16 Cloak-and-Dagger
By Tuesday, October 31, everyone had spoken with everyone else. Toner had spoken with McCauley. McCauley had called D’Angelo. Murati had talked to D’Angelo and McCauley.
That day, the three independent board directors—Toner, McCauley, and D’Angelo—began to meet nearly daily on video calls, agreeing that Sutskever’s and Murati’s feedback about Altman, and Sutskever’s suggestion to replace him, warranted serious deliberation. Sutskever, who was already firmly resolved in his conclusion to fire Altman, sat out of the discussions. Both he and the others felt the independent directors needed to arrive at their own conclusions without his influence. He also had a financial stake in the company, which they didn’t want to sway their decision-making.
The directors would later tell Sutskever a third reason: They had reached such low levels of trust with Altman that one of them wondered whether Altman had in fact sent Sutskever to the board to test their loyalty in order to push out anyone who moved against him.
The independent directors laid out what they knew: This was not the first time that senior leaders had described Altman in this way. In total, the three of them had heard similar feedback from at least seven people within one to two levels of Altman, inclusive of Sutskever, Murati, and Amodei, who oversaw safety and nonsafety parts of the company. Several had described Altman’s behaviors as abuse and manipulation; most had highlighted his lack of honesty and their inability to trust what he said. Then there were the myriad other issues that the independent directors themselves had found, including the disempowerment of the nonprofit; Altman not disclosing his legal ownership of the OpenAI Startup Fund; Altman neglecting to mention Microsoft’s DSB breach; Altman trying to force D’Angelo and now apparently Toner off the board.
They decided not to even touch Annie’s allegations. This was ultimately about Sam’s professional capacity as OpenAI’s CEO.
In that capacity, Murati had said that while some of Altman’s behaviors could be chalked up to typical tech CEO habits, they were still causing major problems. She’d made a comparison to Musk, whom she’d worked with at Tesla: Musk would make a decision and be able to articulate why he’d made it. With Altman, she was often left guessing whether he was truly being transparent with her and whether the whiplash he caused was based on sound reasoning or some hidden calculus. Just as he caused fragmentation in the Microsoft relationship, he caused fragmentation among his own leadership team, scattering information to different people but never giving any one of them the full picture, allowing him to retain full control. Combined with Brockman, the dynamic was disastrous. She disagreed with the Amodei siblings on many things, she’d said, but on this point, their observations had been correct.
And OpenAI was not, in fact, a typical tech company, the independent directors observed. It was arguably the world’s most powerful AI company, overseeing the development of what they felt was one of the most consequential technologies. A fear Sutskever had articulated resonated with them: What did it mean that OpenAI was trying to build AGI when its senior leadership couldn’t trust either basic or critical information coming from the CEO?
But here was another thought experiment: What if OpenAI were a typical tech company? What if it were just a grocery-delivery service like Instacart? Did Altman’s behaviors still warrant his removal? In some cases yes; in some cases no. But it also wasn’t clear that Altman would be the best person to continue running the company anyway, the independent directors thought. He was famous for startups, and OpenAI was rapidly maturing. Did he really have the skills and personality to continue charting the course—and to compensate for the instability he caused?
OpenAI was in a stellar position: It was a hot company. If the board went through a purposeful search process, they could have their pick of phenomenal CEOs with lots of experience running mature companies. Altman wasn’t necessarily essential to OpenAI’s operations, they reasoned; while he globe-trotted, Murati was the one doing the day-to-day heavy lifting within the company and had a strong relationship with Microsoft.
As the independent directors deliberated, Sutskever sent them a series of documents and screenshots that he and Murati gathered in tandem with examples of Altman’s behaviors. They came in long dossiers delivered via two disappearing emails, his icon a mysterious man in a hat. There was, as Sutskever had mentioned, no particularly damning evidence, but an accumulation of many instances of Altman saying different things to different people and stoking intense frustration across management. The screenshots showed at least two more senior leaders, both nonsafety and outside of the seven that the directors were already aware of, noting Altman’s tendency to skirt and ignore processes, whether instituted for AI safety reasons or to smooth company operations. This included, the directors learned, Altman’s apparent attempt to skip DSB review for GPT-4 Turbo by misquoting the legal team to Murati. The problem, which Murati had also raised, was how good Altman was at avoiding putting things in writing. He would deliver most of his communications verbally and wriggle out of agreements by telling other parties that they had simply misremembered what he’d said.
