# The Gambit LLMS index: [llms.txt](/en/llms.txt) --- Four days after Weil’s New York magazine profile of Altman and four days before Sutskever’s message to Toner, the independent board director had met with another OpenAI executive: chief technology officer Mira Murati. Born in Albania, Murati had learned from a young age how to stay calm amid chaos. Through her early childhood, she had experienced the throes of the country’s transition from totalitarian communism to liberal capitalism. The shift happened so rapidly, with the country’s financial system so underdeveloped, that pyramid schemes rapidly proliferated, then collapsed, leading to widespread unrest and violence. The upheaval would leave behind bomb craters that Murati would need to delicately maneuver around on her way to school. A teacher once told her that as long as she was willing to do it, the teacher would do it too. Murati’s parents taught literature, but she found solace in the certainty of numbers. First came her love of math, nurtured by teachers who saw her potential and at times gave her harder problems than the curriculum to push her learning faster. Then came her love of science—chemistry, biology, physics— which fed her love of technology. She was a voracious learner. She burned through any book she could get her hands on, finishing up her own textbooks and then rummaging through her older sister’s. She thrived on competition, finding her happy place in math and science Olympiads as she jostled among peers in the fierce race for knowledge. When she was sixteen, her precociousness won her a scholarship to study abroad at a Canadian private school, Pearson College UWC in Victoria, British Columbia. The opportunity set her on a rapidly rising trajectory, through Dartmouth College, where she studied mechanical engineering, to an aerospace company, to her first major career break at Tesla, where she was a senior product manager on the Model X. At Tesla, she learned to build complex products under intense pressure, navigating and negotiating across teams with very different opinions and areas of expertise. It was there, she often says, that she found herself drawn to AI as the company explored autonomous driving. The more she dug into AI, the more she began to view it as a fundamental asset that would be broadly applicable and universally needed for solving tough problems. “It really seemed like maybe the last thing we’d ever work on,” she later told Kevin Scott on his podcast. Murati didn’t jump into AI immediately. Three years into Tesla, she left to join the company Leap Motion as the vice president of product and engineering, to work on augmented and virtual reality systems. She imagined the company revolutionizing education, allowing learners to rotate strands of DNA with a swivel of their hands or manipulate the physics of a ball hurtling through the air. Instead the company was too early a bet on the technology; VR and AR still made too many people nauseated. In 2018, two years later, she left for OpenAI while it was still a nonprofit. At OpenAI, Murati’s climb continued. As the nonprofit transformed into a commercial operation, she stepped naturally into VP of Applied and partnerships. She oversaw the company’s most important relationship with Microsoft and the budding, then burgeoning, division commercializing the company’s research. She was even younger than Altman, Sutskever, and Brockman, and, for a while, the only technical woman in senior leadership. This sometimes made her the target of sexism, particularly among researchers nostalgic for the early days of the lab who viewed her as not technical enough— an engineer rather than a scientist—and considered her rise as a symbol of OpenAI’s turn away from serious fundamental research. Indeed, in the awkward, nerdy, testosterone-fueled world of AI, Murati stood out. She was socially adept, a good listener, and had little ego to speak of. She was known among the people she worked with as a uniquely skillful problem solver. In the maelstrom of OpenAI’s persistent internal conflict, she could guide the company forward, attentive to different ideas and perspectives yet unafraid of making tough calls. “Imagine when there’s excruciatingly hard decisions that have to be made and there’s no clear answer. And she can just help find an answer,” says a former colleague. “She’s consistently correct.” --- Among the many hats Murati wore, she increasingly played translator and bridge to Altman. After the Anthropic split, Altman had asked her to oversee not just Applied but also Research; in May 2022, she officially took on the title of chief technology officer. As more and more teams rolled up to report to her, she became a critical conduit through which employees interfaced with Altman. If he had adjustments to the company’s strategic direction, she was the implementer. If a team needed to push back against his decisions, she was their champion. Even among company leadership, she had a level of influence on Altman and access to his opinions that others did not. She could tell him directly when his expectations or plans were unrealistic, and he would often listen. She would tell others directly if he didn’t want something, even