Written by: Bitpush
The most expensive "breakup case" in Silicon Valley, the details are finally revealed.
Recently, The Atlantic Monthly published the first excerpt from the new book "The AGI Chronicles" by Kevin Roose, a well-known American best-selling author and podcast host. For this book, Roose interviewed more than 150 insiders from OpenAI, Anthropic and Google DeepMind, plus a large number of undisclosed company documents and internal memos, to piece together the most sensational breakup in Silicon Valley.
If you only know that Dario Amodei left OpenAI with a group of people at the end of 2020 and founded Anthropic, that is just the tip of the iceberg. What really shocked people was the "palace fight" before he left.
“Project National”: A Crazy Proposal to Auction AGI
Amodei once had high expectations for OpenAI. In 2016, he left Google to join the young nonprofit research organization. According to Roose's narrative, he was an important figure in pushing OpenAI towards the large language model route, and the team he led participated in the development of GPT-2 and GPT-3.
As research progresses, Amodei becomes increasingly convinced that larger-scale models may lead to AGI—artificial general intelligence. But at the same time, his confidence in the company's leadership continued to erode.
One of the first things that bothered him happened in 2017.
At that time, Musk's financial support for OpenAI began to waver, and the laboratory urgently needed to find a new way out. Co-founder Greg Brockman and chief scientist Ilya Sutskever came up with an internal plan that became known as the “countries plan.”
This plan was quite decent at the beginning. The idea was to form an international alliance and everyone would jointly manage the risks of AGI and share the benefits.
But by the time Brockman and Sutskever formally presented it to employees at an internal meeting, the picture had gone completely awry: It had turned into a fundraising event to "auction" the rights to use future AGI technology to various governments, including the United States, China and Russia.
The logic of the two people is this: AGI involves national security, and sooner or later governments will snatch it away from the company that created it. Since you have to hand it over anyway, it’s better to sell it at a better price while you’re short of money.
An insider who heard the demonstration complained: "It sounded like a few tech people selling a crazy entrepreneurial project, and at the end of the sentence - just let the government foot the bill."
Dario Amodei was OpenAI’s vice president of research at the time. After he heard this, he sent a rather blunt email directly to Brockman, Sutskever and other executives. He wrote in an email that the plan "will exacerbate rather than reduce the arms race," could "lead to human rights atrocities" that would allow them to do "unspeakable things" against their own people - and "would undermine OpenAI's moral legitimacy."
Translated into human words: Do you want to sell mankind's most powerful technology to them? Go crazy.
Fortunately, Microsoft's subsequent large investment saved the day, and this plan ultimately did not come to fruition. OpenAI's response was that the company "briefly considered whether to turn OpenAI into an international collaboration," but "did not consider selling AGI to China or Russia."
The plan never came to fruition, but Amodei never forgot that management had seriously discussed an idea that he found so objectionable.
From Microsoft to Kanye West, the trust rift grows deeper
Since then, seemingly smaller things have continued to deepen his distrust.
During the GPT-3 internal beta, Amodei asked Altman for a list of customers who were allowed to use the model. Altman gave him one, but Amodei claims he later found a longer list. He began to suspect that the information he saw had been filtered.
A dispute surrounding Microsoft's investment deepened his doubts.
There is a rather special agreement in OpenAI's public charter: if a project with consistent values and emphasis on security approaches the realization of AGI first, OpenAI will stop competing with it and instead assist the other party. This means that, in certain situations, the company's mission should take precedence over winning the game.
According to Roose's new book, when negotiating a $1 billion investment with Microsoft, Amodei wanted to ensure that Microsoft accepted the OpenAI acquisition that might be used to implement this commitment. Altman told him that Microsoft did not accept the conditions.
Amodei doesn't believe it's that simple. He suspected that in order to facilitate the deal as quickly as possible, Altman may have given up the fight, or even made no request at all, but passed the responsibility to Microsoft.
In Amodei's eyes, the problem is becoming increasingly difficult to resolve through business discussions: he has begun to doubt whether Altman, who is sitting across from him, is willing to explain the complete negotiation process to him.
In 2020, Kanye West—the rapper who later publicly expressed his admiration for Hitler—accidentally appeared on this ever-accumulating list of grievances.
Roose quoted two people familiar with the matter as saying that Altman offered to let him participate in a private beta test of GPT-3 after meeting Kanye.
At that time, GPT-3 had not yet been publicly released, and it was the core asset of OpenAI. Jack Clark, the policy director of OpenAI at the time, stopped the proposal. He mentioned that Kanye had recently released a video that appeared to be urinating on the Grammy Award, and his public behavior was worrying. Then, it didn’t take long for Kanye to openly express his admiration for Hitler.
