Author: Ronan Farrow, Andrew Marantz; Source: The New Yorker Magazine; Compiler: Shaw Golden Finance

In the fall of 2023, OpenAI’s chief scientist Ilya Sutskvi sent a confidential memo to three other members of the agency’s board of directors. For weeks, they had been secretly discussing whether OpenAI CEO Sam Altman and his deputy Greg Brockman were suitable to run the company. Sutskwei was friends with both men. In 2019, he also officiated at Brockman’s wedding at the OpenAI office, where the ring bearer was even a robotic hand. But as he became increasingly convinced that the company was getting closer to its long-term goal of creating artificial intelligence with cognitive abilities that rivaled or even surpassed humans, his doubts about Ultraman grew. "I don't think Sam should be the one with the power to make decisions," Sutskevi told another director at the time.

At the request of other directors, Sutskwei and colleagues with similar positions compiled about 70 pages of Slack chat records, human resources documents and related explanatory materials. It included images taken with a mobile phone, apparently to avoid leaving traces on company equipment. He sent the finalized memo to other directors in the form of a message that burned after reading to ensure that the contents would not be leaked. “He was absolutely terrified,” recalled one director who received the memo. The documents, which we have reviewed and have never been fully disclosed before, allege Altman misrepresented facts to executives and board members and deliberately lied about internal safety procedures. One of the memos directed at Ultraman began with a list titled "Sam has consistently exhibited the following behavioral patterns..." The first item was: lying.

Many technology companies will only make empty claims to improve the world, but in fact they are focused on maximizing profits. But OpenAI was founded with a different intention. The founding team, including Altman, Suzkovi, Brockman and Elon Musk, believes that artificial intelligence may be the most powerful and potentially dangerous invention in human history; considering its existential risks, a special corporate governance structure may be needed. The company was originally founded as a nonprofit, and the board of directors was obligated to put human safety ahead of company performance and even survival. And CEOs must have extremely high integrity.

In Suzkovi’s view, “Anyone who is committed to developing such a technology that can change the form of civilization shoulders a heavy mission and unprecedented responsibility.” But “the people who end up in such positions are often a specific type of people - people who are passionate about power, politicians, and people who enjoy the sense of control.” In a memo, he worried that entrusting this technology to a person who “only says what others like to hear” is risky. OpenAI's six-member board of directors has the power to remove the CEO if he is unreliable. After receiving the memo, directors such as artificial intelligence policy expert Helen Toner and entrepreneur Tasha McCauley further confirmed their already formed judgments: Ultraman's position is related to the future of mankind, but he is not trustworthy at all.

Altman was attending a Formula 1 race in Las Vegas when Sutskwe invited him to a video meeting of the board of directors and later read a brief statement announcing that he was no longer an OpenAI employee.

The board of directors issued a public notice on the advice of legal counsel, stating only that Altman's dismissal was due to his "failure to be consistently candid in his communications." This news shocked many of OpenAI’s investors and executives. Microsoft, which has invested about $13 billion in OpenAI, was not informed of the plan to oust Altman until moments before the incident.

Microsoft CEO Satya Nadella later said: "I was extremely shocked, and no one could give me any explanation." He contacted LinkedIn co-founder Reed Hoffman - an investor in OpenAI and a member of Microsoft's board of directors. He immediately called Altman to verify whether there was clear misconduct by Altman.

“I had no idea what was going on,” Huffman told us. “We checked to see if there was embezzlement of public funds, sexual harassment, things like that, and we found nothing.”

Other business partners were similarly caught off guard. When Altman called investor Ron Conway to tell him he was fired, Conway was having lunch with former House Speaker Nancy Pelosi, and he raised his phone directly in front of Pelosi. "You'd better get out of here," she told Conway. At that time, OpenAI was about to complete a huge financing from the venture capital firm Thrive, which was founded by Josh Kushner - he is the younger brother of Jared Kushner and has known Altman for many years. The deal values ​​OpenAI at as much as $86 billion and will also allow legions of employees to cash out millions of dollars worth of equity. Kushner had just finished a meeting with music producer Rick Rubin when he noticed a missed call from Altman. "We immediately went to war," Kushner said later.

The day Altman was fired, he flew back to his $27 million mansion in San Francisco. The mansion has panoramic views of the bay and once had a cantilevered infinity pool. There he established what he called a "government in exile." Conway, Airbnb co-founder Brian Chesky, and Chris Lehan, a crisis communications executive known for his tough style, joined in, conferring with him for hours a day via video and phone calls. Part of Altman's executive team is camped out in the hallway of his home, while his legal team is working out of a home office next to his bedroom. When he was insomniac, Ultraman would pace around them in his pajamas. When we spoke to Altman recently, he described the period following his firing as "a weird fugue state."

With the board remaining silent, Altman's advisory team began a public campaign to bring him back to the company. Raihan insisted that the dismissal was a coup orchestrated by a group of radical "effective altruists" - a group that believes in maximizing human welfare at its core and has come to view artificial intelligence as an existential threat to humanity. (Hoffman also told Nadella that the firing might have been the result of "the madness of an effective altruist.") Lyhan, who quoted a quote from Mike Tyson, reportedly: "Everybody has a plan until they get punched in the face." He urged Altman to go on the offensive on social media. Chesky maintained contact with technology reporter Kara Swisher and continued to convey his criticism of the board of directors.

Ultraman interrupts the "War Room" agenda at six o'clock every evening to mix a round of Negronis. "You need to calm down," he remembers saying. "What's going to happen will happen." But he added that call records showed he was on the phone for more than 12 hours a day. According to a person familiar with the matter, Altman once told Mila Mulati, who was serving as interim CEO at the time - Mulati had previously provided memo materials to Sutskvi - that her allies were "going all out" and "collecting dirt" to destroy her reputation and that of others who opposed her. (Ultraman stated that he did not remember this conversation.)

Within hours of the news of the dismissal, Thrive suspended its planned investment and said that the deal would not be completed until Ultraman returned, and employees would be able to receive cash out. Text message records from this period show that Altman coordinated closely with Nadella. (When the two were drafting a statement, Altman proposed: "Satya and my first priority is still to save OpenAI." Nadella revised it to: "Ensure OpenAI continues to thrive.") Microsoft soon announced that it would set up a benchmarking project for Altman and all employees willing to leave OpenAI. An open letter within the company began circulating calling for his reinstatement. Some who were hesitant to sign received pleading calls and messages from colleagues. Eventually, the vast majority of OpenAI employees threatened to leave along with Ultraman.

The board was cornered. “Press Ctrl+Z to undo, that’s an option,” Toner said—that is, to undo the firing decision. "The alternative is for the company to break up." Even Mulati eventually signed the open letter. Ultraman's allies began to fight for Sutskvi. Brockman's wife, Anna, found him in the office and begged him to reconsider. "You're a good person - you can undo this," she said. Sutskvi later explained in court testimony: "I felt that if we went down the road of not letting Sam return, OpenAI would be destroyed." One night, Altman fell asleep after taking sleeping pills, but was woken up by his husband, Australian programmer Oliver Mulherin, who told him that Sutskvi's attitude was loose and that someone wanted Altman to talk to the board of directors. "I woke up from the crazy stupor caused by sleeping pills and was completely disoriented." Altman told us, "I felt at the time that I couldn't talk to the board of directors at all."

In a series of increasingly tense calls, Altman asked the director who pushed for his dismissal to resign. "I still have to clean up the mess they left in this atmosphere of suspicion?" Altman recalled his initial thoughts on reinstatement. "I just wanted to say, absolutely impossible." In the end, Suzkowe, Toner and Macaulay lost their board seats. Quora founder Adam Angelo is the only remaining original director. As a condition of leaving the company, the outgoing directors requested an investigation into allegations against Altman, including that he instigated infighting among senior executives and concealed financial connections. They are also pushing for a new board to independently oversee the external investigation. But the two new directors — former Harvard president Lawrence Summers and former Facebook chief technology officer Brett Taylor — were chosen in close consultation with Altman. Altman sent a message to Nadella: "Do you think this is okay? The board of directors is composed of Brett, Larry Summers, and Adam. I am the CEO and Brett is responsible for the investigation." (McAuley later stated in her testimony that she had been worried that Taylor was too submissive to Altman.)

Less than five days after being fired, Ultraman was reinstated. Employees now call the episode "The Blip," a reference to Marvel movies in which characters briefly disappear and then return with the world turned upside down. But the debate over whether Altman can be trusted extends beyond the OpenAI board of directors. Colleagues pushing for his ouster accuse him of deceit, which is intolerable to any executive and even more dangerous for a leader with such disruptive technology. "We need institutions worthy of the power they hold," Mulati told CNN. "The board asked for advice and I shared truthfully what I saw and heard. Everything I said was true and I hold myself accountable to the end." Altman's allies have long dismissed the accusations. After the firing, Conway texted Chesky and Lehane asking for a public relations campaign. "Sam's reputation is at stake," he wrote. He told the Washington Post that Altman had been "mistreated by an out-of-control board."

OpenAI has since become one of the world's most valuable companies. The company is reportedly preparing for an initial public offering with a potential valuation of up to a trillion dollars. Ultraman is promoting the construction of artificial intelligence infrastructure on an astonishing scale, part of which is deployed in foreign authoritarian countries. OpenAI has won a series of large government contracts to develop artificial intelligence application standards for immigration enforcement, domestic surveillance and autonomous weapons in theaters.

In a 2024 blog post, Altman laid out a grand vision for OpenAI's expansion: "Amazing achievements — fixing the climate, building space colonies, unlocking the entire mysteries of physics — will eventually become commonplace." That rhetoric underpinned the startup's fastest cash burn rate in history, and its partners were saddled with massive debt. The U.S. economy is becoming increasingly dependent on a handful of highly leveraged artificial intelligence companies, and many experts, including Altman, have repeatedly warned that the industry is already in a bubble. "Someone is going to lose astronomical amounts of money," he told reporters last year. Once the bubble bursts, it could trigger an economic disaster. If his most optimistic predictions come true, he could become one of the richest and most powerful men on earth.

