The Blip

“Mr. Altman’s departure follows a deliberative review process by the board, which concluded that he was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities.” — OpenAI, “OpenAI announces leadership transition,” November 17, 2023

At 12:19 in the afternoon on Friday, November 17, 2023, Sam Altman opened a Google Meet on his laptop and found that the rest of the company he ran was waiting for him. Ilya Sutskever, OpenAI’s chief scientist and the man whose name was on the papers that made deep learning possible, had texted the night before to schedule the call. Altman had assumed it was routine. It was not. On the video grid sat the four board members who, over the previous days, had decided to remove him: Sutskever, the Quora chief executive Adam D’Angelo, the Georgetown researcher Helen Toner, and the entrepreneur Tasha McCauley. Sutskever did the talking. Altman was out, effective immediately. The board no longer had confidence in his ability to lead.

The whole thing took a few minutes. There was no severance discussion, no transition plan, no list of grievances laid on the table for him to answer. Altman, who had spent a year as the most recognizable executive in technology, the public face of the most-discussed product since the iPhone, was being fired by a six-person nonprofit board over a video call, and even as it happened the rest of the world did not yet know. He would later describe the experience as surreal and disorienting, the strangest stretch of his professional life: a coup mounted from inside his own company by people he had helped bring in.

Four minutes after Altman learned he was gone, at 12:23, Sutskever delivered the second blow on a separate call. Greg Brockman, OpenAI’s president, co-founder, and chairman of the board, was being removed from the board. He could keep his job at the company. Brockman, who had helped start OpenAI in 2015 and had famously worked through holidays and weekends with the obsessiveness of a man who believed he was building something that mattered, was told that the chairman of the board had not been informed of the decision to fire the chief executive until the moment the decision was already made. He would not stay long under those terms.

Roughly one minute before OpenAI told the public, it told Microsoft. Satya Nadella, the chief executive of the company that had committed something like thirteen billion dollars in total to OpenAI across several rounds and had wired the startup’s entire trajectory into Azure, learned that his most important partner had just decapitated itself by reading the news essentially as everyone else did. Microsoft owned a large economic stake in OpenAI’s for-profit arm and had bet the future of its consumer and enterprise software on the partnership. It had been given the courtesy of about sixty seconds.

At a few minutes past noon Pacific time, OpenAI published a blog post under the bloodless title “OpenAI announces leadership transition.” The sentence that would be read, parsed, and re-read for months sat in the third paragraph. Altman’s departure, the board wrote, followed “a deliberative review process,” which had concluded that he “was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities.” Mira Murati, the chief technology officer, would be interim chief executive. Brockman would step down as chairman but stay on as president.

“Not consistently candid.” It was a phrase a lawyer might use to avoid a phrase a lawyer would never use. In the days that followed, every reporter, investor, and employee who read it asked the same question, and the board, remarkably, refused to answer it: candid about what? There was no allegation of fraud, no missing money, no safety incident, no specific event the board would name. A company that had raised billions of dollars and employed roughly seven hundred and seventy people had just removed its chief executive in language vague enough to mean everything or nothing, and then declined to elaborate. That vacuum, more than the firing itself, would determine what happened next.

To understand why four people thought they could do this, you have to understand the strange machine they sat atop. OpenAI was not a normal company. It had been founded as a nonprofit in 2015 with a mission written in almost theological terms: to ensure that artificial general intelligence benefited all of humanity. When the cost of training frontier models made pure philanthropy impossible, Altman and his colleagues had bolted a capped-profit subsidiary onto the nonprofit in 2019, a hybrid in which investors could make money up to a ceiling and the nonprofit board retained ultimate control. The board’s job, written into the structure, was to decide whether the company was being faithful to the mission, not to maximize shareholder value. The board had been given, on paper, the power to fire the chief executive if it concluded he was steering the company somewhere dangerous. On November 17, for the first time, it used that power. The question the next five days would answer was whether the power was real.

The people who held it were a thin majority. By late 2023 the board had shrunk through departures to six members. Two were insiders, Altman and Brockman. The other four were Sutskever and three independents: D’Angelo, the Quora chief executive whose company ran an AI chatbot product that competed, awkwardly, with OpenAI’s; Toner, a director at Georgetown’s Center for Security and Emerging Technology who studied AI policy; and McCauley, an entrepreneur with ties to the effective-altruism movement, the community of thinkers preoccupied with whether advanced AI might one day pose an existential risk. Four against two. The independents and Sutskever could outvote the founders, and on November 17 they did.

What had pushed them there is the part of the story that has never fully resolved, and the honest version admits as much. Several frictions had been building. In October, Toner had co-authored an academic paper that, in a passing comparison, praised a rival lab’s safety practices and characterized the competitive scramble that followed ChatGPT’s release as exactly the kind of dynamic responsible labs should avoid. Altman read it as a board member publicly criticizing her own company, and colleagues later said he had moved to push her off the board. Toner argued she had made an academic point, not launched an attack. Separately, senior executives had brought Sutskever their concerns about how Altman managed people, describing a leader who told different people different things and set them against one another. And there was the slow accumulation the board would eventually point to: a chief executive who, in their telling, had withheld information and misrepresented what was happening inside the company often enough that they no longer trusted what he told them.