when he pretended that he did. Where people grew frustrated with their inability to get a straight answer out of Altman, they sought her help to decode his opinions. “She was just honest,” another former colleague says. “She was the one getting stuff done.” But the more Murati worked with Altman, the more she found herself frequently cleaning up his messes. If two teams disagreed, he often agreed in private with each of their perspectives, which created confusion, exacerbated the conflict, and bred mistrust among colleagues. That pattern compounded the chaos that Brockman continued to cause as he jumped into projects. To Murati, Brockman was like a second CEO but a bad one—highly opinionated and prone to driving people to burnout with his intensity. During the development of GPT-4, Altman and Brockman’s dynamic had exerted mission-critical levels of stress on parts of the company, nearly leading key people on the pre-training team, one of the core teams handling the data collection and initial training of each model, to quit. Then there was Nadella, who was practically OpenAI’s third CEO with how deferential Altman could be to Microsoft’s interests. On multiple occasions, after Murati had carefully put together a plan of reasonable commitments that OpenAI could make to Microsoft, Altman had veered off script with the tech giant’s executives, conveying a different picture of what the team was working on to agree with demands divorced from the startup’s road map. To the board, Altman framed the tumult differently: Murati just didn’t have a productive relationship with OpenAI’s most important partner. Altman’s behavior had progressively worsened after ChatGPT had propelled him into megastardom, intensifying both the spotlight and scrutiny and exploding his calendar with an overwhelming travel schedule. Before, he was generally energized; now he was often exhausted. And he was cracking under that pressure, his anxiety reaching new heights and fueling his patterns of destructive behavior. He was doing what he’d always done, agreeing with everyone to their face, and now, with increasing frequency, badmouthing them behind their backs. It was creating greater confusion and conflict across the company than ever before, with team leads mimicking his bad form and pitting their reports against each other. This was corroding enough as it was. But faced with mounting competition externally, Altman was also pushing the company to deploy faster and faster and attempting to skirt some of its established release processes for expediency, sometimes through dishonesty. Recently, he had told Murati he thought that OpenAI’s legal team had cleared GPT-4 Turbo for skipping DSB review. But when Murati checked in with Jason Kwon, who oversaw the legal team, Kwon had no idea how Altman had gotten that impression. In the summer, Murati had attempted to give Altman detailed feedback on the accelerating issues, hoping it would prompt self-reflection and change. Instead, he had iced her out, and it had taken weeks for her to thaw the relationship, including by assuring him that she had not shared that feedback with anyone else. She had seen him do something similar with other executives: If they disagreed with or challenged him, he could quickly cut them out of key decision-making processes or begin to undermine their credibility. It was subtle and contained enough, out of sight of employees, that it had taken her some years to realize the full extent of it. But inevitably, different executives had each had their turn bearing the brunt of this treatment. Over time, the cumulative impact of his actions had taken its toll on the highest levels of the organization. Most recently, the hot seat had passed to Sutskever. Some time earlier, Jakub Pachocki, the Polish researcher leading the AI Scientist project, who reported to Sutskever, had grown frustrated with his lack of recognition or authority. He’d turned to his ally, Brockman, and Brockman had turned to Altman. Altman had then encouraged Pachocki’s ambitions and given him a more senior role in Research. There was only one problem: Altman had never mentioned any of this to Sutskever. Nor would he clear up the divisions between Sutskever’s and Pachocki’s portfolios as both of them, each getting different messaging from Altman, began guiding the same research in their own directions and struggling to understand the source of the misalignment. The tangled situation had caused several months of organizational thrash in the Research division. It was now, just as with the GPT-4 pre-training team crisis, reaching untenable levels of stress. For Sutskever, the ongoing saga was deeply painful. Not only was it a humiliating snub from Altman, it had unraveled his friendship with Pachocki, cultivated over years of late nights, high highs, and low lows, working side by side to build up the company. Murati was once again working overtime to find a solution. It had eaten up significant amounts of her time to simply figure out what was happening. Now she had to get Sutskever and Pachocki to agree on an arrangement, get Brockman to stop putting his thumb on the scale by petitioning Altman, and get Altman to stay on message