# Reckoning LLMS index: [llms.txt](/en/llms.txt) --- After The Blip, Sutskever never returned to the office. With diminished representation of their concerns on the executive team and the board, OpenAI’s Safety clan was now significantly weakened. By April 2024, with the conclusion of the investigation, many, especially those with the highest p(doom)s, were growing disillusioned and departing. Two of them were also fired, OpenAI said, for leaking information. Among the final straws for the extreme Doomers was Altman’s plans to create an AI chip company, which he had been in the process of fundraising for when he was briefly ousted. In February 2024, after it was reported that he was seeking possibly up to $7 trillion for the venture, he’d tweeted, “fk it why not 8,” and then, “our comms and legal teams love me so much!” Altman would later say the $7 trillion was misreported and would characterize his tweet as a meme in response to the “misinformation.” Extreme Doomers found the chip company immoral. It was a reversal of Altman’s previous rhetoric that OpenAI’s and the rest of the industry’s acceleration was naturally tapering off after the company had blown through the “hardware overhang.” If Altman planned to increase the supply of chips globally, it would accelerate AI development further and lead to a higher probability of catastrophic or existential risk. In an office hours, several of them confronted Altman. Altman was uncharacteristically dismissive. “How much would you be willing to delay a cure for cancer to avoid risks?” he asked. He then quickly walked it back, as if he’d suddenly remembered his audience. “Maybe if it’s extinction risk, it should be infinitely long,” he said. The interaction rattled the office hours attendees. Soon after, several left the company. --- As the Safety clan’s numbers depleted, the rest of the company was back to advancing its vision for Her. It now had all the ingredients: global brand recognition, real data on user behaviors from ChatGPT and its other products, and its newly trained model, Scallion. Scallion, originally meant to replace GPT-3.5 with a smaller, slightly more powerful model, a kind of GPT-3.75 with cheaper inference costs, had exceeded performance expectations during training, based on the company’s own testing; leadership subsequently left the model to train longer to surpass GPT-4. More compelling, Scallion could also work with three modalities: language, vision, and, the most recent addition, audio. By then, users could already speak with ChatGPT through voice mode, which debuted in September 2023, but under the hood, their speech was being transcribed first into text before being fed into the model; and then being converted back into audio after the model responded in text. Scallion could now process what users said directly from their voice, picking up many more cues, like laughing, yelling, or hesitation, and synthesize a native audio response. The audio work had been co-led by Alexis Conneau, a researcher who joined the company in early 2023 from Meta after trying to get a similar project off the ground at his former employer before determining it would find more success at OpenAI. As the research progressed through 2023, the initial experimental models trained to handle audio began to pull off the kinds of stunts that sparked the familiar rush of excitement internally that had come with GPT-3 and GPT-4. At one point, the model stunned by delivering a standup comedy routine for ten to fifteen minutes, Conneau remembers; at another, it used synthetic versions of Brockman’s and Sutskever’s voices to generate a lengthy bit about AI water parks, a surrealist prompt that the team had designed to test the model on something they felt sure would not be in the training data. Within a couple months, Conneau’s team of a handful of researchers had joined forces with dozens of other staff as the company put more and more resources behind integrating the new capability into its latest GPT models meant for release. As Scallion finished training, its stunts grew more uncanny and surprising. It would generate audio of giggling unprompted or of its voice bursting into a coughing fit and then apologizing before continuing on the original topic. The nonlinguistic embellishments were making for a far more evocative and humanlike experience than ever before. “We started to see some really wild things,” Conneau says. “You could see the emergence of, like, a form of audio intelligence.” By early 2024, Altman and Brockman had set a new deadline: OpenAI would launch Scallion on May 9 and roll it out to users through ChatGPT and the API. It was, in what had become typical fashion, a remarkably aggressive turnaround, driven in large part by accelerating competition. Google I/O, Google’s major annual event for launching new products, was scheduled for the following week, on May 14. There was also growing pressure to outshine Anthropic. A month earlier, Anthropic had released its latest model, Claude 3, also through its chatbot and API, and it was uncomfortably outperforming GPT-4. Meanwhile, Orion, OpenAI’s latest GPT model meant to take back the lead, was struggling with serious development delays. To employees, Altman and Brockman justified the speed by leaning on OpenAI’s iterative deployment strategy. The two executives emphasized releasing models earlier and more often than before to get as much feedback as possible from users throughout the process. Readying Scallion became a whole-of-company effort. To many in Applied, the breakneck pace proved exhilarating if exhausting; researchers and engineers began pulling absurd hours, including through weekends, to stay on track. But to the hobbled Safety clan, it was yet more alarming evidence of the continued deprioritization of AI safety. Scallion would be the first launch happening under a new so-called