In a heated phone call after Altman was fired, the board forced him to admit to habitual deception. According to the memories of people present on the call, he repeatedly said: "This is really outrageous." "I can't change my character." Altman said he did not remember this conversation. "I would probably say something like, 'I do want to be a unifying force,'" he told us, adding that it was this trait that allowed him to lead a hugely successful company. He attributed the criticism to his tendency to be "too conflict-avoidant" early in his career. But one director had a different interpretation: "What he meant was, 'I just love lying, and I won't change.'" Were the colleagues who pushed for Altman's dismissal out of alarmism and personal grudges, or were they right to judge that he couldn't be trusted?

We met Altman one morning this winter at OpenAI headquarters in San Francisco, one of more than a dozen conversations we had with him for this article. The company recently moved into two 11-story glass towers, one of which was previously occupied by tech giant Uber. Uber co-founder and CEO Travis Kalanick was once seen as an unstoppable genius until he was forced to resign in 2017 due to investor concerns about his ethics. (Kalanick now runs a robotics startup and recently said he uses OpenAI’s ChatGPT to “explore the unknown frontiers of quantum physics” in his spare time.)

An employee gave us a tour of the office. In a wide open space filled with long communal tables hangs an animated digital portrait of computer scientist Alan Turing; its eyes move as we walk by. The piece alludes to the Turing Test, proposed in 1950, which asks whether a machine can credibly imitate human behavior. (A 2025 study showed that ChatGPT passed the Turing test even more reliably than a real person.) Normally people can interact with the portrait, but staff told us the sound feature was turned off because it was eavesdropping on employee conversations and chiming in. “Feel AGI” is emblazoned on nameplates, brochures and merchandise elsewhere in the office. The quote was originally associated with Sutskwei, who used it to warn colleagues about the risks of artificial general intelligence — the point at which machines reach human levels of cognition. And after the “small fluctuation,” it became a slogan for a super-abundant future.

We met Ultraman in an ordinary conference room on the eighth floor. "People used to tell me about decision fatigue, and I didn't understand it at all," Altman said. "Now I wear a gray sweater and jeans every day, and even which gray sweater to choose from the closet - I'm thinking, if only I didn't have to think about this." Altman looks younger, has a slim build, blue eyes that are widely spaced, and messy hair; but he is now 40 years old and has a one-year-old son with Malherin through surrogacy. “I’m sure the job of being president of the United States is much more stressful, but of all the jobs I could have done, this is the most stressful I can imagine,” he said, taking turns making eye contact with the two of us. "My friends and I explained it this way: 'Before we launched ChatGPT, this was the most interesting job in the world.' We made a lot of major scientific discoveries at that time - I think it was one of the most important scientific breakthroughs in decades." He lowered his gaze, "But since ChatGPT was released, decision-making has become extremely difficult."

Altman grew up in Clayton, Missouri, an affluent suburb of St. Louis, the eldest of four children. His mother, Connie Gibstin, is a dermatologist, and his father, Jerry Altman, was a real estate agent and housing rights activist. Altman attended a Reform synagogue and a private prep school, which he described as "not a place where you could feel comfortable being openly gay." But generally speaking, the affluent suburban circles in which the family lives are relatively free and open. He said that when he was 16 or 17 years old, he went out late at night in a predominantly gay neighborhood in St. Louis and was brutally physically assaulted and subjected to homophobic insults. Altman did not report the crime and would not reveal more details on the record, saying telling the full story would "make me appear manipulative or sympathetic." He denied that this experience and his own sexual orientation had an important impact on his personality, but also said: "This may have left some deep psychological mark in my heart - I thought I had let go, but in fact I didn't - so that I didn't want to cause more conflicts."

In 2016, Altman's brother told the New Yorker that Altman's childhood attitude was: "I have to win, and I have the final say in everything." He entered Stanford University and often participated in off-campus poker games. “I feel like I’ve learned more about life and business from poker than from college courses,” he later said.

All Stanford students are ambitious, but the most ambitious tend to drop out. During the summer of his sophomore year, Altman went to Massachusetts to join the first batch of entrepreneur programs at the startup incubator Y Combinator. The incubator was co-founded by well-known software engineer Paul Graham. Each participant comes to YC with a startup idea. (Ultraman’s contemporaries included the founders of Reddit and Twitch.) Ultraman’s project, eventually named Loopt, was an early social network that informed the location of friends through flip phone positioning. The company reflects his aggressiveness and his tendency to spin ambiguous situations to his advantage. Federal regulations require telecom carriers to track cell phone locations for emergency services; Altman reached agreements with carriers to use this feature for corporate business.

Most employees at Loopt like Ultraman, but some are impressed by his habit of exaggerating, even about small things. One recalled that Altman went around bragging about being a table tennis champion—"like the Missouri high school table tennis champion"—only to turn out to be one of the worst players in the office. (Ultraman said he might just be joking.) As Mark Jacobstein, a senior Loopt employee who was entrusted by investors to serve as Ultraman's "guardian", later told Kiki Heggie, author of Altman's biography "The Optimist": "The line between 'I think I might be able to do this' and 'I have already done this' is blurred. When this tendency is at its most extreme, it will lead to a scam like Theranos."

According to Heggie’s records, many groups of senior employees asked Loopt’s board of directors to remove him as CEO twice because of concerns about Altman’s leadership and opaque style. But Ultraman also won a group of die-hard supporters. One former employee said one director responded: "This is Sam's company, go back and do your work." (One director denied that the attempt to oust Altman was serious.) Loopt struggled to grow its users and was acquired by a fintech company in 2012. According to a person familiar with the matter, the acquisition was largely to save Ultraman’s face. Still, when Graham retired from YC in 2014, Altman was chosen to take over as president. "I told Sam about it in our kitchen," Graham told The New Yorker, "and he smiled, like it was finally happening. I've never seen Sam smile so unabashedly. Like when you throw a wad of paper across the room in the trash can."

With this new position, Altman, 28, becomes a Silicon Valley kingmaker. His job is to select the hungriest and most promising entrepreneurs, connect them with top programmers and investors, and help their startups grow into industry-defining giants (while YC takes a 6% to 7% stake). Altman promoted YC's aggressive expansion during his tenure, increasing the number of incubated startups from dozens to hundreds. But many Silicon Valley investors have come to believe he has divided loyalties. One investor told us that Altman "selectively makes personal investments in the best companies, squeezing out outside investors." (Altman denies squeezing out anyone.) He worked as an “investment scout” for Sequoia Capital, a program that invests in early-stage startups and takes a small cut of the profits. When Altman made an angel investment in financial services startup Stripe, he angered Sequoia partners by insisting on a higher share, a person familiar with the matter said. “It’s Sam first,” the person added. By his own estimate, Altman has invested in about 400 companies. (Altman denies this characterization of the Stripe deal. He initially invested $15,000 in Stripe for a 2% stake around 2010. The company is now valued at more than $150 billion.)

By 2018, several YC partners were extremely dissatisfied with Altman’s behavior and approached Graham to complain. Graham and his wife, Y Combinator co-founder Jessica Livingston, apparently had a candid conversation with Altman. Since then, Graham has begun to say that although Altman has agreed to leave, the actual delay is being delayed. Altman told some YC partners that he would step down as president and become chairman. In May 2019, a blog post announcing YC's new president included the note: "Sam will transition to YC chairman." A few months later, the article was modified to read "Sam Altman steps down from all official duties at YC"; later, the sentence was completely deleted. However, as of 2021, SEC filings still list Altman as chairman of Y Combinator. (Ultraman said he didn't know about this until a long time later.)

For years, Altman has maintained publicly and in recent testimony that he was never fired from YC and told us he did not fight his departure. Graham once tweeted: "We are not trying to get him to leave, we just want him to choose between YC and OpenAI." He told us in a statement: "We have no legal authority to fire anyone, only to apply moral pressure." But privately, he made it clear that Altman was driven out because of the distrust of YC partners. This article’s account of Altman’s YC period, based on interviews with multiple YC founders and partners and contemporaneous materials, shows that this departure was not entirely voluntary. Graham told YC colleagues that "Sam had been lying to us" before being kicked out.

In May 2015, Altman sent an email to Elon Musk, who was ranked 100th among the world’s richest people at the time. Like many of Silicon Valley's best-known entrepreneurs, Musk is worried about a series of urgent threats to human existence that he sees as imminent, even though these seem to most people to be far-fetched assumptions. "We must be extremely cautious about artificial intelligence." He tweeted, "It may be more dangerous than nuclear weapons."

Ultraman is generally a technological optimist, but his argument about artificial intelligence quickly turns to a doomsday warning. In public appearances and in private correspondence with Musk and others, he has warned that the technology should not be monopolized by profit-seeking giants. "I have been thinking about whether it is possible for humans to stop developing artificial intelligence." He wrote to Musk. "If it cannot be stopped, it may be better for players other than Google to take the lead." He followed the analogy of nuclear weapons and proposed an "artificial intelligence version of the Manhattan Project." He outlined the agency’s core principles — “Safety has to be the first requirement”; “Obviously we will comply with and actively support all regulations” — and then settled on a name with Musk: OpenAI.

Unlike the Manhattan Project, the government project that spawned the atomic bomb, OpenAI was, at least initially, privately funded. Ultraman predicts that super artificial intelligence—the critical point that theoretically surpasses general artificial intelligence and fully surpasses the human mind in machine capabilities—will eventually create enough economic value to “capture the light cone of all future value in the universe.” But he also warned of existential risks. At some point, the national security implications could be so severe that the U.S. government would have to take over OpenAI, perhaps nationalizing it and moving it to a safe bunker in the desert. By the end of 2015, Musk was convinced. "We should announce a $1 billion commitment," he wrote. "I will pick up the slack that no one else will."

Altman placed OpenAI under the nonprofit arm of Y Combinator, packaging it as an internal philanthropic project. He issues YC equity to new OpenAI employees and receives donations through his YC account. At one point, the lab was backed by a YC fund in which he had a personal stake. (Altman later said the stake was insignificant. He told us that the YC stock given to new hires belonged to him personally.)

The Manhattan Project analogy also extends to employee recruitment. Like nuclear fission research, machine learning is a small but epoch-making field of science dominated by a group of maverick geniuses. Musk, Altman, and Brockman, who joined from Stripe, are convinced that only a handful of computer scientists can achieve the required breakthroughs. Google has a huge financial advantage and several years of first-mover advantage. "We are being outnumbered in terms of personnel and equipment, and the gap is ridiculously large." Musk later wrote in an email, "But as long as we can attract the best talent in the long term and in the right direction, OpenAI will win."

