presentation complained: “It sounded like a few tech people pitching a crazy startup idea, ending with—just make the government pay for it.”
At the time, Dario Amodei was the Vice President of Research at OpenAI. After hearing this, he sent an email directly to Brockman, Sutskever, and other executives with decidedly blunt language. In the email, he wrote that the plan “would escalate rather than reduce an arms race,” could “lead to human rights atrocities”—that entrusting a regime like the Russian government with an AGI that embodies its values would enable them to commit “unimaginable acts” against their own people—and would “undermine OpenAI’s moral legitimacy.”
In plain terms: Are you serious about selling humanity’s most powerful technology to these countries?
Fortunately, Microsoft's subsequent large investment saved the day, and the plan was never implemented. OpenAI's response was that the company "briefly considered whether to transform OpenAI into an international collaboration," but "did not consider selling AGI to China or Russia."
The plan was never implemented, but Amodei didn’t forget that management had seriously discussed an idea he found so objectionable.
From Microsoft to Kanye, the trust gap is growing wider.
Afterward, even smaller things continued to deepen his mistrust.
During the GPT-3 beta, Amodei asked Altman for a list of approved customers using the model. Altman provided him with one, but Amodei claimed he later discovered another, longer list. He began to suspect that the information he had seen had been filtered.
And a dispute surrounding Microsoft's investment deepened his doubts.
One particularly unique provision in OpenAI’s public charter states that if a project aligned with its values and prioritizing safety comes close to achieving AGI, OpenAI will cease competing with it and instead offer assistance. This means that, under certain circumstances, the company’s mission should take precedence over winning the race.
According to Rose’s new book, during negotiations with Microsoft for a $1 billion investment, Amodei wanted to ensure that Microsoft accepted the possibility of an OpenAI acquisition as a means of fulfilling that commitment. Altman told him that Microsoft would not accept those terms.
Amodei doesn’t believe things are that simple. He suspects that Altman, in his haste to close the deal, may have given up on pressing his demands—or never raised them at all—and instead shifted the blame onto Microsoft.
In Amodei’s view, the issues were becoming increasingly difficult to resolve through business discussions: he had begun to doubt whether Altman, sitting across from him, was willing to fully disclose the details of the negotiations.
By 2020, Kanye West—the rapper who later publicly expressed admiration for Hitler—unexpectedly appeared on this growing list of grievances.
Roose cited two sources saying that after meeting Kanye, Altman invited him to join the private beta of GPT-3.
At the time, GPT-3 had not yet been publicly released and was OpenAI’s closely guarded core asset. Jack Clark, then OpenAI’s Director of Policy, intervened to block the proposal, noting that Kanye had recently posted a video疑似urinating on a Grammy award, raising concerns about his public behavior. Shortly after, Kanye began publicly expressing admiration for Hitler.
This invitation was not approved, further deepening Amodei’s dissatisfaction with Altman’s judgment. Roose wrote that individually, similar incidents might not be enough to prompt resignation, but collectively, they have made it difficult for him to believe Altman is suited to control such powerful technology as AGI.
Roose reveals in the book that by 2020, Amodei had become thoroughly disillusioned with OpenAI’s leadership, describing Altman as a “duplicitous manipulator” and Brockman as “power-hungry and reckless.”
The core contradiction is two words: trust.
It was reported that during his time at OpenAI, Amodei privately took notes documenting moments he believed Altman had misled him.
In one of his notes, he wrote: "The problem with OpenAI is Sam himself."
Factions and cliques?
On the other side, some within OpenAI have grown increasingly frustrated with Amodei.
According to Roosevelt's report, Amodei and his sister Daniela gradually brought together a group of like-minded colleagues. They created private chat channels, held separate meetings, and frequently discussed safety issues. To some employees outside this group, these individuals became increasingly difficult to work with.
A former colleague complained to Roose that Amodei often declared a choice morally correct, as if invoking safety should end the debate—frustrating other teams.