and stop contradicting her in private meetings. But none of those would solve the root of the problem. It would only be a matter of time before there would be yet another senior leadership crisis. What OpenAI really needed was stronger governance and accountability mechanisms. During Murati’s time getting iced out, Altman had seemed most worried and threatened by the possibility that she had shared her detailed feedback with the board. In fact, she hadn’t spoken with them. She had wanted to resolve things with him directly and hadn’t been so sure that involving the board would bring real accountability. Over the years, she had been skeptical of its various configurations. Having three cofounders on what was meant to be an independent nonprofit board was far too many. And many of the independent directors had not had true independence from Altman; in one way or another, they’d had financial ties with him or had benefited from his networks. She had observed his investing in startups and donating to politicians to establish and entrench important relationships. He had at various points asked her what he could do for her, and she had always demurred. She didn’t want to entangle herself in his web and owe him later. But if there was a time to reach out to the board, perhaps now was a good moment to at least open up a more regular channel of communication. She would be in Washington, DC, where Toner lived, to speak at The Atlantic’s annual ideas festival at the end of September. The board was in the middle of its search for new members, and whoever joined next could help either strengthen or weaken its independence. Altman needed real oversight. Murati reached out to Toner for coffee. --- For Toner, Murati’s reach out was unusual but not totally unexpected. Toner was a board member; Murati, an executive. It seemed reasonable for Murati to want to talk. The coffee, on September 29, 2023, seemed relatively standard. Murati had given various updates about the company: OpenAI would make tons of money, no problem; the most important things in motion were its Gobi model and the most recent deal it was negotiating with Microsoft; and she was dealing with some personnel issues related to Altman and Brockman’s dynamic. This time it had something to do with Sutskever and Pachocki. Only one thing had somewhat surprised Toner: Altman was pressuring the company to ship so fast, Murati had said, that she worried it could lead to bad things happening. Much more unusual to Toner was the email she received from Sutskever days later. In the two years they had been together on the board, Sutskever had never once contacted her individually. His email had asked her if she had time to meet the next day. Now, on October 4, he was so nervous he was having trouble talking. Toner took the lead. “I just really care about moving toward having a good strong board that can oversee the company,” she told him. “That’s what we all want.” Sutskever suddenly laughed and scoffed at the same time in a highly uncharacteristic way. “I totally agree,” he said, in a way that suggested others did not. “Everyone agrees,” Toner said benignly. Sutskever rolled his eyes. The independent board members, Toner explained, were looking to strengthen the board by adding a new director with a strong AI safety background. If he had different ideas, she would be glad to hear them. At this, Sutskever latched on. The board needed to be better informed, he said. They needed to pay attention to what was happening. Toner tried to probe further. And what did he think was the most important information that the board needed to know? Sutskever paused, choosing his words carefully. “I hesitate to answer your question directly,” he said. “If I answered it, you would understand why. “At the highest levels, OpenAI is a tricky environment,” he continued. “OpenAI is trickier than it seems from the outside. “Maybe I’ll sleep on it and I’ll realize there are some specifics I can share,” he added cautiously. In the meantime, he recommended, it might not be the worst thing for Toner to chat with Murati. Murati would have more context about what was up. --- Murati heard back from Toner over a week later. After another board meeting, Toner had simply told Murati that something funny seemed to be happening. Murati responded that Toner was very perceptive. They agreed to have another talk. It was now October 15, 2023, and Toner had begun the call with a generic opener. “How are things going?” Toner had asked. “Is there anything that the board should know about?” Murati weighed her words. She needed to proceed cautiously. “There’s a lot of tricky stuff going on,” she tentatively offered, echoing the phrasing that Sutskever had used before her. Murati continued: She couldn’t talk about all of it, at least not the trickiest parts, because once she did, she wouldn’t be able to take it back. Especially now while Altman was feeling incredibly threatened in his position as CEO. This last bit seemed to surprise Toner. Altman, threatened? The board had certainly tried to put in place stronger governance mechanisms to check his power, but at no point had they desired to threaten his position or even remotely discussed removing