There were other things the independent directors needed to consider: How would Microsoft react? How would Brockman react? How would employees react? They debated whether to reach out to a third senior leader who Sutskever and Murati said had similar concerns, whether to conduct a more extensive fact-finding process, whether to loop in Microsoft’s executives. After discussion, they decided in each case that it would be better not to. Every new person they clued into the conversation and every new day they spent delaying a decision increased the chances that Altman would find out and, with his maneuvering, make it impossible for them to complete their deliberations.
On November 9, as the independent directors closed in on a final decision, Sutskever had another call with McCauley. “Sam said, ‘Tasha continues to be very supportive of having Helen step off the board,’ ” he told her. It was a balder-faced lie than Altman had told the first time; McCauley had not had any more exchanges with Altman.
The role of Toner’s paper, the directors later felt, would get significantly overplayed in the media, in part because, they were convinced, Altman might have fed it to reporters himself. On the second day of the five-day board crisis, the directors confronted him during a mediated discussion about the many instances he had lied to them, which had led to their collapse of trust. Among the examples, they raised how he had lied to Sutskever about McCauley saying Toner should step off the board.
Altman momentarily lost his composure, clearly caught red-handed. “Well, I thought you could have said that. I don’t know,” he mumbled.
The board directors marveled at his audacity. A few days later, Altman’s initial objections over Toner’s paper appeared in the media.
By Saturday, November 11, the independent directors had made their decision. As Sutskever suggested, they would remove Altman and install Murati as interim CEO. That day, they immediately told Sutskever, then continued to meet daily, just the three of them, with frequent check-ins with Sutskever, to finalize the paperwork for the leadership transition. On the night of Thursday, November 16, all four video called Murati. She picked up the call on her phone from a conference. Upon hearing the news, she looked surprised but receptive.
“He’s so, so paranoid right now,” she said. But did she feel comfortable with the decision? the directors asked. “Completely,” Murati said. She accepted the new role and expressed confidence in her ability to take the decision to the rest of leadership and to Microsoft. She would also loop in Wong, she told them, who could be trusted to help them write the announcement.
Hours later, after the group’s final sprint to finalize the messaging with Wong’s support, the most important thing left was to tell Altman.
In the tense final moments of waiting, none of the board directors fathomed the severity of their miscalculation.
Within hours of the public announcement on Friday, November 17, things had gone significantly south for the independent directors. After what had seemed like an initial period of calm and stability, including Murati having a productive conversation with Microsoft, she had suddenly called them with new stress. Altman and Brockman were telling everyone that Altman’s removal had been a coup by Sutskever, she said. Combined with Sutskever’s ineffectual communication during the employee all-hands, key stakeholders were beginning to turn on the decision.
Shortly thereafter, as the independent directors confronted a hostile leadership over a video call, they realized just how bad the sentiment was. Jason Kwon, chief strategy officer, and Anna Makanju, vice president of global affairs, were leading the charge in furiously rejecting their characterization of Altman’s behavior as “not consistently candid” and demanding evidence to support the board’s decision, which the directors felt they couldn’t provide without outing Murati. But even those in the room who they knew, based on the dossiers, had similar or other reservations about Altman’s leadership were remaining silent. As the night wore on and the hostility mounted, the independent directors’ two most important allies—or at least the two people they thought would be their allies—were beginning to veer off in a different direction.
That first night, faced with the visceral possibility of OpenAI falling apart, Sutskever’s resolve immediately started to crack. OpenAI was his baby, his life; its dissolution would destroy him. While he fiercely stood by everything he’d said about Altman, his intention had been to strengthen OpenAI, not dismantle it. He was shocked, hurt, and disoriented by the reactions of employees and his fellow leaders; this was not an outcome he had anticipated. He began to plead with his fellow board members, and would continue to plead with increasing agony through the weekend, about whether they needed to reconsider their position.