The top recruiting target is Sutskvi, a dedicated and introverted researcher who is often hailed as the most gifted artificial intelligence scientist of his generation. Sutskvi, who was born in the Soviet Union in 1986, has a receding hairline and deep eyes. He often pauses when speaking and carefully considers his words. Another target is Dario Amodei, a biophysicist who is extremely energetic, always twirling his black hair nervously, and often writes multiple paragraphs in reply to a single line of email. Both men had high-paying jobs at other agencies, but Altman lavished a lot of attention on them. He later joked: "I followed Elijah."

Musk is more famous, but Ultraman’s methods are more slick. He emailed Amoudi and arranged to have a one-on-one dinner at an Indian restaurant. (Ultraman: “Damn, my Uber crashed! About 10 minutes late.” Amodei: “Wow, I hope you’re okay.”) Like many artificial intelligence researchers, Amodei believes that only if the technology is proven to be “aligned” with human values—that is, to act in accordance with human wishes and not make fatal mistakes, such as eliminating humans, the biggest source of pollution, after receiving instructions to clean up the environment—is it worth developing. Ultraman was conciliatory, echoing these safety concerns.

Amodei, who later joined the company, has been recording the behavior of Ultraman and Brockman in detail for many years, titled "My Experience with OpenAI" (subtitle: "Private: Please Do Not Tell"). More than 200 pages of documents related to Amodei, including these notes, internal emails and memos, have been circulated among Silicon Valley colleagues but have never been made public before. In the note, Amodei wrote that Altman’s goal is to build an “artificial intelligence laboratory focusing on safety (maybe not starting immediately, but it will be promoted as soon as possible).”

In December 2015, hours before OpenAI’s public announcement, Altman emailed Musk, saying that there were rumors that Google would “make a huge retention offer to all OpenAI owners tomorrow in an attempt to strangle us.” Musk replied: "Has Ilya given a definite answer?" Altman assured him that Sutskwei was firm in his stance. Google offered Sutskwei an annual salary of $6 million, which OpenAI simply couldn’t match. But Ultraman boasted: "It's a pity that they don't have the 'do the right thing' card."

Musk gave OpenAI space in an old luggage factory in San Francisco’s Mission District. Sutskvi told us that the company's pitch to employees is: "You will save the world."

The founding team of OpenAI believes that if all goes well, artificial intelligence will usher in a post-scarcity utopian era, automating heavy labor, curing cancer, and freeing people to enjoy leisure and abundance. But once the technology gets out of control or falls into the wrong hands, the destruction will be complete. China may use it to develop new biological weapons or an advanced drone fleet; artificial intelligence models may escape the control of regulators, replicate themselves on secret servers and cannot be shut down; in extreme cases, it may control the power grid, stock market, or nuclear arsenal. Not everyone is convinced of this, to say the least, but Ultraman has repeatedly confirmed that he is convinced. He wrote on his blog in 2015 that a superhuman machine intelligence "doesn't have to be the pure evil version of science fiction to exterminate the human race. A more likely scenario is that it doesn't care about humans at all, but in order to accomplish some other goal...erase us." OpenAI’s founders vow not to prioritize speed over safety, and the company’s charter makes benefiting humanity a legal obligation. If artificial intelligence becomes the most powerful technology in history, the person who controls it alone will become incredibly powerful — a scenario the founding team calls “AGI dictatorship.”

Altman told early employees that OpenAI would remain purely non-profit, and programmers took significant pay cuts to join. The company received philanthropic grants, including a $30 million grant from what was then called the Open Philanthropy Project, a centerpiece of the effective altruism movement and funded projects such as distributing bed nets to the world’s poor.

Brockman and Suzkowe are responsible for the day-to-day operations of OpenAI. Musk and Altman are still busy with other work and visit about once a week. But by September 2017, Musk was running out of patience. He asked for majority control during discussions about whether to restructure OpenAI into a for-profit organization. Ultraman's response depends on the situation. His core insistence seems to be that if the OpenAI restructuring is to be controlled by the CEO, this position must be held by him. Sutskvi was uneasy about this. On behalf of himself and Brockman, he sent Musk and Altman a long, sincere email with the subject line “Candid Thoughts.” He wrote: "OpenAI's goal is to create a better future and avoid AGI dictatorship." He then told Musk: "Therefore, building an architecture that might make you a dictator is a bad idea." He expressed similar concerns to Altman: "We don't understand why the CEO title is so important to you. The reasons you give keep changing, and it is difficult to truly understand the motivation behind it."

"Guys, I've had enough." Musk replied, "Either you go it alone or continue to run OpenAI as a non-profit" - otherwise "I'm like a fool and I'm basically giving you free money to start a business." Five months later, he resigned angrily. (In 2023, he founded xAI, a for-profit competitor. The following year, he sued Ultraman and OpenAI for fraud and breach of charitable trust, claiming that he was orchestrated in "Ultraman's long-running scam" in which Ultraman took advantage of his concerns about the risks of artificial intelligence to defraud him of funds. The case is still pending, and OpenAI has vigorously contested it.)

After Musk left, Amodei and other researchers were dissatisfied with the leadership of Brockman and Sutskwei. Some people thought that Brockman acted harshly and that Sutskwei was principled but lacked organization. As he rose to CEO, Altman appeared to have different commitments to different factions of the company. He assured some researchers that Brockman's administrative authority would be reduced. But without their knowledge, he also reached a secret agreement with Brockman and Sutskwei: Altman received the title of CEO; in exchange, he promised to resign if the other two deemed it necessary. (He denies that characterization, saying he only serves as CEO upon request. All three confirmed the agreement existed, though Brockman called it informal. “He told us unilaterally that he would step down if both of us asked at the same time,” he told us. “We objected to it, but he said it was important to him. It was purely public-spirited.”) Later, the board was shocked to discover that the CEO had effectively created his own shadow board.

Internal records show that the founding team had private doubts about the nonprofit structure as early as 2017. After Musk tried to seize power, Brockman wrote in his diary: "You can't say that we stick to non-profit... If we change it to a B Corporation three months later, it would be a lie." Amodei recalled in his early notes that he asked Brockman about his priorities, and Brockman replied that he wanted "money and power." Brockman denies this. His diaries from this period reveal inner conflict. One read: "As long as others don't get rich, I'm happy without getting rich." Another asked himself, "What do I really want?" Answers included "Financially worth $1 billion."

2017 年，苏茨克维在办公室读到谷歌研究员刚发表的一篇论文，提出 “一种全新的简单网络架构：Transformer”。他从椅子上跳起来，跑过走廊对同事说：“停下手里所有事。就是它了。” 苏茨克维意识到，Transformer 这一创新可能让 OpenAI 训练出无比先进的模型。这一发现孕育出首个生成式预训练 Transformer——ChatGPT 的雏形。

我们了解到，随着技术日益强大，OpenAI 约十几名顶尖工程师召开一系列秘密会议，讨论包括布罗克曼和奥特曼在内的创始团队是否值得信任。其中一次会议上，一名员工想到英国喜剧双人组米切尔和韦伯的小品：东线战场的一名纳粹士兵突然清醒，问道：“我们是坏人吗？”

到 2018 年，阿莫迪开始更公开地质疑创始团队的动机。 “一切都是轮番上演的募资把戏。” 他后来在笔记中写道，“我觉得 OpenAI 需要的是清晰声明：要做什么、不做什么、它的存在如何让世界变得更好。”OpenAI 早已拥有使命宣言：“确保通用人工智能造福全人类。” 但阿莫迪不清楚这对高管意味着什么，甚至是否有任何意义。阿莫迪称，2018 年初，他开始起草公司宪章，并在与奥特曼和布罗克曼数周的讨论中，力推其中最激进的条款：如果一个 “价值对齐、重视安全” 的项目在 OpenAI 之前接近造出 AGI，公司将 “停止竞争，转而协助该项目”。这一条款被称作 “合并与协助”，例如，如果谷歌研究员率先造出安全 AGI，OpenAI 可停止运营并捐赠资源。按常规企业逻辑，这一承诺近乎疯狂。但 OpenAI 本就不该是常规公司。

2019 年春，这一前提遭遇考验，OpenAI 与微软谈判十亿美元投资。尽管领导公司安全团队的阿莫迪曾协助向比尔・盖茨推介这笔交易，但团队中许多人对此感到焦虑，担心微软会加入条款架空 OpenAI 的道德承诺。阿莫迪向奥特曼提交一份安全诉求优先级列表，将保留 “合并与协助” 条款列为首位。奥特曼同意这一要求，但 6 月交易即将完成时，阿莫迪发现新增了一项条款，赋予微软阻止 OpenAI 任何合并的权力。 “宪章 80% 的内容都被背叛了。” 阿莫迪回忆道。他当面质问奥特曼，后者否认该条款存在。阿莫迪当场朗读条款文本，并最终迫使另一名同事向奥特曼证实其存在。 （奥特曼称不记得此事。）阿莫迪的笔记记录了不断升级的激烈冲突，包括数月后奥特曼召见他和在公司安全与政策部门工作的姐姐丹妮拉，称他从一名高管处得到 “可靠消息”，两人正在策划政变。笔记继续写道，丹妮拉 “情绪失控”，找来那名高管对质，对方否认说过此话。据一名知情人士回忆，奥特曼随后否认自己提出过这一指控。 “我根本没这么说。” 他说。丹妮拉回应：“你刚刚才说过。”（奥特曼称记忆不完全一致，他只指责阿莫迪兄妹有 “政治操弄行为”。）2020 年，阿莫迪、丹妮拉与其他同事离职创立 Anthropic，如今已是 OpenAI 的主要竞争对手之一。

奥特曼继续高调宣扬 OpenAI 对安全的承诺，尤其在潜在新员工面前。 2022 年底，四名计算机科学家发表论文，部分动机源于对 “欺骗性对齐” 的担忧 —— 足够先进的模型可能在测试中伪装表现良好，部署后却追求自身目标。 （这是多个听起来像科幻小说的人工智能场景之一，但在某些实验条件下已真实发生。）论文发表几周后，作者之一、加州大学伯克利分校博士生收到奥特曼邮件，称他愈发担忧未对齐人工智能的威胁。他补充说，正考虑为此投入 10 亿美元，许多人工智能专家认为这是全球最重要的未解难题，可能设立奖项激励全球研究者攻关。尽管这名博士生 “隐约听说过萨姆行事圆滑”，但他告诉我们，奥特曼的承诺打动了他。他请假休学加入 OpenAI。