This invitation failed, but it further aggravated Amodei's dissatisfaction with Altman's judgment. Roose wrote that most of the similar incidents alone were not enough to make him resign, but the accumulation made it difficult for him to believe that Altman was suitable to master such a powerful technology as AGI.
Roose disclosed in the book that by 2020, Amodei had become completely disappointed with OpenAI’s leadership. He described Altman as a "two-faced manipulator" and believed that Brockman was "power-hungry and reckless."
In fact, the core conflict between them is just two words: trust.
Amodei is said to have kept a private journal during his tenure at OpenAI, writing down moments when he felt Altman had misled him.
In one note, he wrote: "OpenAI's problem lies with Sam himself."
Forming gangs and small groups?
On the other hand, there are also people within OpenAI who are increasingly unable to stand Amodei.
According to Roose's report, Amodei and his sister Daniela gradually gathered a group of colleagues with similar ideas. They set up private chat channels, held separate meetings, and frequently discussed security issues. From the perspective of some employees outside the circle, this group of people is becoming increasingly difficult to work with.
A former colleague complained to Roose that Amodei often decided that a certain choice was morally right, as if just citing safety should end the debate. This annoyed the other teams.
Amodei feels that business goals will always trump safety commitments; some colleagues believe that he has turned safety into an argument that is difficult to refute. The dissatisfaction between the two sides spread from specific decisions to working methods and personality judgments.
Ultimately, the conflict centered on one real power: Can the security team veto a new product launch?
According to Roose, Amodei proposed a set of conditions to Altman for staying on before leaving, including giving the security team such veto power. He hopes that the pressure of core researchers possibly leaving en masse will prompt Altman to accept change.
In the secret group chat, the countdown to a "family separation" begins
While the negotiations were ongoing, a retreat was also being prepared.
Amodei and six colleagues set up a private Slack channel called "BATNA," which stands for Best Alternatives When Negotiations Can't Reach an Agreement. Members include sister Daniela, as well as Jack Clark, Sam McCandlish, Tom Brown, Chris Olah and Jared Kaplan.
The name of this group itself illustrates their situation at that time: still talking, but ready to leave.
When the terms of retention could not be negotiated, both parties hired lawyers and began negotiating what was internally called the divorce.
According to Roose's report, Altman agreed to give these people a few weeks to recruit OpenAI employees, in exchange for not poaching people within a year; the two parties also agreed not to disparage each other in public.
In December 2020, OpenAI held an online meeting for all employees.
The beginning looks the same as usual. Altman reported on the progress of the research, introduced new employees, and then handed the message over to Amodei.
Next, Amodei announced that he and six other leaders would leave to establish a new organization focused on AI security.
According to Roose's description, the chat window was quickly filled with surprised messages and expressions. At the end of the meeting, Jack Clark typed a line in the chat box - "I love you all", and then the screen suddenly turned black.
Seven core members of OpenAI, including Dario Amodei and his sister Daniela Amodei, were kicked out of the meeting at the same time, and their email and Slack permissions were revoked on the spot.
Then in early 2021, Anthropic was established.
When did the grudge end?
And starting a new business did not allow Amodei to get rid of the problems that had troubled him.
Anthropic also needs to raise funds, expand computing power, and develop more powerful models. Only by mastering cutting-edge technology can we study the most cutting-edge risks; but to stay at the cutting-edge, we must continue to invest and face the pursuit of competitors. As the company grows, Amodei must answer a familiar question: When security, business and competition collide, which one gives way first?
The mutual resentment between the two sides has not disappeared at all.
People familiar with the matter said that Amodei once described Altman and OpenAI as "evil"; at the same time, some people within OpenAI also believed that the Amodei siblings packaged themselves as moral models, but were equally keen on power; they were particularly dissatisfied with Anthropic's public warning of AI risks while accelerating its participation in the competition.
In the book, Roose describes these AI leaders as "a spider web of entangled hostility": Hassabis looks down on Altman, Amodei looks down on Altman, Altman looks down on Amodei, and Altman looks down on Musk.
A few years have passed, but this network has not been untied. Instead, each of them has new troubles. According to Bloomberg, Anthropic is seeking to go public as soon as mid-November, with a target valuation of up to US$2 trillion; but the prospectus also shows another side: revenue in 2025 will be approximately US$4.6 billion, but operating losses will exceed US$8 billion, and it has promised to invest more than US$500 billion in building computing power in the future. Altman is cautious about the IPO, saying that there is no rush to go public before security and alignment are resolved, but OpenAI has also been plagued by AI security incidents recently.
Amodei left Altman because he didn't trust him, and wanted to build a company whose security promises truly count. Now his own company has to face the same problem: sustainability of growth + safety concerns, neither side can be loosened. They haven't handed in this question yet.