There was also a rumor, and it deserves to be named precisely because it was so widely repeated and so thinly supported. In the days after the firing, a Reuters report described a letter from staff researchers to the board warning of an internal breakthrough, code-named Q-star, that the researchers said could threaten humanity by improving the system’s ability to reason through math. The detail was irresistible: a secret AI so capable that its discovery had triggered a coup. It was almost certainly wrong as an explanation. OpenAI never confirmed that Q-star had anything to do with the firing, the board members themselves said it did not, and the later independent review found no safety breakthrough behind the decision. The reasoning research that the codename pointed toward would surface, much later and far less dramatically, as a real and useful line of work. But on the weekend it mattered, Q-star was a ghost story that filled the silence the board had left, and the silence was the actual problem.

By Friday evening, the silence had begun to cost the board. Brockman, who had been told he could keep his job, did not want it. He posted a detailed timeline of the firing on the platform formerly called Twitter, laying out the minute-by-minute sequence, and announced that he was quitting. He and Altman, he wrote, were shocked and saddened. Several of OpenAI’s most senior researchers resigned in solidarity within hours. The board had removed two people. By the end of the first night it had lost more, and the people it was losing were precisely the ones who could not be replaced by a recruiter.

Saturday belonged to the investors. As the scope of what had happened became clear over the weekend, Microsoft and the other backers who had put money into OpenAI’s for-profit arm began pressing the board for answers and for a reversal. Nadella and his chief technology officer, Kevin Scott, wanted Altman back. Thrive Capital and others wanted Altman back. They wanted, at a minimum, to be told what he had done. And here the board’s refusal to specify became something close to fatal, because investors who are given no concrete reason for an action that threatens billions of dollars do not conclude that the action was wise and discreet. They conclude that there was no good reason, and they move to undo it. Even inside the company, the people now charged with running it were working to bring back the man who had just been fired. Murati, the interim chief executive, was among those trying to reinstate Altman.

Sunday was the day the deal that might have ended it quietly fell apart. Altman came back to OpenAI’s offices for negotiations about returning. The terms were the sticking point. Altman and his allies wanted the board that had fired him to resign; the board wanted assurances and a governance arrangement it could live with; the two sides could not agree on who would hold power afterward. As the hours passed and Sunday night arrived, the board did the opposite of reconciling. It announced that Murati was being replaced as interim chief executive by Emmett Shear, a co-founder and former chief executive of the streaming service Twitch. Shear had not built large language models. He had, however, written publicly about wanting to slow down the pace of AI development, which told employees a great deal about what the board thought it was doing. It was now installing its second interim leader in three days, over the objection of nearly everyone who actually worked there.

In the small hours of Monday morning, the conflict escalated past the point of repair. Nadella posted that Altman and Brockman, “together with colleagues,” would be joining Microsoft to lead a new advanced AI research team, and that Microsoft remained committed to OpenAI regardless. It was a stunning move and a deliberate one. Microsoft was offering a landing pad large enough to absorb two founders and the entire company behind them, and it was signaling to every OpenAI employee that there was a place for them, with funding, if they wanted to follow. The board had tried to remove one man. It now faced the possibility that the man would simply reconstitute OpenAI inside its largest investor, taking the talent, the momentum, and the mission with him, and leaving the nonprofit in control of an empty building and a logo.

Shear, for his part, did not behave like a man certain of the board’s case. He said publicly that he had taken the job, that he intended to hire an independent investigator to establish what had actually happened, and that he would resign if the board could not produce written evidence of wrongdoing. By his own description he doubted the board had handled the firing correctly. The interim chief executive the board had installed was, within hours of taking the role, publicly questioning the decision that had created the role.

Then the employees moved, and that is what decided it. On Monday a letter began circulating among the staff, addressed to the board. It demanded that the board resign and reinstate Altman and Brockman, and it threatened the only thing employees of a startup can credibly threaten: a mass walkout, in this case to Microsoft’s new subsidiary, which had just promised to take them. The signature count climbed through the day in a way that became its own form of theater, the numbers passing from a few hundred to well over six hundred to, by most accounts, more than seven hundred of the company’s roughly seven hundred and seventy people. The exact figure varied by the hour and by the source, which is why the safest way to describe it is the way the company itself effectively did: nearly everyone. The letter’s most quoted line went straight at the board’s competence. The signatories wrote that the board’s conduct had made clear it was not capable of overseeing OpenAI.

Two signatures on that letter mattered more than the rest. One was Murati’s. The interim chief executive the board had named on Friday and replaced on Sunday signed the letter demanding the board’s resignation. The other was Sutskever’s. The man who had delivered the firing to Altman over a video call on Friday signed, on Monday, the letter repudiating the decision he had cast a vote for. And then he went further. On the same platform where Brockman had posted the timeline, Sutskever wrote that he deeply regretted his participation in the board’s actions, that he had never intended to harm OpenAI, that he loved everything the company had built together, and that he would do everything he could to reunite it. The decisive insider vote had publicly recanted within seventy-two hours.