Even more complicated, Murati was also acting differently than the directors had expected. As the negotiations with leadership unfolded that night, she kept in touch over calls to privately relay them information. Yet despite the confidence she had expressed to get people’s buy-in, she now remained unwilling to explicitly throw her own weight behind the board’s decision. In the room with other leadership, during their series of escalating face-offs with the board, she at times even said things that made her appear as if she were just as confused as everyone else about what exactly was happening.
To Murati, the intensifying revolt of her fellow leaders and employees seriously challenged her position as interim CEO. In rapid succession that Friday, Brockman had quit in protest, then his allies Pachocki and Szymon Sidor, along with Aleksander Mądry, the professor on leave from MIT who is also Polish and close with Pachocki. It had led to a conflagration of anger not only within the company but that was also spreading to a growing circle of investors that increasingly made her doubt her ability to effectively hold together and continue to lead the organization. She began to waver on her commitment to take over the role. While she supported the directors’ decision, she had not been part of their deliberations. If they wanted her to have any chance of succeeding at taking the reins, she felt, it was now their burden to bear to justify their decision to the company first.
No longer certain about whether they could rely on Murati, the three independent directors pushed ahead with searching for a new interim CEO or new board members. D’Angelo, in particular, the board’s only Silicon Valley insider, made dozens of calls through Saturday and Sunday, putting out as many feelers as possible to his sprawling network. At one point, the directors called Dario Amodei about the interim chief position. Amodei wasn’t interested. But to others that weekend, he seemed almost giddy with excitement about the overall situation.
On Sunday, D’Angelo finally found someone to take the board up on one of its offers: Emmett Shear, the cofounder of Twitch, appeared willing, and cooperative, to temporarily take over the company. But soon enough he, too, began to deviate for seemingly inexplicable reasons. The Wall Street Journal would later report that Shear, who had been YC batchmates with Altman, was also a friend and mentor of YC alumnus Airbnb cofounder Brian Chesky. All weekend, Chesky, among Altman’s most trusted friends, had worked the phone lines along with Reid Hoffman, a Microsoft board member, to calm investor nerves, align Microsoft’s messaging, and mount the pressure further. Chesky quickly reached Shear, who then sided with Altman.
By late Sunday night, after reassurances from Chesky and Hoffman, Microsoft had also thrown its weight behind Altman, with Nadella announcing that Altman and Brockman were joining to lead a new advanced AI research division. Other AI labs were circling OpenAI like vultures, intent on poaching away their share of talent in the carnage. It became clear to Murati that with the board unable to legitimize their decision, the dynamics now threatened to leave behind a shell of a company. She fell on the side of her fellow leaders. To salvage the situation, Murati wanted Altman back.
Overnight, as employees put together the open letter protesting the board’s decision and threatening to quit and join Microsoft, Murati put her name first. Many senior employees were more loyal to Murati than to Altman. Murati was the one in the trenches with them day in and day out. Murati was the one whom they trusted to act not in her self-interest but in the best interests of the company. Others also believed that Altman’s close relationship with investors and singular ability to fundraise massive amounts of capital made him the best, if not only, person to keep OpenAI’s ongoing tender offer on track, promising them a chance to cash in as much as millions of dollars of their equity, as well as to secure the finances for the company’s long-term success. Still others were keenly aware of Altman’s uniquely expansive network and ability to make people’s careers in Silicon Valley. With the key executives and senior staff bought in, the signatures on the letter rapidly snowballed. In the wee hours of the morning, Sutskever, who could no longer see another path forward without the company collapsing, added his name to the list. “Without Mira, I don’t think Sam would have been able to pull off what he did,” a researcher says. To the independent board directors, her equivocating felt like a self-fulfilling prophecy.
The New York Times would later break the story of Murati’s role in the ouster. To employees, Murati would defend herself. “Sam and I have a strong and productive partnership and I have not been shy about sharing feedback with him directly,” she wrote. “I never reached out to the board to give feedback about Sam. However, when individual board members reached out directly to me for feedback about Sam, I provided it—all feedback Sam already knew.”