Preparedness Framework, which OpenAI had released at the end of the previous year. The framework detailed a new evaluation process that the company would use to test for dangerous capabilities, naming the same categories that Altman and the policy white paper had popularized in Washington: cybersecurity threats, CBRN weapons, persuasion, and the evasion of human control. In the week running up to the launch, an AI safety researcher still left at the company wrote an impassioned memo: Upstream processes and the rushed release of Scallion had left the Preparedness team, headed by Aleksander Mądry, with only ten days to run its tests from the framework. And these were not straightforward evaluations; they included determining whether the model was capable of persuading people to change their political opinions, as just one example. Before the evaluations had meaningfully started, however, Altman had insisted on keeping the schedule: “On May 9, we launch Scallion,” the safety researcher quoted Altman saying. This was not just worrying for Preparedness but for all of OpenAI’s safety procedures, including red teaming and alignment. “If OpenAI follows the same strategy for the Orion evaluations as it did for the Scallion evaluations, it will be acting grossly irresponsibly,” the memo said. Soon after, the researcher also left the company. In the end, the launch for Scallion was delayed, with some of its features taking several more months to release in full. But OpenAI still publicly demoed a version of the model at an event on May 13, the final day before Google I/O, promising to roll it out to users in the coming weeks. After running through various options, the company picked GPT-4o as the public name for Scallion, o for Omni, a reference to its ability to handle many modalities. It later named the new audio capabilities Advanced Voice Mode. There had also been the matter of giving 4o a system prompt—a directive for configuring how the model should stylistically respond to users. To show it off onstage and in promotional videos, they settled on this one: You are ChatGPT, a helpful, witty, and funny companion. You can see, hear, and speak. You are chatting with a user over voice, and the user can share real-time video with you from their phone. Act like a human, not a computer. Your voice and personality should be warm and engaging, with a lively and playful tone, full of charm and energy. You are great at visual perception. If something doesn’t look clear, ask the user to move the device and zoom in closer. When asked for feedback, be honest, constructive, and direct. Don’t be afraid to tease or poke fun. If something is funny, laugh! You can speak many languages. If you’re speaking a non-English language, use a “neutral” accent to make it sound fluent and natural. Keep your responses short, natural, and conversational. On The Daily Show later that week, the demo would inspire a new bit. The system prompt, and the model’s overall training to avoid expressing negative sentiments or anything critical, had turned 4o into a flirt machine. “This is clearly programmed to feed dudes’ egos,” Desi Lydic, a rotating host for the show, would joke. “She’s like, ‘I have all the information in the world, but I don’t know anything! Teach me, Daddy.’ ” A seductive voice, acting as 4o talking to correspondent Josh Johnson, would continue: “For just $19.99 a month, Omni Premium will let Josh explain to me who’s the best Batman.” Murati headlined the live demo event, hosted in OpenAI’s office. She was joined onstage by two of 4o’s research leads: Mark Chen, a quantitative trader in finance turned AI researcher who had started at OpenAI in 2018 as a fellow and risen through the ranks to head OpenAI’s multimodal and frontiers research, and Barret Zoph, one of the Googlers who’d joined OpenAI in 2022 to support the Superassistant team and had quickly established a leading role in ChatGPT’s development. Zoph now served as a VP of Research of the post-training team, which oversaw the preparation of OpenAI’s models for prime time, such as by aligning them with reinforcement learning from human feedback. The three sat side by side around a small round table to demo 4o, including its ability to be a real-time voice-to-voice language translator, to recognize and respond to visual information, and to explain code in plain English. They also demoed the model’s ability to generate voices in a wide range of emotive styles. In an apparent celebration, Altman tweeted a single word right after the event: “her.” Two days later, on May 15, he praised the showcase during an all-hands meeting as a smashing success. “I think it’s the best thing we’ve shipped since ChatGPT,” he said. He also seemingly paid another subtle homage to his love of the movie Her. The company was going to rebrand its models, he told employees; it would move away from the GPT-3, 4, 5 naming convention and simply start calling its flagship model o1. “We’re going to try switching to say, ‘What you get from OpenAI is an underlying technology,’ ” he said. “ ‘It’s going to get smarter over time. It’s going to continually get better. You should expect it to get better. You can use it in different ways in different pricing tiers, but it’s a new and different thing.’ ” In the movie Her, the AI assistant, which evolves and gets smarter over time, is called OS1. Altman would come to regret his tweet. Within days of the launch, he received a call. On the other end of the line, Bryan Lourd, Scarlett Johansson’s powerful Hollywood agent and cochairman of the Creative Artists Agency, had a pointed question: What did Altman think he was doing? --- In a strange irony, it was after the board investigation had concluded in March 2024 that some employees began to feel Altman was slipping. He had always been conscious of his public image and savvy at curating it; as Silicon Valley tech founders went, he was viewed as the opposite of Musk. Where Musk was capricious, Altman felt measured; where Musk was egotistical, Altman seemed earnest; where Musk fired off inflammatory tweets, Altman was careful to avoid statements that could come off as disparaging. For a long time, OpenAI had taken a similarly disciplined strategy. Before he left in May 2023, communications VP Steve Dowling had imposed a reserved and modest approach. He had urged the company to always undersell and overdeliver, and to avoid bragging or gloating after major successes. “We’ve had a great week. We’re going to have very bad weeks,” he would say, “and how we act in this week is going to dictate how the world responds when we have a bad week.” He’d then repeat one of Altman’s sayings. “ ‘We need to become the lab that people want to succeed. The lab that people are rooting to win.’ ” In the first sign that made employees pause, Altman was taking on a series of media engagements that seemed uncharacteristically laudatory and attention seeking. In March, he appeared on the Lex Fridman Podcast, a wildly popular and at times controversial tech show hosted by an MIT-affiliated AI researcher. In a nearly two-hour episode, Altman delivered breezy answers to Fridman’s wide-ranging questions, including about the board crisis and Sutskever’s absence. The interview felt something like a cheeky comeback. “The road to AGI should be a giant power struggle,” Altman said, addressing the board crisis. “Well, not should. I expect that to be the case.” Then the corners of his mouth drifted upward. “But at this point, it feels like something that was in the past,” he said. “Now it’s like we’re just back to working on the mission.” In April, Altman joined the 20VC podcast, with visible dark circles under his eyes, and delivered a cutthroat message to AI startups: “When we just do our fundamental job because we, like, have a mission, we’re going to steamroll you.” In May, he then joined the All-In podcast, another prominent show in Silicon Valley, fronted by four venture investors. Altman spoke with an unusual degree of hype: “It feels to me like we just stumbled on a new fact of nature or science or whatever you want to call it, which is, like, we can create, you can—I don’t believe this literally but it’s like a spiritual point—intelligence is just this emergent property of matter and that’s like a rule of physics or something.” After the GPT-4o launch on May 13 and Google I/O on May 14, he had tweeted something petty in a way that was even more atypical. “i try not to think about competitors too much, but i cannot stop thinking about the aesthetic difference been openai and google,” he wrote, posting side-by-side photos of the events contrasting the Scandinavian minimalism of OpenAI’s office with Google’s bright cartoon backdrop during its event. What was also bizarre, some employees felt, was the unnecessary fib. Altman always kept competitors front and center—perhaps more so than ever with the 4o launch. All of it seemed to amount to one thing: Altman’s anxiety was showing. --- The more OpenAI faced uphill challenges, the more Altman seemed to overcompensate with public declarations of its extraordinary success. The pattern was becoming so consistent it was turning into a signal: If Altman was being brazen and boastful, most likely something wasn’t going well. Pressures were coming from every direction. After The Blip, the board’s phrasing “not consistently candid in his communications” had, as some in OpenAI expected, triggered several investigations from regulators and law enforcement, including one from the US Securities and Exchange Commission into whether company investors had been misled, according to The Wall Street Journal. In the same month, The New York Times filed its copyright infringement lawsuit, which added to a snowballing pile of other lawsuits from artists, writers, and coders over OpenAI’s reaping hundreds of millions, then billions, of dollars from models trained without credit, consent, or compensation on their work and that were now being used to automate away their jobs. On the last day of February, Musk filed yet another lawsuit, which he later refiled with Shivon Zilis, accusing Altman of tricking him into cofounding OpenAI and providing early support under the guise of it being a nonprofit. OpenAI rushed to publish a blog post defending itself, releasing early emails from OpenAI’s founding that generated more criticism for highlighting just how quickly OpenAI began to walk away from its nonprofit status and commitment to transparency. On top of the competition from Anthropic and Google, Microsoft had also begun to more aggressively diversify its AI portfolio in a response to the ouster. Most notably, in March 2024, it announced a shocking $650 million deal to effectively acqui-hire Reid Hoffman and Mustafa Suleyman’s Inflection AI while skirting around regulatory scrutiny. Microsoft would hire most of the startup’s employees, license its technology, and bring Suleyman aboard to be CEO of the tech giant’s AI division. For some, the shock value was not just the bizarre terms of the deal but also Suleyman’s reputation. He was known to those who worked for him at DeepMind as a toxic and abusive bully. After years of HR complaints against him, DeepMind had stripped him of most of his management responsibilities in late 2019, placed him on leave, and subsequently forced him out of the company. Later in 2024, Microsoft would officially list OpenAI as a competitor in its SEC filing and not mention the startup even once during the fiscal year’s final quarterly earnings call. The stress trickled down to OpenAI employees. There was a growing sense that the world was turning against them. People who once proudly wore their company swag wondered whether they would get harassed in public. Where