Murati gave the revolt its slogan. “OpenAI is nothing without its people,” she wrote, and the line propagated across the company’s accounts as employee after employee posted the same five words, a coordinated message aimed at a board that had spent the weekend learning, in real time, exactly what the sentence meant. The company’s value did not sit in its servers or its model weights or its brand. It sat in a few hundred researchers and engineers who could walk to Microsoft on Monday afternoon, and who were now telling the board, in unison, that they would.

By Tuesday the board’s position had collapsed. It had lost the founders, the executives, the chief scientist who had been its own decisive vote, the largest investor, and the workforce. There was no version of events in which it kept control of a functioning company. The negotiations restarted, but this time the subject was not whether Altman would come back. It was the terms under which he would.

Late on Tuesday night, in the early hours of Wednesday, November 22, OpenAI posted the resolution. It had reached an agreement in principle for Sam Altman to return as chief executive, under a new initial board: Bret Taylor, the former co-chief executive of Salesforce, as chair; Larry Summers, the former Treasury secretary; and Adam D’Angelo, the one member of the firing faction who would remain. D’Angelo stayed, by most accounts, to provide continuity and to represent the old board in vetting the new one. Toner, McCauley, and Sutskever left the board. Sutskever stayed on as an employee, at least for the time being. The crisis that had started at 12:19 on Friday afternoon ended sometime after midnight on Wednesday, roughly five days, start to finish. Emmett Shear had been interim chief executive for something under three days, one of the shortest tenures in the history of the role.

In the narrowest sense, the board had accomplished nothing it set out to do. The chief executive it removed was back. The independent members who had voted him out were gone. The nonprofit’s much-discussed power to fire the leader in service of humanity had been exercised exactly once, and the exercise had demonstrated that the power could not survive contact with the people, the investors, and the platform partner whose interests it had crossed. Microsoft, which had been told one minute in advance, came out of the weekend with its partner restored and, before long, a non-voting observer seat on the new board, a small structural acknowledgment that being blindsided by a thirteen-billion-dollar investment is not something a company allows to happen twice.

The board’s defenders argued, then and afterward, that the substance had been right even if the execution had been a catastrophe. Months later an independent review commissioned by the company and conducted by the law firm WilmerHale traced the firing to a breakdown in trust between the board and Altman, ruling out concerns about product safety, security, finances, or the pace of development, and concluded that his conduct had not required his removal. The phrasing cut both ways. It cleared Altman of the dramatic charges the rumor mill had supplied. It also confirmed that the board’s central complaint, that it could not trust what the chief executive told it, was a real and documented thing, not an invention. The following spring, Toner gave her account in public for the first time. She said the board had learned that ChatGPT was launching not from Altman but from social media; that he had not disclosed his ownership of an investment fund that bore OpenAI’s name; that he had given the board inaccurate information about the company’s safety processes more than once. For years, she said, Altman had made it difficult for the board to do its job by withholding information and, in some cases, lying to it. OpenAI’s new chair pushed back, pointing to the review that had cleared him.

Both things could be true. The board could have had genuine cause for alarm and still have detonated it so incompetently that the explosion served no one. By refusing to name a single concrete reason, the four had asked roughly seven hundred and seventy people, plus Microsoft, plus the investors, to take on faith that a decision with no stated basis was nevertheless sound. In a startup, faith of that kind goes to the person who hands out the equity and ships the product, never to a board. That person was Altman. The structure had been built on the premise that the nonprofit board was the adult in the room, the check that capital and ambition could not override. The weekend proved that the check existed only as long as no one with real leverage objected to it, which is another way of saying it did not exist at all.

Insiders took to calling the episode “The Blip,” a name with a deliberate shrug in it, as if five days that nearly dismembered the world’s most valuable AI startup had been a momentary glitch in an otherwise smooth ascent. The name was wishful. The questions the weekend raised did not blip away. The legal power to control a company that says it is building artificial general intelligence had sat with the board. The real power, the weekend had shown, sat with the people who could credibly threaten to leave en masse and the investor who could credibly offer to take them. A governance structure designed expressly to let conscience override commerce had failed the first time conscience tried to use it. The crisis settled the immediate matter and left the larger one open, and the larger one would be fought over the next several years rather than a single weekend, in the slow grinding form of lawyers and attorneys general and the question of whether the nonprofit at the center of OpenAI would keep any real control at all. Sutskever would leave for good the following spring, taking his worries about safety with him. The board that fired Sam Altman had wanted to prove that the mission came before the man. It had proved the opposite, and everyone who built AI for a living was watching when it did. If a company’s own conscience could not restrain it, the job of restraint would have to fall to someone outside the building, and the only institutions large enough to try were governments that had spent the same weekend learning, like everyone else, how little they understood the thing they were now expected to regulate.