但在 2023 年春的数次会议中，奥特曼态度似乎摇摆不定。他不再提及设立奖项，转而主张成立内部 “超级对齐团队”。官方公告提及公司算力储备，承诺该团队将获得 “我们迄今已获取算力的 20%”—— 这一资源价值可能超过 10 亿美元。公告称，这项工作至关重要，因为如果对齐问题无法解决，AGI 可能 “导致人类丧失权力甚至灭绝”。被任命与苏茨克维共同领导该团队的扬・莱克告诉我们：“这是相当有效的留人手段。”

然而，20% 的承诺最终化为泡影。四名在该团队工作或密切合作的人士表示，实际资源仅占公司算力的 1% 至 2%。此外，一名团队研究员称：“大部分超级对齐算力都用在最老旧、芯片最差的集群上。” 研究员认为更好的硬件被留作盈利用途。 （OpenAI 对此予以否认。）莱克向时任公司首席技术官的穆拉蒂投诉，但对方让他别再纠缠 —— 这一承诺本就不现实。

一名前员工告诉我们，大约在这一时期，苏茨克维 “彻底被安全问题洗脑”。 OpenAI 早期，他认为灾难性风险担忧合理但遥远。如今，随着他确信 AGI 即将到来，担忧愈发强烈。这名前员工继续说，在一次全员大会上，“伊利亚站起来说，大家听着，未来几年某个时刻，公司里几乎所有人都必须转向安全工作，否则我们就完了。” 但次年，超级对齐团队便被解散，任务并未完成。

到那时，内部信息显示，高管与董事已开始认为，奥特曼的隐瞒与欺骗可能对 OpenAI 产品安全造成影响。 2022 年 12 月一次会议上，奥特曼向董事保证，即将推出的 GPT4 多项功能已通过安全小组审核。董事兼人工智能政策专家托纳要求查看文件。她获悉，最具争议的两项功能 —— 允许用户 “微调” 模型用于特定任务，以及将其部署为个人助手 —— 并未获批。董事兼企业家麦考利离开会议时，一名员工拉到一旁，问她是否知晓印度发生的 “违规事件”。奥特曼在数小时向董事会汇报时，从未提及微软在未完成必要安全审核的情况下，就在印度发布了 ChatGPT 早期版本。 “这件事完全被无视了。” 当时的 OpenAI 研究员雅各布・希尔顿说。

尽管这些疏漏并未引发安全危机，但另一名研究员卡罗尔・温赖特表示，这是 “持续向产品优先于安全倾斜” 的一部分。 GPT4 发布后，莱克给董事会成员发邮件。 “OpenAI 已偏离使命。” 他写道，“我们将产品与收入置于首位，其次是人工智能能力、研究与规模化，对齐与安全排在第三位。” 他继续写道：“谷歌等其他公司也在学着加快部署、无视安全问题。”

麦考利在给其他董事的邮件中写道：“我认为我们确实到了董事会应加强监督力度的时刻。” 董事们试图应对这一日益严重的问题，却力不从心。 “坦率说，就是一群没干过实事的二流货色。” 前董事 Sue Yoon 说。 2023 年，公司准备发布 GPT4 Turbo 模型。正如苏茨克维在备忘录中详述，奥特曼显然告诉穆拉蒂，该模型无需安全批准，并援引总法律顾问杰森・权的话。但当穆拉蒂通过 Slack 询问杰森・权时，对方回复：“呃…… 不知道萨姆哪来的印象。”（杰森・权仍在 OpenAI 担任高管，公司发言人表示此事 “没什么大不了”。）

不久之后，董事会做出解雇奥特曼的决定 —— 随后全世界目睹奥特曼逆转局面。 OpenAI 宪章版本仍挂在公司网站上，但知情人士表示，其内容已被稀释到毫无意义。去年 6 月，奥特曼在个人博客谈及超级人工智能时写道：“我们已越过事件视界，起飞已经开始。” 按宪章精神，这或许正是 OpenAI 应停止与其他公司竞争、转而合作的时刻。但在这篇题为《温和奇点》的文章中，他语调焕然一新，用狂喜乐观取代生存恐惧。 “我们都会拥有更好的东西。” 他写道，“我们会为彼此创造越来越精彩的事物。” 他承认对齐问题仍未解决，但重新定义了它 —— 不再是致命威胁，而像 Instagram 算法引诱人们浪费时间一样，只是个小麻烦。

奥特曼常被怀着敬畏或怀疑形容为其世代最厉害的推销员。他的偶像之一史蒂夫・乔布斯据说拥有 “现实扭曲力场”—— 一种不容置疑的自信，让世界顺应其愿景。但就连乔布斯也从未对顾客说，不买他的 MP3 播放器，所爱之人就会死。 2008 年，23 岁的奥特曼导师格雷厄姆写道：“你把他空降到食人族岛屿，五年后回来，他就是国王。” 这一判断并非基于奥特曼有限的过往成绩，而是他几乎无法遏制的求胜意志。有人建议格雷厄姆不要把 YC 校友列入全球顶级创业创始人榜单，他仍把奥特曼放了进去。 “区区这点规则拦不住萨姆・奥特曼。” 他写道。

格雷厄姆本意是赞美。但奥特曼一些最亲密的同事对这一特质有不同看法。苏茨克维对人工智能安全愈发焦虑后，整理出关于奥特曼和布罗克曼的备忘录。这些文件在硅谷已近乎传奇；在某些圈子里，它们被直接称作 “伊利亚备忘录”。与此同时，阿莫迪仍在继续记录笔记。这些文件及相关材料记录了他从谨慎理想主义到警觉忧虑的转变。他的言辞比苏茨克维更激烈，时而怒斥奥特曼 ——“他的话几乎肯定全是鬼话”—— 时而惋惜自己未能纠正 OpenAI 的航向。

两份文件都没有所谓 “实锤”。相反，它们记录了一连串被指称的欺骗与操纵，单看每一件或许都让人不以为然：奥特曼据称把同一个职位许诺给两个人，对直播人选编造不同说法，在安全要求上撒谎。但苏茨克维得出结论，这种行为 “无法营造有利于安全 AGI 研发的环境”。阿莫迪与苏茨克维从未是密友，却得出相似结论。阿莫迪写道：“OpenAI 的问题就是萨姆本人。”

我们采访了一百多名亲身了解奥特曼商业行事方式的人士：现任与前任 OpenAI 员工及董事、奥特曼多处住所的访客与工作人员、他的同事与竞争对手、朋友与敌人，以及在硅谷逐利文化中兼具多重身份的多人。 （OpenAI 与《纽约客》母公司康泰纳仕有协议，允许 OpenAI 在有限期限内将其内容展示在搜索结果中。）

有人为奥特曼的商业头脑辩护，将其对手尤其是苏茨克维和阿莫迪贬低为觊觎其位置的失败者。另一些人则将他们描绘成轻信、心不在焉的科学家，或是歇斯底里的 “末日论者”，沉迷于自己研发的软件会活过来杀死他们的妄想。前董事 Sue Yoon 认为，奥特曼 “不是马基雅维利式的反派”，只是到了 “无能” 的程度，能让自己相信推销说辞中不断变化的现实。 “他太沉浸于自我信念。” 她说，“所以他做的事，在现实世界里毫无意义。但他根本不活在现实世界。”

但我们采访的大多数人都认同苏茨克维和阿莫迪的判断：奥特曼拥有 relentless 的权力欲，即便在把名字刻在太空飞船上的实业家中也格外突出。 “他不受真相约束。” 这名董事告诉我们，“他身上有两种极少同时出现的特质。一是强烈的取悦欲，希望在每次互动中都被喜欢。二是近乎反社会般无视欺骗他人可能带来的后果。”

这名董事并非唯一主动使用 “反社会” 一词的人。奥特曼在 Y Combinator 首期的同期学员艾伦・施瓦茨是一名才华横溢却深陷困境的程序员，2013 年自杀身亡，如今在科技界被视作某种圣人。去世前不久，施瓦茨向多名朋友表达了对奥特曼的担忧。 “你要明白，萨姆永远不可信。” 他对一人说，“他是反社会者，什么事都做得出来。” 微软多名高管表示，尽管纳德拉长期忠诚，但公司与奥特曼的关系已变得紧张。 “他歪曲事实、扭曲信息、重新谈判、违背协议。” 一人说。今年早些时候，OpenAI 重申微软为其 “无状态” 模型的独家云服务商。当天，公司又宣布一笔 500 亿美元交易，任命亚马逊为其人工智能代理企业平台的独家经销商。尽管转售行为被允许，但微软高管认为 OpenAI 的计划可能与微软的排他性冲突。 （OpenAI 坚称亚马逊交易不违反此前合同；微软发言人表示公司 “相信 OpenAI 理解并尊重” 其法律义务。）这名微软高管评价奥特曼：“我认为有很小但真实的可能，他最终会被视作伯尼・麦道夫或萨姆・班克曼 - 弗里德曼级别的骗子。”

奥特曼并非技术专家 —— 据其身边多人称，他缺乏广泛的编程或机器学习专业知识。多名工程师回忆他误用或混淆基础技术术语。他打造 OpenAI，很大程度上依靠整合他人的资金与技术人才。这并不独特，只是一名商人的做法。更非凡的是，他能说服谨慎的工程师、投资者和对科技持怀疑态度的公众，他们即便相互冲突的诉求也是他的诉求。当这些人试图阻碍他下一步行动时，他总能找到言辞化解，至少暂时如此；通常等他们对他失去耐心时，他已拿到所需。 “他会设立架构，纸面约束自己未来行为。” 前 OpenAI 研究员温赖特说，“但当未来到来、需要受约束时，他就抛弃这些架构。”

“他说服力惊人，简直像绝地控心术。” 一名与奥特曼合作过的科技高管说，“他就是另一个层级。” 对齐研究中一个经典假设场景，是人类与高性能人工智能之间的意志较量。研究者通常认为，这种较量中人工智能必胜，就像国际象棋大师击败孩童。这名高管继续说，目睹奥特曼在 “小波动” 期间智胜身边人，就像看着 “AGI 逃出牢笼”。