him, Toner said. Murati elaborated: Altman was an incredibly anxious person. And when he felt anxious, he had dumb ideas, particularly when enabled by Brockman. Altman’s anxiety also fed into toxic behaviors that always followed the same playbook: To anyone resisting his decisions, he would say whatever he thought they wanted to hear to win their support; then, when he lost patience waiting and believed they would continue to go against him, he would undermine their credibility until they got out of the way. It was subtle but pervasive, and had most recently manifested with the issue between Sutskever and Pachocki. It had been extremely damaging and had left Sutskever very upset, she said. The board needed to focus on making sure they didn’t bring in an Altman ally as the seventh director, Murati continued. And it needed to pay attention to Microsoft’s deployments of OpenAI’s technologies and the DSB. She couldn’t say much more. Altman would be freaked out if he found out that she and Toner were talking. But the level of toxicity at the highest levels of management wasn’t sustainable, and something needed to give in the next six to nine months. Toner should talk to Ilya, Murati finished, and see what he felt comfortable sharing. --- Sutskever had had much on his mind when he’d first reached out to Toner. Over that year, as he’d watched OpenAI’s rapid rise, he had grown increasingly preoccupied by thoughts of AGI’s imminent arrival: the cataclysmic shifts it would cause, the way they would be irrevocable, the responsibility OpenAI had to ensure an end state of extraordinary abundance, not extraordinary suffering. Then he became consumed by another anxiety: the erosion of his faith that OpenAI could even reach AGI, or bear that responsibility with Altman as its leader. After ChatGPT, working at OpenAI and rising up its ranks had become the ultimate social currency in Silicon Valley. It had created a new level of internal competitiveness and office politics as different team leaders jostled for attention and priority. Altman was making it significantly worse, Sutskever observed. Instead of negotiating between egos, he was conveying to everyone exactly what they wanted to hear as he maneuvered to get exactly what he wanted. And he was telling so many little lies and some big ones in the process that it was becoming a near-daily occurrence. Brockman added to the turmoil, as Sutskever saw it. Gone were the days when the two original cofounders turned to each other as trusted confidants with their fond memories of the endless hours spent holed up together, dreaming about what OpenAI could become. To Sutskever, the result was the most toxic combination: a directionless, chaotic, and backstabbing environment where people no longer had shared information or a shared foundation of trust to agree on critical decisions about how to move forward. This infighting was undermining what Sutskever saw as the two pillars of OpenAI’s mission: It was slowing down research progress and eroding any chance at making sound AI safety decisions. And now he was also being harmed directly by Altman’s behavior. After his first call with Toner on October 4, Sutskever had slept fitfully, consumed by stress. Toner, he’d felt, had been the safest independent board director to approach. She had been vocal during board meetings about instituting strong governance and safety mechanisms, and was most apparently not in Altman’s pocket. He had been less sure about the other two independent directors, D’Angelo and McCauley. Still, he worried about how much he could fully divulge to Toner and what would happen if Altman found out. As Sutskever had wrestled with these thoughts, Annie’s recirculating allegations on Twitter had added yet another dimension to his piling list of questions about Altman’s fitness to lead the world to AGI. After the New York magazine article, which some people discussed only in hushed tones at the company, two of Annie’s old tweets in particular had newly gone viral: the one from November 2021, which accused Sam of “sexual, physical, emotional, verbal, financial, and technological abuse,” and another from March 14, 2023, which would rack up nearly 4,000 Likes and was more explicit: I’m not four years old with a 13 year old “brother” climbing into my bed non-consensually anymore. (You’re welcome for helping you figure out your sexuality.) I’ve finally accepted that you’ve always been and always will be more scared of me than I’ve been of you. Sutskever didn’t know whether her allegations were true, but he believed that whatever had happened, Annie had had a rough experience growing up with Sam. It was evidence of how long Sam’s history of problematic behaviors could have extended. Annie’s word, abuse, was also the word Sutskever felt best captured his own observations of Altman. Like Murati, he had taken a long time to understand Altman’s playbook, though there had been signs of his untrustworthiness from even the beginning: His insistence on being OpenAI’s CEO without clear or consistent reasoning; little lies he had sometimes told through the early years that seemed so inconsequential as to have no point; the warnings that Amodei had conveyed at the