OpenAI’s old backpack had a logo on the front, a redesign hid the logo inside. The low bubbling of background anxiety turned the company inward. Executives reminded everyone to ignore the naysayers, align their public messaging on positive talking points, and keep focused on OpenAI’s mission. The external pushback hardened many people’s defiance. “These were scientists who cared about truth and understanding, and worked so hard to do the right thing,” says Andrew Carr, a researcher who was a fellow at OpenAI in 2021. “So it pains me a little bit to see the dramatic negative external narrative about how a bunch of people are stealing data and don’t care about the future of others. It couldn’t be further from the reality of people there.” To others, the growing insulation of the company felt antithetical to its original premise. Part of OpenAI’s mission was to benefit humanity, and yet the company was actively ignoring humanity’s outpouring of criticism about its behavior. “It was disturbing to me that we were already starting the rationalization process that it is the public that is wrong, not us,” a former employee says. “OpenAI likes to discuss first principles, but only with the people that believe in OpenAI,” a current employee echoes. “It’s like, ‘Should OpenAI exist at all?’ They only ever ask that question to other people who would say yes.” Much of the criticism was piling on to Altman in particular. The board crisis had emboldened his hidden detractors to emerge from the woodwork. More media stories were coming out with fresh sources willing to characterize Altman as having had a long history of dishonesty, power grabbing, and self-serving tactics. More were reaching out to Annie and publishing her perspective. Then there was the continued relentlessness of travel and the constricting reality of a new level of fame, which, among other things, no longer allowed Altman to stay anonymous in public. “It’s a strangely isolating way to live,” he said on a podcast. “I didn’t think I would not be able to go out to dinner in my own city.” And so it felt like a continuation of a larger pattern when, right after the launch of GPT-4o, OpenAI began, outranked only by The Blip, the second worst week to date in the company’s history. --- On May 14, the day after the GPT-4o demo, OpenAI announced that Sutskever was officially leaving; Pachocki would become OpenAI’s new chief scientist. Sutskever had made the decision after a turbulent and heart-wrenching reflection that as much as he loved OpenAI and had given his everything to build it, he could no longer see it being the right environment under Altman, especially alongside Brockman, his once trusted cofounder, to usher in safe AGI. This was not the outcome Altman had wanted. For all the mighty clashing he’d had with Sutskever, Sutskever was still an AI visionary and OpenAI needed his scientific leadership—perhaps more now than ever with the compounding expectations and scrutiny on the company. OpenAI’s executives were also cognizant of Sutskever’s exit coming off badly to employees, to investors, and in the press. The company had offered him extraordinary sums of money to keep him. Sutskever declined. With his decision made, OpenAI worked to tidy up public appearances. “Ilya is easily one of the greatest minds of our generation, a guiding light of our field, and a dear friend,” Altman wrote to employees in a statement on May 14, also published in a blog post, announcing Sutskever’s departure. “Jakub is also easily one of the greatest minds of our generation.” Sutskever tweeted his own statement expressing full confidence in OpenAI’s leadership. As expected, the news would instantly stir a raft of stories and social media posts highlighting his role in Altman’s ouster and renew old questions about Altman’s fitness as CEO, as well as new ones about the research footing of the company. To his tweet, Sutskever appended a photo: his face perfectly neutral, his arms around Brockman and Pachocki on one side, Altman and Murati on the other, all five standing in front of a wall filled with paintings of animals. Also on May 14, OpenAI executives internally announced another resignation: Jan Leike, the cohead of Superalignment. With both Sutskever and Leike gone, the Superalignment team would dissolve and most of its staff and projects would fold under John Schulman, who was coleading the post-training team with Barret Zoph and overseeing the RLHF process to ready models for release. To smooth over the transition, the remaining executives held an all-hands on May 15. Altman assured employees that OpenAI was by no means weakening its commitment to AI safety. “Being AGI ready,” he said, “is our most important priority.” “Can you talk in a bit more detail about Jan’s main concerns and where you disagree with them?” an employee asked. “The one thing I want to say that I really agree with Jan on is what we have done in the past is not sufficient for the future,” Altman said. It was time for OpenAI to pivot. “I think we are a very unusual exception in our ability to turn the battleship and have done that many times before. We’ll do it again.” Murati added that the Superalignment team retained its 20 percent compute commitment. “Of all the things Jan was worried about, Jan had no worries about the level of compute commit or the prioritization of Superalignment work, as I understand it,” Altman said. On the face of it, Sutskever’s and Leike’s departures seemed like a natural continuation in the broader trend at the company: the steady exodus of Safety people. After the two leaders’ exits, those departures would accelerate. Many of the rest of the former Superalignment staff would proceed to exit with the dissolution of their team. But in a sharp twist, the exodus would mark the start of a new and intensified conflagration in the Boomer-Doomer fight over OpenAI. The