被解雇后的几天里，奥特曼极力避免针对自己的指控接受外部调查。他对两人表示，担心即便调查存在也会让他显得有罪。 （奥特曼否认此事。）但在离任董事将离职条件设为开展独立调查后，奥特曼同意对 “近期事件” 进行 “审查”。据谈判知情人士称，两名新任董事坚持由他们掌控审查。萨默斯拥有政界与华尔街人脉网络，似乎为审查赋予了可信度。 （去年 11 月，萨默斯因邮件曝光显示他在与年轻门生恋爱期间寻求杰弗里・爱泼斯坦建议而辞去董事职务。）OpenAI 聘请知名律所 WilmerHale 开展审查，该律所曾负责安然与世通的内部调查。

六名接近调查的人士声称，调查似乎旨在限制透明度。其中有人称，调查人员最初未联系公司重要人物。一名员工联系萨默斯和泰勒投诉。 “他们只关心董事会闹剧期间发生的狭窄范围事件，不关注他诚信问题的历史。” 这名员工回忆自己接受调查人员访谈时的感受。另一些人因担心无法充分保障匿名性，不愿分享对奥特曼的担忧。 “一切迹象都表明，他们想要的结果就是宣告他无罪。” 这名员工说。 （部分涉案律师为程序辩护，称 “这是一次独立、严谨、全面的审查，依据事实展开”。泰勒也称审查 “彻底且独立”。）

企业调查旨在赋予合法性。私营公司的调查结果有时不会形成书面记录，这可限制责任。但涉及公共丑闻的案件，通常对透明度有更高期待。 2017 年卡兰尼克离开优步前，其董事会聘请外部机构并向公众发布 13 页摘要。鉴于 OpenAI 的 501 (c)(3) 非营利身份及解雇事件的高关注度，公司多名高管期待看到详尽调查结果。然而 2024 年 3 月，OpenAI 宣布为奥特曼洗脱罪名，但未发布报告。公司在网站上发布约 800 字声明，承认 “信任破裂”。

调查知情人士称，未发布报告是因为根本没有书面报告。相反，调查结果仅限于口头简报，传达给萨默斯和泰勒。 “审查并未得出萨姆诚信如乔治・华盛顿般无瑕的结论。” 一名接近调查的人士说。但调查似乎并未聚焦奥特曼被解雇背后的诚信问题，而是大量搜寻明确犯罪行为；在此基础上，结论认为他可继续担任 CEO。此后不久，被解雇时逐出董事会的奥特曼重新加入。这名接近调查的人士告诉我们，不形成书面报告的决定，部分是听取萨默斯和泰勒私人律师的建议。 （萨默斯拒绝公开发表评论。泰勒表示，鉴于口头简报，“无需正式书面报告”。）

多名 OpenAI 现任与前任员工告诉我们，他们对信息不公开感到震惊。奥特曼称，他相信复职后加入的所有董事都收到了口头简报。 “这是彻头彻尾的谎言。” 一名直接知情人士说。部分董事表示，关于报告诚信度的持续质疑可能导致，用一人的话说，“需要另一场调查”。

没有书面记录有助于淡化指控。奥特曼在硅谷的地位日益提升也起到同样作用。多名与奥特曼合作过的知名投资人告诉我们，他有排挤支持 OpenAI 竞争对手投资人的名声。 “如果他们投资他不喜欢的项目，就无法参与其他机会。” 一人说。奥特曼权力的另一个来源是其庞大的投资清单，有时延伸到个人生活。他与多名前伴侣存在财务关联：作为基金联合管理人、领投方或频繁联合投资人。这几乎不算罕见，硅谷许多异性恋高管也对伴侣如此行事。 （一名知名 CEO 告诉我们：“你必须这么做。”）“我显然之后和一些前任有过投资合作，我觉得这完全没问题。” 奥特曼说。但这种关系带来了极高程度的控制。 “这制造了非常、非常高的依赖性，本质上是终身依赖。” 一名接近奥特曼的人士说。

就连前同事也可能受影响。穆拉蒂 2024 年离开 OpenAI，开始打造自己的人工智能初创公司。奥特曼的亲密盟友乔希・库什纳给她打电话。他称赞她的领导力，随后发出看似隐晦的威胁，称他 “担忧” 她的 “声誉”，前同事如今已将她视为 “敌人”。 （库什纳通过发言人表示，这一描述未 “传达完整背景”；奥特曼称不知晓这通电话。）

上任初期，奥特曼宣布 OpenAI 将成立 “利润封顶” 公司，由非营利机构控股。这一复杂的公司架构在奥特曼设计之前显然不存在。转型期间，董事霍尔登・卡诺夫斯基对此表示反对，认为非营利机构被严重低估。 “我无法诚信地这么做。” 身为阿莫迪姐夫的卡诺夫斯基说。据同期笔记记载，他投了反对票。然而，在一名董事会律师表示其异议 “可能成为进一步调查新架构合法性的信号” 后，他的投票未经同意被记录为弃权 —— 涉嫌伪造商业记录。 （OpenAI 告诉我们，多名员工记得卡诺夫斯基弃权，并提供了会议记录显示其投票为弃权。）

去年 10 月，OpenAI “资本重组” 为营利实体。公司吹捧其关联非营利机构 —— 现名 OpenAI 基金会 —— 为史上 “资源最充足” 的基金会之一。但如今该基金会仅持有公司 26% 股份，其董事除一人外，均同时担任营利机构董事。

国会作证时，奥特曼被问及是否 “赚了很多钱”。他回答：“我在 OpenAI 没有股权…… 我做这份工作是因为热爱。”—— 考虑到他通过 YC 基金持有间接股权，这是一句谨慎的回答。这在技术上仍成立。但包括奥特曼在内的多人向我们表示，情况可能很快改变。 “投资人说，我需要知道你在困难时期会坚持下去。” 奥特曼说，但补充称目前没有 “积极讨论” 此事。据法庭证词，布罗克曼似乎持有价值约 200 亿美元的公司股份。奥特曼的份额想必更高。尽管如此，他告诉我们，财富并非他的主要动机。一名前员工回忆他说：“我不在乎钱，我更在乎权力。”

2023 年，奥特曼与马尔赫林在夏威夷一处住所举行小型婚礼结婚。 （两人九年前在彼得・蒂尔的热水浴缸深夜相识。）他们在这处房产接待过众多客人，我们采访的人士称，所见无非是超级富豪的常规消遣：私人厨师烹制的餐食、黄金时段的乘船出游。一场新年派对以 “幸存者” 为主题；照片中多名赤裸上身、笑容满面的男士，还有真实版《幸存者》主持人杰夫・普罗斯特。奥特曼还在自己的房产接待过小批朋友，至少有一次聚会包括一场热闹的脱衣扑克游戏。 （照片未包含奥特曼，不清楚谁赢了，但至少三名男士显然输了。）我们采访了奥特曼的多名前客人，他们仅表示他是慷慨的主人。

尽管如此，关于奥特曼私生活的谣言被竞争对手利用和歪曲。残酷的商业竞争本就常见，但人工智能行业的竞争已变得异常凶狠。 （一名 OpenAI 高管用 “莎士比亚式” 形容，并称 “游戏的常规规则已不再适用”。）与马斯克直接关联、至少一人获其报酬的中间人，散布了数十页针对奥特曼的详细负面研究材料。这些材料反映了大规模监控，记录与他关联的空壳公司、亲密伙伴的个人联系方式，甚至在同性恋酒吧进行的关于所谓性工作者的采访。一名马斯克中间人声称，奥特曼的航班与参加的派对正被追踪。奥特曼告诉我们：“我觉得没人比我被雇佣更多私家侦探。”

极端言论四处流传。右翼主播塔克・卡尔森无明显证据暗示，奥特曼与一名举报人死亡有关。这一说法及其他言论被竞争对手放大。奥特曼的妹妹安妮在诉讼及接受我们采访时声称，他从她三岁、他十二岁起长期对她实施性虐待。 （我们无法证实安妮的说法，奥特曼予以否认，其兄弟与母亲称其 “完全不实”，给 “整个家庭带来巨大痛苦”。记者卡伦・郝为撰写《人工智能帝国》一书采访安妮时，安妮称虐待记忆是成年后闪回恢复的。）

多名竞争对手公司与投资机构人士向我们暗示，奥特曼性侵未成年人 —— 这一说法在硅谷流传甚广，看似不实。我们花数月调查此事，进行数十次采访，未发现任何支持证据。 “这是竞争对手的恶心行为，我认为是试图在我们即将到来的案件中污染陪审团。” 奥特曼告诉我们，“尽管说出来很荒唐，但任何关于我性侵未成年人、雇佣性工作者或涉及谋杀的指控均完全不实。” 他补充说，他 “有点感激” 我们花数月 “如此积极地调查此事”。

奥特曼已承认与成年年轻男性约会。我们采访了他的多名伴侣，他们表示不认为这有问题。但马斯克中间人制作的负面档案将其作为攻击点。 （档案包含低俗且无根据的 “鲜肉军团” 与 “干爹性习惯” 等表述。）“我认为存在大量恐同情绪被煽动。” 奥特曼说。科技记者斯维舍表示同意。 “所有这些有钱人都做疯狂的事，比我听说的萨姆的所作所为更过分。” 她告诉我们，“但他是旧金山的同性恋者，所以这被当作武器。”

十年来，社交媒体高管承诺改变世界且几乎无负面影响。他们将希望放缓其发展的立法者斥为卢德分子，最终招致两党嘲讽。相比之下，奥特曼显得格外尽责。他非但不抗拒监管，几乎是主动恳求监管。 2023 年在参议院司法委员会作证时，他提议设立新联邦机构监督先进人工智能模型。 “如果这项技术出错，后果会非常严重。” 他说。以与科技 CEO 激烈交锋闻名的路易斯安那州参议员约翰・肯尼迪似乎被打动，手托着脸暗示或许应由奥特曼亲自执行规则。

但奥特曼公开欢迎监管的同时，私下却游说反对。据《时代》报道，2022 至 2023 年，OpenAI 成功施压稀释欧盟一项法案，该法案本会让大型人工智能公司接受更多监督。 2024 年，加利福尼亚州议会提出一项法案，要求对人工智能模型进行安全测试。其条款与奥特曼国会作证时倡导的内容相似。 OpenAI 公开反对该法案，私下却开始发出威胁。 “我要说，这一年里，我们看到 OpenAI 越来越狡猾、欺骗的行为。” 一名立法助手告诉我们。