By the Monday morning of the crisis, the independent directors knew they had lost. Murati and Sutskever had flipped sides, and the employee protest letter made clear the destabilization at the company had become untenable. Altman would come back; there was no other way to preserve OpenAI. The one silver lining: After holding out long enough, Altman also seemed ready to make concessions. With that, the independent directors switched their focus to saving what mechanisms they could for continuing to hold him accountable: to keep at least one of them on the board for continuity, to find two more independent directors who could truly be independent against Altman, and to get Altman to submit to an investigation.
But as they neared a final agreement, there was one more person who wanted to stir the pot. It was none other than OpenAI’s spurned former cochairman, Elon Musk.
All through the weekend, X had become the breeding ground for every possible theory and conspiracy theory about the OpenAI board drama. It had also been a busy weekend for Musk. He had overseen a SpaceX launch and finalized a lawsuit against the nonprofit Media Matters for its report on antisemitic content on X. He had then turned to his platform, which he’d bought in an ill-conceived deal in April 2022, to defend his record against antisemitism, to assert free speech absolutism, and to criticize the woke mob and the mainstream media. In between it all, he was reply guying to other people’s commentary and memes, and at times tweeting himself, about OpenAI and Altman.
“I am very worried,” he wrote on Sunday, November 19, in the afternoon. “Ilya has a good moral compass and does not seek power. He would not take such drastic action unless he felt it was absolutely necessary.” Later, at around 2:00 a.m. Pacific time, he provocatively tweeted out a YouTube clip of the famous baptism scene from The Godfather, when Michael Corleone transforms from a morally conflicted son to a ruthless new don by murdering the heads of all the other families.
But on Tuesday afternoon, Musk sought to tip the scale more explicitly. “This letter about OpenAI was just sent to me,” he tweeted, with a link. “These seem like concerns worth investigating.” It was a different letter than the one circulating among current employees but was also addressed to OpenAI’s board of directors. It began:
We are writing to you today to express our deep concern about the recent events at OpenAI, particularly the allegations of misconduct against Sam Altman.
We are former OpenAI employees who left the company during a period of significant turmoil and upheaval. As you have now witnessed what happens when you dare stand up to Sam Altman, perhaps you can understand why so many of us have remained silent for fear of repercussions. We can no longer stand by silent.
The letter presented a series of demands—chiefly, for interim CEO Emmett Shear to expand his investigation into Altman’s behaviors to include OpenAI’s earlier history and its corporate restructuring away from the nonprofit. “We believe that a significant number of OpenAI employees were pushed out of the company to facilitate its transition to a for-profit model,” it said. It also presented a series of allegations, describing “a disturbing pattern of deceit and manipulation by Sam Altman and Greg Brockman.”
At the bottom of the letter was a section called “Further Reading for the General Public,” which listed three links. One was an X thread from Geoffrey Irving, the AI safety researcher who had left in 2019 for DeepMind, saying that Altman had “lied to me on various occasions” and “was deceptive, manipulative, and worse to others.” The other two were journalism articles. Both of them were mine.
By the time I saw it, Musk’s tweet had already gained over ten thousand retweets and several times more likes. I found the reading selection surprising: There had been plenty of pieces written about OpenAI. Why had they chosen to link only the two written by me? The first was my 2020 OpenAI profile for MIT Technology Review; the second was a piece I had just written with my colleague Charlie Warzel for The Atlantic, providing context to the board crisis with a window into the ideological polarization that had inflamed within the company after ChatGPT.
Above the section, there was a Tor email for exchanging encrypted messages. “We encourage former OpenAI employees to contact us,” it read. I wondered if, by picking my pieces, the authors of the letter were trying to reach me. I created my own Tor email and typed up a message:
Hi—I am the journalist who wrote both pieces linked in your note, and I believe you are trying to get in touch.
I added my contact information and pressed send. Within minutes, my Signal lit up with a notification: Someone had responded.
In one of the strangest reporting experiences of my career, the person who responded was just as confused as I was about what was happening. He was also a former employee at OpenAI, but not one who had been involved in writing the letter. He had simply received a copy of the letter in his personal email with zero explanation; when he tried to follow up with questions, he received another mysterious response: a link to the Tor inbox with the phrase capped_profit, which seemed to be a username, followed by what looked like a password.