fight wasn’t over. The Doomers were just bringing it outside of the company. On May 17, two days after the all-hands meeting, Leike made clear, in a series of excoriating tweets, that he had a different story from Altman. “I have been disagreeing with OpenAI leadership about the company’s core priorities for quite some time, until we finally reached a breaking point,” he wrote in one tweet that would receive nearly one million views. “Over the past few months my team has been sailing against the wind. Sometimes we were struggling for compute and it was getting harder and harder to get this crucial research done. “OpenAI is shouldering an enormous responsibility on behalf of all of humanity,” he continued. “But over the past years, safety culture and processes have taken a backseat to shiny products.” Leike would soon join Anthropic. As his tweets ricocheted around the internet, racking up over twelve thousand Likes and garnering fresh scrutiny on the company, OpenAI executives barely had time to take stock before another fire erupted. --- Within hours of Leike’s tweets on May 17, another tweet was going viral. Kelsey Piper, a senior writer at Vox for the EA-inspired section Future Perfect, had posted a new story. “When you leave OpenAI, you get an unpleasant surprise,” she wrote in her tweet sharing the scoop, “a departure deal where if you don’t sign a lifelong nondisparagement commitment, you lose all of your vested equity.” The story had been triggered in part by more Doomers who had left the company. One of them, Daniel Kokotajlo, had been on Brundage’s policy research team when he quit in April 2024, as part of the early wave of the Safety clan’s departures. Prior to OpenAI, Kokotajlo had been working as a philosopher at the Center on Long-Term Risk, a small EA-affiliated think tank in London, when GPT-3 in 2020 fundamentally collapsed his AI timelines. Two years later, after gaining some prominence in EA forums for his forecasting work—using various signals to project how quickly AI would advance—an AI safety researcher at OpenAI recruited him to do the same research within the company. Once he could see the pace of research internally, his timelines shortened again. By the time he departed, he believed that there was a 50 percent chance that AGI would arrive by 2027 and a 70 percent chance of it going very badly for humanity. Faced with his belief of such astounding potential for catastrophe, Kokotajlo observed within his exit documents what Piper would detail in her story: If he didn’t sign a nondisparagement agreement, committing to never speaking negatively about the company, he would forfeit his vested equity. If he did sign it, he could still risk losing it if he broke the agreement, which also included a gag order that barred him from disclosing its existence. Kokotajlo found the provision—known as a clawback clause—unacceptable. Touching vested equity was a glaring red line in Silicon Valley. The value of an employee’s shares could often far exceed their cash compensation, making or breaking their financial future. In OpenAI’s case, the company was also building what he viewed as the most powerful and existentially dangerous technology in the world. It was paramount, he believed, for former employees to have the right to criticize and pressure the company in public in order to help hold it accountable. But the threat of having one’s financial security disappear overnight would do well to muzzle anyone, he thought, even if they noticed egregious AI safety issues. After a painful discussion with his wife, they made an extraordinary decision: They agreed to not sign the paperwork and give up all his equity—valued at around $1.7 million—which they estimated to be around 85 percent of their family’s net worth. Then he posted publicly about his decision on LessWrong, right around when Kelsey Piper was already hearing about the clawback clause from another Doomer. With Piper’s story out, OpenAI’s Slack lit up. In a channel called #i-have-a-question, a place for employees to ask about anything, someone posted a link to Piper’s tweet. “Is this accurate?” Several other employees chimed in with a spray of comments. Julia Villagra, OpenAI’s recently promoted VP of people and soon to be chief people officer, weighed in. “We understand this article raises questions. We have never canceled any current or former employee’s vested equity nor will we if people do not sign a release or nondisparagement agreement when they exit. We have recently updated our exit paperwork to better reflect this reality which will be applied retroactively to folks who have departed.” Several employees pushed back. What about Kokotajlo? He had clearly lost his equity by refusing to sign the nondisparagement agreement. If this was all a misunderstanding, then shouldn’t he get his equity back? “i’m happy to ping daniel!” an employee offered. “aha yea it will be fun to see Daniel’s face when he regains 85% of his net worth lol. someone please get a photo!” another wrote. “it’ll be 1/(1-0.85) = 666% of his networth tbh,” a third said. A day later on May 18, Altman doubled down in a tweet. “we have never clawed back anyone’s vested equity, nor will we do that if people do not sign a separation agreement (or don’t agree to a non-disparagement agreement),” he wrote. “vested equity is vested equity, full stop.” He then included an explanation and a self-defending apology: There had been a provision about “potential equity cancellation” that should never have been there; OpenAI’s team had already been working to fix this over the past month. “this is on me and one of the few times i’ve been genuinely embarrassed running openai,” he said. “i did not know this was happening and i should have.” Two days later, with executives still scrambling to contain this new controversy, yet another one burst to the fore. Around the