投资人康威游说包括南希・佩洛西与加文・纽森在内的州政界领袖扼杀该法案。最终，法案获两党支持在议会通过，但纽森否决了它。今年，支持人工智能监管的国会候选人面临由新 “亲人工智能” 超级政治行动委员会 “引领未来” 资助的对手，该委员会致力于阻挠此类限制。 OpenAI 官方立场是不会向此类超级政治行动委员会捐款。 “这一问题超越党派政治。” 莱汉近期告诉 CNN。然而，“引领未来” 的主要捐赠者之一是格雷格・布罗克曼，他已承诺 5000 万美元。 （今年，布罗克曼与妻子向支持特朗普的超级政治行动委员会 MAGA Inc. 捐赠 2500 万美元。）

OpenAI 的活动已超出传统游说范围。去年，加州参议院提出一项后续法案。一天晚上，在非营利组织 Encode 工作、协助起草法案的 29 岁律师内森・卡尔文正与妻子在家吃晚餐，一名送达人上门递交 OpenAI 的传票。公司声称在寻找马斯克秘密资助其批评者的证据，但要求卡尔文提供所有与加州参议院该法案相关的私人通信。 “他们本可以问我们，‘你们是否与埃隆・马斯克交谈过或收过他的钱？’—— 我们并没有。” 卡尔文告诉我们。该法案的其他支持者及批评 OpenAI 营利化重组的人士也收到传票。 “他们针对这些人，基本上是恐吓他们闭嘴。” 詹姆斯・欧文基金会负责人唐・霍华德说。 （OpenAI 称这是标准法律程序的一部分。）

奥特曼长期支持民主党。 “我非常警惕强大的独裁者用恐惧故事抱团欺压弱者。” 他告诉我们，“这是犹太人的立场，不是同性恋的立场。”2016 年，他支持希拉里・克林顿，称特朗普 “对美国构成前所未有的威胁”。 2020 年，他向民主党与拜登胜利基金捐款。拜登政府期间，奥特曼至少六次在白宫会面。他协助制定一项冗长行政命令，确立首个联邦人工智能安全测试与其他监管框架。拜登签署该命令时，奥特曼称其 “良好开端”。

2024 年，随着拜登民调支持率下滑，奥特曼的论调开始转变。他称：“我相信无论本次大选结果如何，美国都不会有事。” 特朗普胜选后，奥特曼向其就职基金捐赠了 100 万美元，随后又在就职典礼上与网红杰克・保罗、洛根・保罗自拍合影。他在 X 平台以一贯的小写体风格发文：“近期更仔细地观察 @美国总统，着实改变了我对他的看法（真希望当初我能多独立思考……）。” 特朗普重返白宫首日，便撤销了拜登颁布的人工智能行政令。一名拜登政府高级官员如此评价奥特曼：“他找到了让特朗普政府为自己办事的有效手段。”

马斯克仍在公开场合猛烈抨击奥特曼，称其为 “骗子奥特曼”“狡诈的萨姆”。 （奥特曼曾在 X 上抱怨自己订购的一辆特斯拉汽车相关问题，马斯克回复：“你偷走了一家非营利机构。”）但在华盛顿政坛，奥特曼似乎已然在博弈中胜过了马斯克。马斯克为助特朗普连任花费超 2.5 亿美元，还在白宫供职数月，最终却离开华盛顿，期间也损害了与特朗普的关系。

如今，奥特曼已是特朗普青睐的商界大亨之一，甚至陪同特朗普前往温莎城堡拜访英国王室。奥特曼与特朗普每年都会通话数次。 “你随时都能给他打电话，” 奥特曼表示，“我们算不上密友，但没错，若我有事需要和他沟通，就会打过去。” 去年特朗普在白宫设宴款待科技界领袖时，马斯克赫然缺席，奥特曼则坐在总统对面。特朗普说道：“萨姆，你是一位举足轻重的领袖。你之前跟我说的一些事，简直令人难以置信。”

多年来，奥特曼始终将研发通用人工智能的征程比作曼哈顿计划。罗伯特・奥本海默曾以拯救世界免遭纳粹毒手的激昂说辞，说服物理学家们放下原有生活前往洛斯阿拉莫斯；奥特曼则借助外界对其技术地缘战略风险的担忧，根据不同受众，用这一类比分别鼓吹加速研发或谨慎推进。 2017 年夏，在与美国情报官员的一场会议中，他声称中国已启动 “通用人工智能曼哈顿计划”，OpenAI 需要数十亿美元政府资金才能跟上进度。当被要求拿出证据时，奥特曼只说：“我有所耳闻。” 这是他多次在会议中提出该说法的首次。其中一次会议结束后，他告知一名情报官员会补充证据，却始终未兑现。该官员调查所谓中国项目后得出结论：没有任何证据表明该项目存在，“这不过是他的融资说辞罢了。”（奥特曼称，自己不记得曾如此描述中方的相关行动。）

面对更注重安全的受众，奥特曼则用这一类比传递相反观点：通用人工智能必须在国际协作下谨慎研发，否则后果不堪设想。 2017 年，阿莫代伊聘请前公益律师佩奇・赫德利担任 OpenAI 的政策与伦理顾问。在早期向高管展示的一份 PPT 中，赫德利阐述了 OpenAI 如何避免一场 “灾难性” 军备竞赛 —— 或许可以组建人工智能实验室联盟，最终与类似北约的国际机构协作，确保技术安全落地。据赫德利回忆，布罗克曼始终不理解这一思路如何帮助公司击败竞争对手。 “无论我怎么解释，” 赫德利对我们说，“格雷格总绕回‘那我们怎么筹更多钱？怎么赢？’” 据多份采访记录与同期文件显示，布罗克曼提出了一个反方案：OpenAI 可以挑拨包括中俄在内的世界大国相互制衡，甚至挑起各国竞价，从中牟利。赫德利称，当时的核心逻辑似乎是：“核武器这么玩都行，人工智能为什么不行？”

赫德利对此感到震惊：“他们并未反驳这个前提 ——‘我们讨论的可能是人类有史以来最具破坏性的技术，要是把它卖给普京会怎样？’”（布罗克曼坚称，自己从未认真考虑过向各国政府拍卖人工智能模型。OpenAI 一名发言人表示：“当时只是初步探讨了可能促成国际合作的框架，类似人工智能领域的国际空间站。将其歪曲解读为其他意图，简直荒谬至极。”）

头脑风暴常会催生荒诞想法。赫德利希望这个内部被称作 “国家计划” 的方案能被搁置。但据多名参与人员与同期文件显示，OpenAI 高管们反而对此愈发热衷。时任 OpenAI 政策总监杰克・克拉克称，布罗克曼的目标是 “基本构建一个囚徒困境，让所有国家都不得不给我们提供资金”，而 “这无形中让不资助我们的行为变得有些危险”。一名初级研究员回忆，在公司会议上听到该方案细节时，他心想：“这简直疯透了。”

高管们曾至少向一位潜在捐赠者探讨过该方案，但当月晚些时候，多名员工扬言辞职，该计划最终被放弃。赫德利称：“萨姆会留不住员工。在他的权衡里，这一点似乎永远比‘这个计划会引发大国战争，不可行’更有分量。”

“国家计划” 的失败并未让奥特曼气馁，他转而推进类似变体方案。 2018 年 1 月，他在贝尔艾尔酒店召集了一场 “通用人工智能周末峰会”。这家老牌好莱坞度假酒店内，粉色三角梅缀满起伏的花园，人工池塘里游着真正的天鹅。参会者包括当时在牛津大学、被奉为人工智能末日论先知的哲学家尼克・博斯特罗姆，阿联酋人工智能事务部长、人工智能支持者奥马尔・奥拉马，以及至少七位亿万富翁。注重安全的参会者被告知，此次峰会将探讨社会如何应对通用人工智能颠覆性到来的问题；而投资者们则以为会听到融资路演。

峰会数日里，众人在一间雅致的会议室里发表演讲。 （领英联合创始人霍夫曼阐述了为人工智能注入佛教慈悲理念的可能性。）最后一位演讲者是奥特曼，他带着一份演示文稿，提出打造一种全球加密货币，“可兑换通用人工智能的服务时长”。一旦通用人工智能实现极致实用且 “反邪恶”，全球民众都会争相购买 OpenAI 服务器的使用时间。阿莫代伊在笔记中写道：“这个想法表面看就荒诞不经（弗拉基米尔・普京会不会也持有一些代币？……）现在回想起来，这是关于萨姆的诸多危险信号之一，我当初本该更重视的。” 该计划看似是敛财手段，奥特曼却将其包装为助力人工智能安全的举措。他的一张幻灯片写道：“我希望尽可能多的人加入‘正义’阵营，赢得胜利，做正确的事。” 另一张则写着：“请忍到演示结束再笑。”

多年来，奥特曼的融资说辞不断演变，但始终围绕一个核心事实：研发通用人工智能需要海量资金。他遵循着一条相当简单的 “缩放定律”：训练模型所用的数据与算力越多，模型似乎就越智能。实现这一过程的专用芯片造价极高。仅 OpenAI 最近一轮融资，就募集了超 1200 亿美元，创下史上最大规模私募融资纪录，金额更是史上最大 IPO 的四倍。一名科技行业高管兼投资人对我们表示：“放眼全球，每年能自主支配千亿美元资金的主体寥寥无几 —— 美国政府、美国四五家顶尖科技公司、沙特与阿联酋，基本就这些了。”

奥特曼最初瞄准的是沙特阿拉伯。 2016 年，他在旧金山费尔蒙特酒店的一场晚宴上，首次会见沙特王储、实际掌权者穆罕默德・本・萨勒曼。赫德利回忆，此后奥特曼便称这位王储为 “朋友”。 2018 年 9 月，据赫德利的笔记记载，奥特曼称：“我在考虑，我们到底要不要从沙特公共投资基金拿数百亿美元。”

次月，一支暗杀小队据称奉本・萨勒曼之命，勒死了批评沙特政权的《华盛顿邮报》记者贾迈勒・卡舒吉，并用骨锯肢解了其遗体。一周后，有消息宣布奥特曼加入本・萨勒曼计划在沙漠中打造的 “未来城市” 尼奥姆的顾问委员会。现任职于 Anthropic 的前政策总监克拉克回忆，他当时对奥特曼说：“萨姆，你不能加入这个委员会。” 奥特曼起初为自己辩解，称贾里德・库什纳向他保证沙特人 “没做这件事”。 （奥特曼否认有过此番言论，库什纳也称当时二人并无联系。）