After testing out the credentials and successfully logging into the inbox, he saw many other emails to former employees with the body of the letter in the account’s Sent folder. As he continued to look around, media inquiries came pouring in, including from The New York Times, The Washington Post, and The Information. Other emails were filtering in from people identifying as current and former employees. One read:
current employee here.
have worked directly with leadership
your message resonates with me.
what is your plan?
He found that one particularly funny. Though there was no reason to believe he was actually behind the message, Altman was known for always writing in lowercase letters.
The former employee began to take screenshots of everything. There was at least one other person logged in to the account at the same time. At various points during his sifting, inbound emails would populate with responses. He had no skin in the game, he told me. But what he didn’t like was that it seemed as if these former employees had referenced in their allegations the experiences of others without their consent; many of the allegations, he felt, had also been distorted to fit a particular narrative. He responded to my email because he recognized my name. He sent me all of the screenshots.
As I went through them, a few stood out. In response to various journalists, there were several emails that said the letter had been prematurely posted and had gone viral before it was ready. One also responded to a reporter with a specific number of coauthors: “At the moment of its publication, 13 individuals had contributed to its writing.”
Then in the Drafts folder, there was an email that hadn’t yet been sent and wasn’t meant to be. It was a message intended for just those logged in to the inbox.
SUBJECT LINE: Cease contact with media
Elon’s involvement has rendered this into a conspiracy-theorist character attack piece.
Let’s refocus on privately and individually communicating with the board to indicate that there’s plenty of available evidence to indicate the great lengths the leadership team went to ensure the for-profit future of OpenAI from early on.
Sign when read.
-1
-2
-3
-4
-5
The group likely couldn’t have known that the negotiations would end only hours later and OpenAI would announce Altman back as CEO. But in their final coordinated attempt to keep Altman out, I had somehow accidentally stumbled into their temporary control center. From the phrasing of the letter, many would surmise to me, as would I, that it was likely written by former employees in OpenAI’s Safety clan, known for harboring some of the harshest opinions about Altman and the unraveling of the nonprofit. After Altman’s return, it was people in Safety as well who would feel most betrayed by the board, believing that the fiasco of the ouster had delivered the biggest blow yet to their fight to yank back OpenAI’s trajectory toward the original Doomer-rooted ethos of the nonprofit.
But no one could—or would—tell me which people had written the letter. And I never found out who they were.
On December 6, 2023, two weeks after the board crisis, OpenAI employees gathered for an all-hands at San Francisco’s Palace of Fine Arts Theatre, an architectural landmark with an open rotunda and stone columns reminiscent of the Greco-Roman empire. Several members of leadership, including Murati, who was back to being CTO; Bob McGrew, who was now chief research officer; COO Brad Lightcap; and Jason Kwon, delivered an update on their division’s 2024 plans.
Altman looked despondent, in a way employees had never seen him before. Sutskever was also notably absent, which Altman addressed directly. “Look, I know people are sad Ilya is not here. I am sad too,” he said. In Sutskever’s place was Pachocki, who argued in a halting, stuttering talk that OpenAI’s latest research advancements brought the company closer than ever to building Turing’s decades-old dream of thinking machines.
During the board crisis, one media report in particular had sparked a fresh wave of frenzied speculation: a Reuters article stating that the directors had received a letter from employees days before Altman’s firing about a supposedly new research breakthrough, an algorithm called Q*. Q* had not factored into the board’s decision. But, as Pachocki alluded to in his talk, Research was indeed treating the new algorithm with intense importance.
The algorithm had been a brainchild of Sutskever’s, rooted in research he had been developing since 2021 to advance OpenAI’s models without a need for more data. The idea was to get a deep learning model to make better use of its existing data by using more compute at inference time to deliver better results. It broke the logic of the original scaling laws, which tied together a model’s performance with three inputs used during training: data, parameters, and compute. With this new method, Sutskever hoped to improve a model’s performance by further scaling the one ingredient he had always believed to be the most important: compute—but inference compute rather than training compute; in other words, the compute used to generate the model’s responses. He would later explain his thinking behind this approach during a keynote at NeurIPS in December 2024, after having one of the papers he’d coauthored win a Test of Time Award for the third year in a row. “While compute is growing through better hardware, better algorithms, and larger clusters,” he would say, “the data is not growing because we have but one internet.”