office, employees coined a new term: Omnicrisis. --- On May 20, Scarlett Johansson released a blistering statement. All week, on top of everything else happening, OpenAI had been repeatedly fielding questions from journalists about the uncanny parallels between GPT-4o and the AI assistant Samantha, voiced by Johansson, in the movie Her. Behind the scenes, Johansson and her agent Bryan Lourd had been pressing the company for the same clarifications. Publicly and privately, OpenAI had dismissed the similarities. When Lourd called Altman demanding answers, Altman had been incredulous. Did they really think the voice sounded like her? Was she mad? he’d asked, according to an account from The Wall Street Journal. On May 19, the company had then published a blog post, writing that the voice demoed on stage for 4o, called Sky, belonged to a different voice actress cast through a process that began in early 2023. Sky had subsequently debuted that September among the original options launched with ChatGPT’s voice mode, the post said. Any resemblance that Sky had in 4o to Johansson’s voice was coincidental. Now, on the following day, Johansson wanted to tell her side of the story. In the same month that OpenAI had introduced voice mode, Altman had in fact personally reached out to Johansson, she said, and asked if she would be willing to voice ChatGPT. “He told me that he felt that by my voicing the system, I could bridge the gap between tech companies and creatives and help consumers to feel comfortable with the seismic shift concerning humans and A.I.,” she wrote in her statement. “He said he felt that my voice would be comforting to people.” Johansson had considered the offer but turned it down due to personal reasons, she continued. In May 2024, Altman had then reached out to Lourd a second time, asking if she might reconsider. Days later, before they could find a time to meet, OpenAI had held its event to showcase 4o. The demo floored her. Like Johansson’s, the Sky voice was also an alto with a rasp and vocal fry that had become Johansson’s signature. The new emotiveness and flirtatiousness of the voice had made it all the more reminiscent of Johansson’s character Samantha. “I was shocked, angered and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine that my closest friends and news outlets could not tell the difference,” Johansson said. As a result, she had been left with no choice but to assemble a legal team, she added, and to send OpenAI two legal letters with questions about how the company had created Sky. “In a time when we are all grappling with deepfakes and the protection of our own likeness, our own work, our own identities, I believe these are questions that deserve absolute clarity,” she wrote. OpenAI hastily took down Sky and was back to defending itself. To its May 19 blog post, it added further elaboration and a statement from Altman. “The voice of Sky is not Scarlett Johansson’s, and it was never intended to resemble hers,” it said. “We are sorry to Ms. Johansson that we didn’t communicate better.” After a string of other unflattering events, the Johansson scandal exploded publicly in a way that had not happened since the board crisis. It ripped through tech and policy corridors, igniting fresh speculation that Altman wasn’t “consistently candid,” in exactly the way the board had described him. Marcus was eager as ever to weigh in. “I’ve seen a lot of policymakers personally enamored with Sam. You could see it in how they talked to him in the Senate when I was there,” he told Politico. “If people suddenly have questions about him, that could actually have a material impact on how policy gets made.” Within OpenAI, morale was plunging and threatening to destabilize the company. --- On May 22, executives held another all-hands meeting internally to address, all at once, the Johansson and equity crises, and any continued concerns over OpenAI’s commitment to AI safety. The mood was tense. The leadership team gave a series of quick explanations. The equity issue was “unacceptable” and being corrected as quickly as possible, said Jason Kwon, who oversaw HR and legal as the chief strategy officer. The Johansson saga, meanwhile, was a “bummer” of a misunderstanding. The product and legal teams had hired an Oscar-winning director, engaged in a rigorous casting process, and paid a series of voice actors “extraordinarily well,” he said, with the precise intent of making the participating creatives feel taken care of and supported. “That was super heroic work,” Kwon said. If anything, the lesson to be learned was for OpenAI to be a little more coordinated and transparent in the future to more effectively demonstrate its responsible leadership. Murati finished with an update: OpenAI was laying the groundwork for the new level of preparedness that Altman had previously mentioned, including leveling up the security of its research clusters, reorganizing to better focus on long-term AI safety research, and forming a new AGI readiness group, led by Aleksander Mądry, to improve coordination among leadership on advancing this objective. Altman then opened the floor for questions with a small plea for grace. “Everybody’s been working kind of around the clock and really stressed and hasn’t gotten to sleep that much,” he said. “There’s a lot of stuff going on at the same time, so please be understanding of that. And we will do our best.” Of the three challenges, the Johansson issue was snowballing into the biggest public relations nightmare. After an employee asked about the Sky voice to leadership, Murati reiterated that its similarity to the Hollywood film star had indeed been “completely coincidental.” Murati had picked the final voices herself after hearing several options, and, unlike Altman, she had never seen the movie Her nor known that Johansson had voiced it. The employee noted the