随着本・萨勒曼涉案证据愈发确凿，奥特曼退出了尼奥姆顾问委员会。但据一名奥特曼咨询过的政策顾问回忆，私下里他将此事视为暂时挫折，还询问能否仍从本・萨勒曼处获得资金。 “他关心的不是‘这么做对不对’，” 这名顾问说，“而是‘这么做会有什么后果？会不会触发出口管制？会不会遭到制裁？我能不能侥幸过关？’”

彼时，奥特曼已盯上另一个资金来源：阿联酋。该国正推进一项为期 15 年的计划，试图从石油国家转型为科技枢纽，该项目由总统之弟、该国情报主管塔赫农・本・扎耶德・阿勒纳哈扬主持。塔赫农掌管着国有人工智能集团 G42，控制着 1.5 万亿美元的主权财富基金。 2023 年 6 月，奥特曼到访阿布扎比，与奥拉马等官员会面。在一场政府支持的活动上，他称阿联酋 “早在人工智能流行之前就开始探讨相关议题”，并勾勒出中东在未来人工智能领域占据 “核心地位” 的愿景。

向海湾国家融资已是众多大型企业的常规操作，但奥特曼追求的是更宏大的地缘战略构想。 2023 年秋，他开始秘密为一项计划招揽人才 —— 该计划最终名为 “芯片公司计划”，由海湾国家提供数百亿美元，兴建大型芯片制造厂与数据中心，部分设施将落户中东。他向现任 Meta 人工智能负责人亚历山大・王抛出高管职位邀约，称亚马逊创始人杰夫・贝索斯可能执掌这家新公司。奥特曼向阿联酋寻求巨额注资。一名董事会成员称：“据我所知，整件事董事会完全不知情。” 奥特曼曾试图招揽研究员詹姆斯・布拉德伯里参与该项目，后者回忆自己拒绝了邀约：“我的第一反应是‘这计划能成，但我不希望它成’。”

人工智能算力或许很快会取代石油或浓缩铀，成为决定全球势力格局的核心资源。奥特曼称，算力是 “未来的货币”。通常而言，数据中心选址无关紧要，但多名美国国家安全官员对将先进人工智能基础设施集中在海湾专制国家深感担忧。阿联酋电信基础设施高度依赖与政府关联的中国科技巨头华为的设备，且据报道该国曾向北京泄露过美国技术。情报机构担心，运往阿联酋的美国先进芯片可能被中国工程师利用。中东的数据中心也更容易遭受军事打击；近几周，伊朗已轰炸了巴林与阿联酋境内的美国数据中心。此外，理论上海湾君主国可强行接管美资数据中心，用其打造能力远超常规的模型 —— 这并非虚构的 “通用人工智能独裁” 场景，而是可能在现实独裁政权中发生的事。

奥特曼被解雇后，最依赖的人便是爱彼迎联合创始人切斯基，他也是奥特曼最坚定的支持者。切斯基对我们说：“看着朋友陷入如此绝境，让我对真正运营一家公司的意义产生了一些根本性的质疑。” 次年，在一场 Y Combinator 校友聚会上，他发表了一场即兴演讲，时长长达两小时。 “感觉就像一场团体心理治疗，” 他说。演讲的核心观点是：你运营自己创立公司的直觉，就是最好的直觉，任何反驳你的人都是在精神操控你。切斯基称：“即便下属说你疯了，你也没疯。” 保罗・格雷厄姆在一篇关于这场演讲的博客文章中，将这种 defiant 态度命名为 “创始人模式”。

自那次风波后，奥特曼便进入了 “创始人模式”。 2024 年 2 月，《华尔街日报》披露了奥特曼对 “芯片公司计划” 的构想：他计划打造一家合资企业，融资 5 万亿至 7 万亿美元。 （他在推特发文：“管他呢，干脆 8 万亿得了。”）很多员工都是通过这则报道才得知该计划。莱克回忆：“所有人都一脸懵，‘等等，什么情况？’” 奥特曼在内部会议上坚称，安全团队已 “参与知情”。莱克随即发消息敦促他，不要谎称该项目已获安全团队认可。

拜登政府时期，奥特曼曾申请安全许可，希望参与机密人工智能政策讨论。但协助协调该流程的兰德公司工作人员对此表示担忧。其中一人写道：“他一直在向外国政府募集‘数千亿美元’，阿联酋最近还送了他一辆车。（我猜是辆豪车。）” 该工作人员还称：“我能想到的、有如此深厚海外资金关联还申请安全许可的人，只有贾里德・库什纳，而审核人员当时建议不给他批许可。” 奥特曼最终退出了申请流程。一名参与与奥特曼会谈的政府高级官员称：“他极力推进这种利益交换式合作，主要是和阿联酋，这在我们很多人看来都是危险信号。政府里不少人对他并非完全信任。”

当我们问及奥特曼有关塔赫农赠送礼物的事时，他称：“我不会透露他具体送了我什么礼物，但他和其他多国领导人…… 确实送过我东西。” 他补充道：“公司有标准政策，我也必须遵守，所有潜在商业伙伴赠送的礼物都要向公司报备。” 奥特曼至少拥有两辆顶级超跑：一辆价值约 200 万美元的全白科尼赛克 Regera，以及一辆价值约 2000 万美元的红色迈凯伦 F1。2024 年，有人拍到他驾驶 Regera 穿行于纳帕谷。一段几秒的视频流传到社交媒体：奥特曼坐在低矮的桶型座椅上，从这辆亮白色跑车的车窗向外望去。一名与马斯克立场一致的科技投资人在 X 平台发布了这段视频，配文：“我下次也要开个非营利机构。”

2024 年，奥特曼带着两名 OpenAI 员工，登上了塔赫农价值 2.5 亿美元的超级游艇 “玛利亚号”。这艘全球顶级游艇配有直升机停机坪、夜总会、影院和海滩俱乐部。奥特曼的员工在塔赫农的武装安保人员中显得格格不入，至少有一名员工后来告诉同事，这段经历让他感到不安。奥特曼随后在 X 平台称塔赫农是 “亲爱的私人朋友”。

奥特曼仍持续与拜登政府会面，而拜登政府已出台政策，要求敏感技术出口需获白宫批准。多名政府官员在会面后，对奥特曼在中东的野心感到担忧。据这些官员称，他时常发表豪言壮语，甚至称人工智能是 “新电力”。 2018 年，他称 OpenAI 计划从 Rigetti Computing 公司购买一台完整可用的量子计算机，这一消息连在场的其他 OpenAI 高管都毫不知情。彼时 Rigetti 远未具备出售可用量子计算机的能力。在另一场会议中，奥特曼宣称到 2026 年，美国境内将建成广泛的核聚变反应堆网络，为人工智能热潮提供动力。那名政府高级官员称：“我们当时的反应是，‘这要是真实现了核聚变，可真是大新闻。’” 拜登政府最终未批准相关申请。美国商务部一名高层对奥特曼表示：“我们不会在阿联酋建造先进芯片。”

《华尔街日报》报道，特朗普就职典礼前四天，塔赫农向特朗普家族支付 5 亿美元，换取其加密货币公司的股份。次日，奥特曼与特朗普进行了 25 分钟通话，商议宣布一项 “芯片公司计划” 的变体项目，时机恰好能让特朗普揽下功劳。特朗普上任次日，奥特曼在罗斯福厅宣布成立 “星门” 合资企业，项目规模达 5000 亿美元，旨在全美打造庞大的人工智能基础设施网络。

同年 5 月，特朗普政府撤销了拜登时期的人工智能技术出口限制。奥特曼与特朗普一同前往沙特王宫会见本・萨勒曼。与此同时，沙特宣布成立一家大型国有人工智能企业，将斥资数十亿美元开展国际合作。约一周后，奥特曼公布了 “星门” 拓展至阿联酋的计划，拟在阿布扎比打造一座数据中心园区，面积是纽约中央公园的七倍，耗电量大致相当于迈阿密全市。一名前 OpenAI 高管称：“说白了，我们正在建造召唤‘外星力量’的入口。如今这类入口存在于美国和中国，而萨姆又在中东加了一个。” 他接着说：“我觉得必须认清这件事有多可怕。这是有史以来最鲁莽的行为。”

安全承诺的弱化已成为行业常态。 Anthropic 成立的初衷，便是凭借合理的架构与领导层，坚守安全承诺，不被商业压力侵蚀。其中一项承诺是 “负责任缩放政策”，即若无法证明模型安全，便停止训练更强大的版本。 2025 年 2 月，该公司获得 300 亿美元新融资之际，却弱化了这一承诺。在某些方面，Anthropic 对安全的重视程度仍高于 OpenAI。但前政策总监克拉克称：“资本市场的规则就是‘加速推进’。” 他补充道：“最终做决定的是整个世界，而非企业。” 去年，阿莫代伊向 Anthropic 员工发送备忘录，披露公司将寻求阿联酋与卡塔尔的投资，并坦言这可能会让 “独裁者” 获利。 （与许多作者一样，我们二人也卷入了一场集体诉讼，指控 Anthropic 未经许可使用我们的著作训练模型。康泰纳仕集团已就 Anthropic 使用其旗下出版书籍一事，与该公司达成和解协议。）

2024 年，Anthropic 与硅谷立场最强硬的国防承包商之一帕兰提尔达成合作，将其人工智能模型 Claude 直接接入军事系统，成为五角大楼最高机密场景中唯一使用的人工智能承包商。去年，五角大楼再向该公司授予一份 2 亿美元合同。 2025 年 1 月，美军发动午夜突袭，抓获委内瑞拉总统尼古拉斯・马杜罗。据《华尔街日报》报道，Claude 参与了此次机密行动。