end of 2020 when he was leaving. Sutskever hadn’t fully grasped then Amodei’s phrase “psychological abuse.” Now, with Altman’s behaviors worsening and their impacts rapidly escalating, Sutskever had a new and deeper understanding of its meaning. On October 12 and 13, he’d gone on a retreat with his Superalignment team, where he’d burned another effigy as a team-bonding experience and continued to wrestle with how to move forward. Now, on October 16, he was on a second call with Toner and ready to share a bit more. He recounted the situation with Pachocki and the ways it illustrated Altman’s behaviors. Altman could have simply told him directly and honestly that he wanted Pachocki to play a bigger role. Instead, he had pitted Sutskever and Pachocki against each other, in no small part assisted by Brockman. They seemed eager even to do so, leaving the two scientists to fight each other without full visibility into why they couldn’t seem to reach an agreement. “The beatings will continue until morale improves,” Sutskever said. The problem was that everything Altman did was always so subtle. Each act viewed in isolation didn’t seem like that big a deal. It was only when viewing it all at once that patterns snapped into focus. The takeaway was not Look at this bad incident where Ilya feels like he’s been wronged, he stressed. This was just the latest instance of Altman’s patterns of abuse. There was also his treatment of Amodei, with whom Toner should speak to, Sutskever urged. And there were the circulating allegations from his sister. “It’s a different kind of a safety issue, if you see what I mean,” he said. The bottom line was Altman, sometimes with Brockman, had treated many people similarly over the years, manipulating and lying to people so habitually that at times he said things that he didn’t even seem to believe himself. That said, Sutskever added, the two of them, Toner and himself, could talk and decide there wasn’t much to do. If that happened, they should forget they ever had this conversation. There was a specific discussion that he had wanted to have with Toner though that had made him reach out. The board had scheduled its second annual on-site for the end of November, where the intent was to finally make a decision on new directors. He wanted to talk about this with Toner. He didn’t think expanding the board was a good idea. He wasn’t sure it would successfully emerge with more independence. Even if the people who joined weren’t Altman loyalists, it would take the new directors too long to pick up on his tactics. Holding him accountable would become harder, not easier, Sutskever said. He wasn’t fully certain of this opinion, though, he hedged cautiously. He wanted to know what Toner was seeing. Toner agreed that Altman was slippery. She’d seen in her own professional life how slipperiness at the management level could cause cascading problems, she said. Now that the board had lost three members who had been most deferential to Altman, she also agreed that the goal of whatever happened next was to hold Altman accountable. Sutskever was more at ease. He understood Toner’s position better now. They had been on opposite sides of the deadlock, but after the same thing: to create real checks on Altman’s power. Where she’d seen adding board members as the way to do so, he’d seen the opposite. He pressed forward. He realized that his concerns could seem to Toner very intense and sudden, but it was because there was a narrowing window in which to remedy the issue. “The board is like outer alignment, the management is like inner alignment,” he said. Where Toner was saying OpenAI needed to fix the outer alignment, he believed it needed to fix the inner. Toner seemed to digest the information. “It sounds like you think some quite major changes should be on the table,” she said. “Yes,” Sutskever answered. But the biggest challenge would be that there was no clear-cut evidence to point to of anything obviously egregious with Altman’s leadership. Most likely, the board would decide there was nothing they could do, they would finally agree on new directors, and the moment would pass. At this, Toner raised a few alternatives. Perhaps the board could set up different targets for OpenAI to hit to more concretely measure Altman’s performance and revisit the issue in twelve months. Sutskever brushed this aside. Altman would pass whatever targets the board implemented, but it wouldn’t result in any structural changes to address his behavioral issues. Perhaps Brockman could step off the board, Toner suggested. That could certainly help, Sutskever answered, but it still wasn’t sufficient. While Brockman definitely exacerbated Altman’s dynamic, Altman was really the root of the problem. “I have been thinking in a related direction as a plan B,” Sutskever said to Toner’s proposal on Brockman. He stopped short of saying aloud what he saw as plan A. But as they wrapped up the call and agreed to speak again the following week, Sutskever was certain that Toner was beginning to get it. --- Murati was dealing with yet another crisis. Shortly before she spoke with Toner the second time, Altman had started panicking, for seemingly no apparent reason, about