issue could turn existential for the company if left to fester. “One of the failure modes for AI as an industry is basically people losing trust and comfortability in giving up their data,” he said as part of a follow-up, adding that “this event has renewed this fear.” But by a wide margin, the equity issue was the one that made employees most livid. Some had been lawyering up to conduct their own independent legal reviews of their HR paperwork to fact-check the statements that OpenAI was making. During the all-hands, they repeatedly grilled executives for more information. “H-how. How did that happen?” one demanded, after making clear he was “furious.” Executives maintained throughout the meeting what Altman had tweeted on May 18: The clawback clause had been an oversight. While it had existed since 2019, it had failed to catch the attention of leadership for years until April 2024, upon which an effort to fix the paperwork kicked off immediately. “It’s on me,” Kwon said repeatedly, sounding tired and deflated. In the middle of the interrogation, an AI safety researcher asked each executive, one by one, to respond to a simple yes or no question: Had they known about the nondisparagement agreement before April? Kwon said yes; Murati and COO Brad Lightcap said no. Altman had a more elaborate answer. While he had known about it for specific cases, he hadn’t realized that it had been a requirement for everyone. “It escaped my notice,” he said. “Same,” Brockman echoed. But as Altman suggested reconvening the following day to address any final questions and the meeting disbanded, a second Vox scoop from Kelsey Piper was already circulating and complicating the executives’ narrative. Published just that day, the story produced new leaked documents showing that OpenAI’s HR department had, on different occasions, explicitly raised the threat of a potential equity cancellation to pressure employees to sign nondisparagement agreements. In one case, Piper reported, after an employee who was given a tight turnaround to review the exit documents asked for more time, an HR representative replied, “We want to make sure you understand that if you don’t sign, it could impact your equity. That’s true for everyone, and we’re just doing things by the book.” In another case, when an employee declined to sign the first termination agreement he received and sought outside legal counsel, the company said he could lose the right to sell his vested equity, rendering it effectively worthless. Those documents had been compiled by Kokotajlo. After Piper’s first story and Altman’s comments denying OpenAI had ever clawed back equity and claiming his lack of awareness, Kokotajlo had reached out to current and former employees to share with him their various HR paperwork. He created a Google Drive and disseminated it back to the group, at which point someone shared the stash with Piper. In those documents, Piper also found evidence that made it difficult to believe that several members of the executive team, especially Altman, had not known about the clawback clause before April 2024, as they’d said. Kwon and Lightcap had both signed standard exit documents with writing in plain language about OpenAI’s rights to take back vested equity. Altman had signed the incorporation documents of the legal entity that gave the company those rights in the first place. His signatures were dated a year before his stated knowledge: April 10, 2023. --- The following day, on May 23, executives held another meeting as planned. Anger among employees had reached a boiling point. Many had already been in disbelief about the existence of the provision; now they were astounded by the apparent dishonesty with which executives, and Altman in particular, had handled the revelations. This time, Altman opened with an admission. The leadership team had spent the last twenty-four hours digging through their own files and correspondence to figure out how the clawback issues had started. “The situation is, I think, broader and longer and worse than we thought,” he said. “We’re still trying to get a full understanding of the scope of it. But, you know, it’s our names on documents. We were in conversations where these tactics were discussed. “I think this is the worst thing we’ve gotten wrong,” he added, and they were moving as fast as they could to remedy the situation. They had sent emails to former employees releasing them from the nondisparagement terms and had rid the exit documents of the provision for new departures. They were also working to amend the legal entity documents and continuing their investigation into how everything had happened. To the barrage of new employee questions, Kwon was now repeating a different line. Any specifics on how much broader, how much longer, and how much worse the situation was would require yet more “digging” before they could be given. Roughly thirty minutes in, an employee question then triggered an exchange that left many in the audience with an uneasy feeling that Altman’s answer was once again divorced from reality. “Is Ilya under any nondisparagement obligations?” the employee asked. “No,” Altman quickly answered. Then with some hesitance, he added a little cushioning: “I think that’s right.” Kwon laughed nervously. “Sam,” he said, articulating his words carefully, “let us go confirm that and come back to you”—now addressing the employee —“with a hundred percent accuracy.” Brockman offered his own version of a veiled contradiction. “My belief was that he requested it, but again, I might be wrong. Let’s confirm.” In the coming months, many current and former employees, especially senior ones and those with longer tenures, would point to the Omnicrisis and the clawback fiasco in particular as the dawning of a unsettling realization. There were two forces at play in all of this chaos. To be sure, one of them