OpenAI’s Research division believed Q* would finally allow the company to develop models with stronger reasoning abilities—that critical elusive ingredient to unlocking AGI. Q* was so important, in fact, that after the project leaked, leadership implemented its most aggressive strategy to clamp down on yet more leaks to the media. They siloed the company completely, splitting Research into its own Slack group, restricting access to all Q*-related Google docs, and renaming the project Strawberry. The renaming was an attempt to make it harder for outsiders to recognize and track internal projects if they ever overheard OpenAI researchers talking. In a similar way, after the Arrakis project leaked to The Information, desert names for models were largely abandoned.
The frenetic Q* discourse and OpenAI’s reaction were a strange demonstration of how much the foundations of scientific inquiry in the AI field had eroded. Science is a process of consensus building. The significance of any advance—whether in AI or otherwise—tends to be highly subjective the moment that it happens. Only through peer review, the test of time, and sustained impact does a particular advance become elevated to “a breakthrough.” With OpenAI performing its work in secrecy—and the rest of the industry now following—the “breakthrough” label could really only be treated as a matter of the company’s opinion.
By the following all-hands in January 2024, Altman seemed mostly back to his old self. He discussed with new energy the plans for the first half of the year, including beginning to train what he hoped could become GPT-5 in Arizona. The project was code-named Orion, after the constellation. Soon, internal memes spawned about Orion looking like the word onion. A smaller model in the Orion series was subsequently named Scallion.
A month later, it was as if The Blip—as employees began to call it—had never happened. OpenAI teased Sora, a new video-generation model, built on diffusion, explaining in its blog that video was an even better way than images of developing complex multimodal models. Research was continuing its progress with Strawberry and AI Scientist and advancing other methods for improving compute efficiency. Applied was back to rapidly prototyping different product ideas.
On March 8, 2024, the new board’s investigation into Altman officially concluded. Toner’s and McCauley’s replacements, directors Larry Summers and Bret Taylor, oversaw the process. In 2022, Taylor had played a critical role in a different corporate drama, brokering the final sale of Twitter to Musk as its board chairman, which included Taylor leading a lawsuit against Musk to force the deal through after the South Africa-born billionaire attempted to ditch the original agreement. When the sale was completed, the Twitter board dissolved. Soon after, Taylor cofounded an AI agent startup, Sierra, that would fast become one of the most highly valued AI startups by the fall of 2024.
For the OpenAI investigation, Summers and Taylor hired the law firm WilmerHale to conduct the independent review, during which it said it pored over more than thirty thousand documents and conducted dozens of interviews with the previous board members, executives, and other relevant people and scoped the examination to how the board made its decision to fire Altman. The resulting report was never released to the public or employees. Summers would tell people privately that the investigation had found many instances of Altman saying different things to different people, but to a degree that the new board decided didn’t preclude him from continuing to run the company; it was thus not worthwhile to release any details to sow doubt about Altman’s leadership and risk breaching the confidentiality of people whose testimonials had contributed to the report.
In a blog post, Taylor, now OpenAI’s board chairman, released a statement with a resounding vote of confidence. “We have unanimously concluded that Sam and Greg are the right leaders for OpenAI,” he wrote. Altman would return to the board, and three new independent directors were being added: Sue Desmond-Hellmann, former CEO of the Bill & Melinda Gates Foundation; Nicole Seligman, former EVP and global general counsel of Sony and former president of Sony Entertainment; and Fidji Simo, the CEO and chair of Instacart.
That night, Toner and McCauley released a statement of their own. “Accountability is important in any company, but it is paramount when building a technology as potentially world-changing as AGI,” they wrote. “We hope the new board does its job in governing OpenAI and holding it accountable to the mission. As we told the investigators, deception, manipulation, and resistance to thorough oversight should be unacceptable.”
Among many employees, the conclusion delivered the final assurance they needed. The crisis was over. And then, suddenly, it was not.