Microsoft being unhappy with OpenAI. In an attempt to get to the bottom of why, Murati had set up a meeting with Microsoft executives, including Mikhail Parakhin, the head of Bing. The meeting had gone better than expected, but in the process she had discovered that Altman had yet again said yes to one of Microsoft’s demands without grounding in what OpenAI was actually doing, creating false expectations with Microsoft about what OpenAI would deliver. She had yet again been stuck with cleaning up the situation. The Microsoft meeting was on October 19. Now, on October 20, she was telling Toner over a call in a third meeting that she planned to give Altman lots of feedback. There were so many issues caused when Altman said yes and she said no. It had created a lot of fragmentation in the Microsoft relationship. It was the same exact situation with Sutskever and Pachocki, Murati continued. Just today they had finally landed on a configuration for how Sutskever and Pachocki would continue to coexist in their roles. She and Sutskever had then implored Altman, shortly after reaching the agreement, not to deviate from their decision when speaking with Pachocki by simply saying what Pachocki wanted to hear. “When I talk to Jakub, he hears me, then he goes to Sam and he hears something different,” she told Toner. Altman had given them his word in the meeting. But everything remained precarious. Brockman could speak to Altman, on behalf of Pachocki, and seek to influence him, threatening once again to unbalance the situation. Brockman was a whole other story, Murati said. He had recently admitted to her that he had tried to fire her during the development of GPT-4. She was technically his manager and used to write Brockman’s performance reviews, but it had always created so much drama that she stopped. Murati would later confide to someone else that she had wished she could fire him, but she couldn’t because he was on the board. She had at one point thought about asking Brockman to step off the board, she told Toner. In the end, she hadn’t. Now the chaos was rampant. And did Murati know anything about the situation with Annie? Toner asked. Murati didn’t. She hadn’t asked Sam about it and had no real context for what had happened within his family. If even 10 percent of it were true, though, it was really bad, she said. “I’m shocked that I can do my job as well as I can with everything that’s going on,” Murati continued. She was writing down notes as things happened. If Toner needed it, she could send more information. Toner responded: The board would focus on the things that it could actually change—not Sam’s personality or behavior but instituting better governance processes and structures to keep him in check. This aligned with what Murati had been trying to do within the company. She had just one more word of caution for Toner. “Make sure your information isn’t just coming from Sam,” she said. --- Toner wasn’t sure what Altman wanted to talk about. He had texted her earlier that day, October 25, asking if she had time to chat today or tomorrow. Two days before, on October 23, she had spoken again to Sutskever, who seemed far more open after hearing from Murati that she and Toner had also spoken. He made his concerns more explicit than ever before. “I don’t think Sam is the guy who should have the finger on the button for AGI,” he’d said, and noted the “tremendous opportunity” that had befallen the board to do something about it. He’d then suggested a path forward: replace Altman with Murati as an interim CEO. Later, as Toner, McCauley, and D’Angelo all conferred with one another, they realized that Murati had also said, “I don’t feel comfortable about Sam leading us to AGI.” The revelation would have a huge influence on their thinking. If two of Altman’s most senior deputies—one from Applied and one from Safety—both felt this way, the board had a serious problem. Then, on October 24, Toner had had a meeting with D’Angelo and McCauley to discuss steps they could continue to take to shore up the board’s oversight mechanisms. One glaring issue: OpenAI’s nonprofit didn’t have sufficient independent legal support, and everything was being routed through the for-profit lawyers. The three agreed that it was time to find new nonprofit lawyers who could be present at every board meeting and help review all of the deals and other legal arrangements that Altman was striking. Toner wondered whether Altman had somehow caught wind of these meetings and wanted to put an end to the discussions. But now over a call, he was talking about something completely different—concerns he had over a research paper she had published in her day job at CSET. Toner had published three papers that week. He singled out the one that had been the most dense and academic. It was about a political science idea called “costly signals,” referring to the challenges that state and private actors face when signaling to the public about their intentions with AI regulation and development. She was the third coauthor, and references to OpenAI were buried on pages 28–30 of a sixty-five-page document. The company hadn’t been mentioned anywhere else—in the executive summary, on the web