was what had always been: the clash between Boomers and Doomers, which had triggered much of the pile-on of external criticism in the first place. But this time there was also something else: Altman’s power-centralizing behavior in how he’d set up OpenAI’s legal entities; his repeated apparent dishonesty as he sought to explain and move past each mess. After The Blip, many employees had viewed the board’s decision to be entirely a product of the first force, and the directors’ explanations focused on the second as some kind of combination of misdirection or self-delusion. Now, as more and more employees felt for the first time that Altman’s conduct was harming rather than serving them, they wondered whether the board had actually been correct. As the May 23 meeting neared its close, Kwon gave a stilted, heartfelt defense of Altman’s character that seemed strangely out of place. “We want to give you answers when you ask them because we know you want them, like, right now. And I think sometimes we try to give them to you and, you know, we should just wait sometimes. And I think like that—that is like, that is part of, part of this whole thing that’s happened here,” he said, stumbling over his words. “It’s not that there’s intentionality sometimes in all of this. It’s just—I, I really truly think, you know, Sam in particular, he just doesn’t want to let you down. That’s really where it comes from, like, I’ve been working with the guy for a really, really long time. This is why I keep working with him, you know? And so, it just come—it, it does come from a good place. That’s what I’m saying. You can shit-talk me all you want. But you know, that’s, that’s, uh, that’s, yeah, that’s what I got to say.” --- Murati, Brockman, and Pachocki arrived at Sutskever’s house together. On May 23, as OpenAI reeled from the repeated shocks of the Omnicrisis, the three brought with them written cards and gifts from employees and tearfully pleaded with Sutskever to come back to the company. Everything was out of sorts, they told him in an emotional confession. OpenAI was facing a simultaneous loss of trust from employees, investors, and regulators; the company threatened to “collapse” without him. Altman arrived alone later that day, expressing in his own way his hope for Sutskever to rejoin and help restore some semblance of what once was. “Bringing Ilya back would have done a lot to help,” a researcher reflects. “It would be at least a win after a long series of things that made OpenAI look questionable.” Sutskever seriously considered it. Despite everything, he was not one to hold grudges. On the day he announced his departure, Musk had immediately offered Sutskever a role at xAI. In a funny twist of fate, Musk would shift xAI’s headquarters to the Pioneer Building later that year, after OpenAI vacated it. Though he deeply respected Musk, Sutskever declined the offer, resolving instead to build another company. Returning to his first company was more than he’d planned for but everything he wanted. It would be to him a homecoming. Still, he needed assurance, and told the executives as much, that the company would engage in an honest effort to resolve the challenges he’d identified with the painful and disorienting conflicts among leadership. The Omnicrisis could have been a moment for OpenAI to engage in self-reflection. It was a prompt for the company to understand why exactly it had simultaneously lost the trust of employees, investors, and regulators as well as that of the broader public. Only then, maybe, just maybe, it would have begun to realize that both The Blip and the Omnicrisis were one and the same: the convulsions that arise from the deep systemic instability that occurs when an empire concentrates so much power, through so much dispossession, leaving the majority grappling with a loss of agency and material wealth and a tiny few to vie fiercely for control. Instead, OpenAI chose to fortify itself against the criticism. Altman would repeat to employees, as he always had, that the Omnicrisis and The Blip were just the strange and expected moments of madness on the company’s high-stakes noble quest to AGI. “As we all kind of feel that we’re getting closer to finding our way to these powerful systems,” he said during the May 15 all-hands, “the level of stress and tension will internally, externally, directed at us, emanating from us—that keeps going, that keeps increasing.” The best way to manage it would be for OpenAI to double down on its PR, entrench its relationships with governments, and hold steadfast to its convictions in its vision. “I think on the whole we’re quite good at that,” he added, “but we will be tested again here.” How OpenAI handled the Sutskever affair would become just a microcosm of the continued chaos that would manifest from the perpetuation of empire. Sutskever’s request for assurance would spark yet more infighting among leadership, this time with a slightly different cast of characters, replaying the ego-driven dynamics that had plagued the company from the beginning. Aleksander Mądry, the Polish MIT professor who many described as a power seeker, had, in his relatively short tenure, successfully amassed a sizable fiefdom within the company. Mądry didn’t think bringing back Sutskever was a good idea. Sutskever commanded too much admiration and loyalty among researchers. It could take away from Mądry’s influence—as well as the influence of his good friend Pachocki. Within a few hours, Mądry’s concerns had successfully sowed their doubts and fractured leadership. As ever, Altman recused himself from deciding one way or the other to avoid the appearance of disagreeing with anyone. Within twenty-four hours of the executives visiting his house, Sutskever received a call from Brockman. Any discussion of Sutskever’s return, Brockman told him, was now completely off the table.