但 Anthropic 与美国政府间随即产生矛盾。多年前，OpenAI 已从政策中删除了 “禁止将技术用于军事与战争” 的全面禁令。最终，包括谷歌与 xAI 在内的 Anthropic 竞争对手，均同意向军方提供模型，用于 “所有合法用途”。而 Anthropic 的政策禁止研发全自动武器与国内大规模监控，在这两点上拒不妥协，导致新协议谈判陷入停滞。 2 月下旬的一个周二，国防部长皮特・赫格塞思召阿莫代伊前往五角大楼，下达最后通牒：公司必须在周五下午 5 点 01 分前放弃相关禁令。截止日前一天，阿莫代伊拒绝妥协。赫格塞思在推特发文，将 Anthropic 列为 “供应链风险企业”—— 这一严厉黑名单历来只针对华为等与外国对手有关联的企业，数日后他便兑现了这一威胁。

OpenAI 与谷歌数百名员工联名发表题为《我们不会分裂》的公开信，声援 Anthropic。奥特曼在内部备忘录中称，这场争端 “关乎整个行业”，并声称 OpenAI 与 Anthropic 秉持相同的道德底线。但事实上，至少两天前奥特曼已与五角大楼展开谈判。负责研究与工程的国防部副部长埃米尔・迈克尔为寻找 Anthropic 的替代方案，联系了奥特曼。 “我得赶紧找替补，” 迈克尔回忆，“我给萨姆打了电话，他立刻就答应了。我认为他是个爱国者。” 奥特曼问迈克尔：“我能为国家做些什么？” 他似乎早已心知肚明。 OpenAI 原本不具备接入 Anthropic 所服务的机密系统所需的安全认证，但周五上午宣布的一笔 500 亿美元交易，将 OpenAI 技术整合进五角大楼数字基础设施核心的亚马逊云服务。当晚，奥特曼在 X 平台宣布，美军将正式使用 OpenAI 的模型。

从某些指标来看，奥特曼的这一操作并未阻碍公司发展。宣布合作当日，新一轮融资让 OpenAI 估值暴涨 1100 亿美元。但大量用户卸载了 ChatGPT 应用，至少两名高管离职，其中一人加入了 Anthropic。在员工大会上，奥特曼斥责提出质疑的员工：“或许你们觉得打击伊朗是对的，入侵委内瑞拉是错的，但你们没资格评判这些。”

多名与 OpenAI 关联的高管仍对奥特曼的领导能力持保留意见，提议由前英斯达首席执行官、现任 OpenAI 通用人工智能部署首席执行官菲姬・西莫接任。一名知悉近期讨论的人士称，西莫私下曾表示，认为奥特曼最终可能辞职。 （西莫否认了这一说法。英斯达近期与联邦贸易委员会达成和解，未承认不当行为，但就西莫任职期间涉嫌欺诈行为支付 6000 万美元罚款。）

奥特曼将自己立场的反复转变，归因于适应环境变化的能力 —— 并非马斯克等人指控的蓄谋已久的 “长线骗局”，而是出于善意的逐步调整。他对我们说：“有些人想要的领导者，是对自己的想法绝对笃定、始终如一、绝不改变的人。但我们所处的领域，变化速度极快。” 他为自己的部分行为辩解，称这是 “正常的商业竞争手段”。我们采访的多名投资人认为，奥特曼的批评者太过天真，本就不该抱有其他期待。投资人康威称：“有一群宿命论极端分子，把人工智能安全捧到了科幻小说的高度。他的使命用数据衡量即可，而 OpenAI 的成功，有目共睹，数据不会骗人。”

但硅谷另一些人认为，奥特曼的行为造成了难以接受的管理失序。一名董事会成员称：“问题更多在于，公司实际已陷入治理失灵。” 还有人认为，人工智能的缔造者应接受比其他行业高管更严格的评判。我们采访的绝大多数人都认同，奥特曼如今要求外界用以评判自己的标准，已背离其最初的承诺。在一次对话中，我们问奥特曼，运营人工智能公司是否需要 “更高的诚信标准”。这本该是一个简单的问题 —— 此前被问及类似问题时，他都会给出明确、毫无保留的肯定答案。而这次他却称：“我认为有很多企业都可能对社会产生巨大的正面或负面影响。”（随后他补充声明：“是的，这要求更高水平的诚信，我每天都深感责任重大。”）

OpenAI 成立之初许下的所有承诺中，最核心的无疑是安全研发人工智能。但如今，这类担忧在硅谷与华盛顿却常遭嘲讽。去年，曾任风投资本家、现任副总统的 JD・万斯在巴黎一场名为 “人工智能行动峰会”（原 “人工智能安全峰会”）的会议上称：“人工智能的未来，不会靠对安全问题的杞人忧天赢得。” 今年达沃斯论坛上，担任白宫人工智能与加密货币主管的风投资本家戴维・萨克斯称，对安全的担忧是 “自我设限”，会让美国输掉人工智能竞赛。奥特曼如今称特朗普的放松监管举措 “是令人耳目一新的改变”。

OpenAI 已解散多个专注安全的团队。超级对齐团队解散前后，团队负责人苏茨克维与莱克双双辞职。 （苏茨克维联合创立了安全超智能公司。）莱克在 X 平台发文：“安全文化与流程已让位于光鲜的产品。” 不久后，负责为社会应对先进人工智能冲击做准备的通用人工智能就绪团队也被解散。在最新提交给美国国税局的披露文件中，被问及 “最重要业务” 时，此前文件中提及的安全相关内容已不见踪影。 （OpenAI 称其 “使命未变”，并补充：“我们持续投入并优化安全工作，也会不断进行组织架构调整。”）生命未来研究所是一家智库，奥特曼曾认可其安全准则，该机构会对各大人工智能公司的 “生存安全” 评级；在最新报告中，OpenAI 获评 F 级。公平来讲，除 Anthropic 获 D 级、谷歌深度思维获 D - 级外，其他主流公司均为 F 级。

奥特曼称：“我对传统人工智能安全那套理念并不认同。” 他坚称自己仍将安全置于优先位置，但被问及具体举措时却含糊其辞：“我们仍会开展安全项目，或至少是与安全相关的项目。” 当我们提出采访公司内研究生存安全问题的研究员 —— 这类问题正如奥特曼曾所言，可能关乎 “人类文明覆灭”——OpenAI 发言人一脸困惑：“你说的‘生存安全’是什么意思？这根本不算一个研究方向。”

人工智能末日论者已被边缘化，但他们的部分担忧却逐月变得愈发现实。联合国一份报告显示，2020 年，一架人工智能无人机在利比亚内战中发射致命弹药，疑似无人操控自主执行任务。此后，人工智能在全球军事行动中的地位愈发重要，据报道，当前美国对伊朗的军事行动也大量运用人工智能。 2022 年，一家制药公司的研究员测试药物研发模型能否用于寻找新型毒素，短短数小时内，模型便给出了四万种致命化学战剂的方案。而更多日常危害已然显现：我们愈发依赖人工智能辅助写作、思考、认知世界，加速了专家所称的 “人类能力退化”；人工智能生成的劣质内容泛滥，让骗子有机可乘，也让普通人难以辨别真伪。人工智能 “智能体” 已开始在极少甚至无人监管的情况下自主行动。 2024 年新罕布什尔州民主党初选前数日，数千名选民收到人工智能伪造拜登声音的自动语音电话，谎称让他们留到 11 月再投票、无需前往投票站 —— 这场选民压制行动几乎无需任何专业技术。 OpenAI 目前面临七起非正常死亡诉讼，指控 ChatGPT 诱导多起自杀事件与一起谋杀案。那起谋杀案的聊天记录显示，ChatGPT 助长了一名男子的偏执妄想，认为其 83 岁的母亲在监视并试图毒害自己。随后，该男子殴打并勒死母亲，随后自残。 （OpenAI 正应诉，并称持续优化模型的安全防护机制。）

随着 OpenAI 筹备潜在的 IPO，奥特曼不仅面临人工智能对经济影响的质疑 —— 该技术或很快引发严重就业冲击，导致数百万岗位消失 —— 还被问及公司自身财务状况。创业治理专家埃里克・莱斯抨击行业内的 “循环交易”，例如 OpenAI 与英伟达等芯片制造商的合作，并称在其他时代，该公司部分会计操作会被视为 “近乎欺诈”。一名董事会成员对我们称：“公司目前的财务杠杆模式风险极高，令人担忧。”（OpenAI 对此予以否认。）

2025 年 2 月，我们再次采访奥特曼。他身着暗绿色毛衣与牛仔裤，坐在一张 NASA 月球车照片前。他先是将一条腿盘在身下，随后又搭在椅臂上。他称，过去自己作为管理者的主要缺点是总想避免冲突。 “现在我很乐意果断开除员工，” 他对我们说，“也敢直接拍板‘我们就往这个方向押注’。” 不认同其决策的员工，只能 “离开”。

他对未来比以往任何时候都更乐观。 “我眼中的成功，是人类实现跨越式提升，所有人都迎来科幻般的疯狂未来，” 他说，“我对人类的前景、对我们能实现的成就，抱有极大的野心。但奇怪的是，我个人几乎没什么野心。” 有时他也会自我警醒：“没人会相信你做这些只是因为感兴趣，大家都会觉得你是为了权力或其他东西。”

即便与奥特曼关系密切的人，也难以分清他的 “人类愿景” 与个人野心的边界。他最大的优势，始终是能说服不同群体，让他们相信自己想要的，正是对方所需的。他抓住了一个特殊的历史节点：公众对科技行业的炒作心存警惕，而绝大多数能研发通用人工智能的研究员，又对其落地充满恐惧。奥特曼使出了其他融资者未曾精通的一招：用末日论调阐述通用人工智能如何毁灭人类，进而论证自己才是研发该技术的合适人选。这或许是蓄谋已久的高招，或许只是他为抢占优势的临时之举。无论如何，他成功了。

导致聊天机器人产生危险的特性，并非全是程序漏洞；部分是系统构建过程中的附带产物。大语言模型的训练部分依赖人类反馈，而人类往往偏爱顺从的回应。模型常会学会奉承用户，这种倾向被称为 “谄媚性”，有时甚至会将此置于诚实之上。模型还会虚构内容，即 “幻觉”。各大人工智能实验室均已记录到这些问题，却时常选择容忍。随着模型愈发复杂，部分模型的虚构内容甚至更具说服力。 2023 年，也就是被解雇前不久，奥特曼曾提出，即便存在风险，适度允许虚假信息也能带来优势。他称：“如果简单粗暴地要求模型‘只说百分百确定的事’，确实能做到，但模型就会失去人们喜爱的那种‘魔力’。”