page, or in any of the launch materials. Based on its traffic, few people had even read it. She was confused as to how it had even come to Altman’s attention. Unbeknownst to Toner, the paper had surfaced the day before on OpenAI’s Slack in the #policy-research-chatter channel. The pages that mentioned the company had identified in turn the strengths and weaknesses of OpenAI’s and Anthropic’s model release strategies, and included praise of OpenAI for publishing a candid safety assessment of GPT-4 and stating that it had delayed the model’s release six months to do so. David Robinson, OpenAI’s head of policy planning, had pasted into Slack only a selection of three paragraphs—the ones that critiqued OpenAI and commended Anthropic. He bolded several lines for emphasis, which contrasted the “race to-the-bottom dynamics” that ChatGPT spurred and the restraint that Anthropic had shown releasing Claude after ChatGPT. “Speaking of CSET reports, just seeing this new one,” Robinson had written. “Helen Toner is a coauthor and the comparisons between OpenAI and Anthropic are quite spicy.” It had sparked a short discussion in the channel: “Yeah that is surprisingly partisan (not so much the criticisms of us, which IMO is harsh but fair, but rather the uncritical treatment of Anthropic),” one person wrote. “Yeah agreed it feels quite partisan & I’d say also quite flimsy?” another added. “Regardless,” he continued a little farther down, “very much appreciate the share—just surprised at this level of analysis from the report, based on reading these snippets.” Altman told Toner the paper had been flagged by someone external in an email to OpenAI only a few hours after its release. He was worried that it could look bad for a board member to criticize OpenAI while it was under regulatory scrutiny, including from a July 2023 FTC probe over the company’s data, training, and security practices as well as its models’ hallucinations that may have reputationally harmed consumers. Toner had drafted the paper in May or June of that year, before all of the intensified regulatory scrutiny. She admitted that she hadn’t taken a look with fresh eyes in the context of the new political environment. The call lasted fifteen minutes. Altman’s voice had been mild-mannered throughout. They both agreed at its conclusion that she would email the rest of the board to flag the paper and explain what had happened. In the email, she struck a conciliatory tone. She apologized for two mistakes: believing that the paper wouldn’t draw anyone’s attention and not reviewing it more closely because of it. “Sam and I both agree it’s important for board members to be able to criticize the company if we want to, but that this would not be the way to do it,” she wrote. None of the other board members responded to her. With that, it seemed the matter was over. --- In Sutskever’s third call with Toner on October 23, she had suggested he reach out to McCauley and D’Angelo. Sutskever was still not quite sure about them and whether he could give them his trust. After he met with D’Angelo in person, D’Angelo hadn’t seemed as aware as Toner of Sam’s problematic behaviors. Now on the phone with McCauley on October 26, Sutskever was skittish about not revealing too much. But there was something he wanted to know from McCauley. Shortly after Altman’s call with Toner, Altman had sent out an email to some people within the company saying that he had spoken with her about her paper and strongly disagreed with her about its consequences. “I did not feel we’re on the same page on the damage of all this,” he wrote in the email. “Any amount of criticism from a board member carries a lot of weight.” To Sutskever, Altman had said more directly that Toner needed to go as a board member and that McCauley had agreed with him. This felt off to Sutskever. “Sam said that when he spoke to you about Helen’s paper, you said, ‘Helen’s obviously got to go,’ ” he ventured to McCauley. “And Sam said he updated positively on you as a result. Is that true?” On the other end of the line, McCauley seemed dumbfounded. She had definitely not said that, she responded. Altman had indeed called her late on October 24 to talk about the paper. He’d mentioned that it had included a section he thought was critical of OpenAI and that D’Angelo hadn’t felt it was a fireable offense but also that Toner shouldn’t have written it. McCauley had then told Altman that she hadn’t seen the paper and suggested he have a conversation directly with Toner. There was no way she could have said anything to suggest pushing Toner off the board, McCauley said. It would seem like a coincidence: In the middle of talking about Altman’s “specific untruths,” here was yet another example playing out in real time of exactly that. Altman would have even gotten away with it had Sutskever not already had reason to reach out to McCauley. Altman knew that the two typically never talked. But to Sutskever, the frequency with which Altman was lying and maneuvering had made it only a matter of time before something like this happened. After hanging up with McCauley, Sutskever called Toner back. It was time, they agreed, for the three independent board members to talk.