Amodei believed that business goals would always override security commitments; some colleagues felt he turned security into an irrefutable justification. Mutual dissatisfaction between the two sides spread from specific decisions to judgments about work styles and personal character.
Ultimately, the conflict centers on a tangible power: Can the security team veto a new product launch?
According to Roose, before his departure, Amodei presented Altman with a set of conditions for staying, including granting the safety team such a veto power. He hoped that the pressure of core researchers potentially leaving en masse would prompt Altman to accept changes.
In a private chat, a "split" is counting down
While negotiations are underway, backup plans have also been put in place.
Amodei and six colleagues created a private Slack channel called “BATNA,” which stands for Best Alternative to a Negotiated Agreement. The members included his sister Daniela, as well as Jack Clark, Sam McCandlish, Tom Brown, Chris Olah, and Jared Kaplan.
The name of this group itself reflects their situation at the time: still negotiating, but already prepared to leave.
When the retention terms could not be agreed upon, both parties brought in lawyers to begin negotiating what they internally called "the divorce."
According to Roosevelt's report, Altman agreed to grant this group a several-week window to recruit OpenAI employees, in exchange for a one-year pledge not to poach further; both parties also agreed not to publicly disparage each other.
In December 2020, OpenAI held a company-wide virtual meeting.
It started out just like usual. Altman reported on research progress, introduced new employees, and then handed the floor to Amodei.
Next, Amodei announced that he and six other leaders would leave to establish a new organization focused on AI safety.
According to Roose’s description, the chat window was quickly flooded with messages and emojis of surprise; as the meeting neared its end, Jack Clark typed a line into the chat: “I love you all,” and then the screen suddenly went black.
Seven core members of OpenAI, including Dario Amodei and his sister Daniela Amodei, were simultaneously removed from the meeting, with their email and Slack access immediately revoked. They gave this departure a blunt codename: the divorce.
In early 2021, Anthropic was founded.
When will this grudge end?
Starting from scratch also did not free Amodei from the problems that had troubled him.
Anthropic also needs funding, increased computing power, and the development of more powerful models. To study the most advanced risks, one must master cutting-edge technology; but to remain at the forefront, continuous investment is required, along with the challenge of competitors catching up. As the company grows, Amodei must also answer a familiar question: when safety, business, and competition conflict, which one should yield first?
The mutual resentment between both sides has not been resolved at all.
According to insiders, Amodei previously described Altman and OpenAI as "evil" in internal messages; meanwhile, some within OpenAI believe the Amodei siblings have portrayed themselves as moral exemplars while similarly craving power; they are particularly frustrated that Anthropic publicly warns about AI risks while simultaneously rushing to compete.
Roose describes these AI leaders in the book as “a web of mutual hostility”: Hassabis looks down on Altman, Amodei looks down on Altman, Altman looks down on Amodei, and Altman also looks down on Musk.
Years have passed, and the web hasn’t loosened—instead, each side has encountered new troubles. According to Bloomberg, Anthropic is seeking to go public as soon as mid-November, with a target valuation of up to $2 trillion; however, the prospectus also reveals another side: projected revenue of about $4.6 billion in 2025, yet an operating loss exceeding $8 billion, alongside commitments to invest over $500 billion in computing infrastructure. Meanwhile, Altman remains cautious about an IPO, stating that上市 is not urgent until safety and alignment are resolved—but OpenAI has recently been entangled in AI safety incidents as well.
Back then, Amodei left because he didn’t trust Altman, aiming to build a company where safety commitments truly mattered. Now, his own company faces the same challenge: growth must be fast, safety must be solid—neither can be compromised. Neither of them has turned in their answer yet.
Author: Seed.eth
Twitter: https://twitter.com/BitpushNewsCN
BitPush Telegram community: https://t.me/BitPushCommunity
BitPush TG subscription: https://t.me/bitpush