Part IThe Story
The Second Source
The question was asked with the flatness lawyers use when they already know the answer and want the jury to hear it in someone else's voice. "As you understand it, was Mr. Altman telling the truth when he made that statement to you?"
"No."
The answer was one syllable. It was played on a screen in a federal courtroom in Oakland on Wednesday, May 6, 2026, during the second week of Musk v. Altman, the lawsuit in which Elon Musk sought up to $180 billion and an order forcing OpenAI to unwind its conversion from a nonprofit into something closer to an ordinary company. The witness was not in the room. She was on tape, as she had been, in one way or another, for most of the previous three years: on panels, at developer conferences, on magazine covers. Two days earlier Mira Murati had been photographed on the steps of the Metropolitan Museum of Art at the Met Gala, in a gown the New York Post called "glam." Now she was a deposition clip, explaining in the measured, faintly accented, paragraph-length sentences that are her native register why she had stopped believing the man she once served as chief technology officer.
The statement at issue was narrow and procedural, which is what made it damning. According to her testimony, Altman told her that OpenAI's legal department had determined that a new model did not need to go through the company's deployment safety board. The model has been reported to be GPT-4 Turbo. Murati did not argue with Altman, and she did not go to the board. She did something smaller and more characteristic. She asked a second person.
The second person was Jason Kwon, who had joined OpenAI in 2021 as its general counsel and who would go on to become its chief strategy officer. A lawyer who became a strategist, which is to say a man whose whole career has been spent in the space between what is said and what is binding. Murati put the same question to Kwon that she had put to Altman. "I confirmed that what Jason was saying and what Sam was saying were not the same thing," she testified. There was, in her word, a "misalignment." So she sent the model through the safety board anyway, quietly, without a memo to the world, and moved on to the next problem on a list that was always long.
Silicon Valley has a mythology of the founder's monologue, and in that mythology, the CEO who says the true thing loudest wins. Murati's characteristic act is different. It is the follow-up call. It is the checking of one account against another, the refusal to let a single source stand, and then the absence of any announcement at all. "My concern," she said in the same testimony, "was about Sam saying one thing to one person and completely the opposite to another person." She was careful, almost fussy, about the scope of her complaint. It was "completely management related." She had, she said, "an incredibly hard job to do in an organization that was very complex."
In 2023, before any of this was public, Wired's Steven Levy wrote that you could call her OpenAI's "minister of truth." He meant the phrase as a description of a job: she ran the teams responsible for making sure ChatGPT and DALL-E did not mislead people, show bias, or, in Levy's words, "snuff out humanity altogether." He could not have known how literally the title would eventually be tested, or that the system whose candor she would end up having to audit most carefully was not the model.
By the Numbers
Mira Murati, 1988–present
16Her age when she left Vlorë, Albania, on a United World Colleges scholarship
100MChatGPT monthly active users within two months of its November 2022 launch
3 daysHer tenure as OpenAI's interim CEO in November 2023
$2BThinking Machines Lab's seed round, the largest in history
$12BValuation at that seed, five months after launch
100×Voting weight of founding shares versus regular shares
975BParameters in Inkling, the lab's first open-weights model
Mathematics as a Place to Stand
She was born Ermira Murati on December 16, 1988, in Vlorë, a port city on Albania's southern coast, in the last years of a regime so sealed off from the world that she would later reach for the only comparison an American audience could grasp. "I was born in Albania, just after the fall of communism," she told Andreessen Horowitz's Martin Casado in 2023. "Very interesting times in this very isolated country. It's similar to North Korea today."
The transition that followed was sudden and, as Fortune put it with some understatement, "somewhat chaotic." Both of her parents, according to Fortune, were high school teachers who taught literature. That detail sits strangely next to everything that comes after it, because the daughter of two literature teachers grew up to say that the humanities, in the country she came from, could not be trusted.
— Mira Murati, in conversation with Martin Casado, a16z, September 2023The humanities, like history and sociology, were a bit questionable. The source of information and truthfulness was hard. It was ambiguous, so I got very interested in math and sciences. That's what I pursued relentlessly.
It is worth sitting with that sentence. A child grows up in a place where the official history has just been revealed as a fiction and the replacement history has not yet been written, where the books her parents assigned may have been revised by the state and the state itself is dissolving. And she concludes, reasonably, that the safest ground is the ground where a proof is a proof regardless of who is in power. Mathematics as a place to stand. She told Microsoft's chief technology officer, Kevin Scott, that she had an "organic interest towards math and science," and she competed in Olympiads throughout her schooling. Later she would say, with a certain flatness, "I'm still working fundamentally in mathematics." She said it as CTO of a company whose product was, of all things, language.
She credited the old regime with exactly one thing. When everything else was equal (and in that system, materially, everything else was equal), there was, she said, an intense competition for knowledge. Fortune reported that she credited boredom, too. The internet, when it arrived, was slow. She tinkered with computers anyway, and wondered, among other things, how the human brain works. Nobody remembers a teenager's questions, but that one would turn out to have a long half-life.
At sixteen she left. A scholarship from the Davis United World College Scholars Program took her to Pearson College UWC on Vancouver Island, one of a network of schools that gathers students from more than eighty countries on the premise that teenagers who share dormitories across borders will grow into adults less likely to go to war with one another. Her parents stayed in Albania. It is the kind of decision that families in small countries make constantly and that biographies tend to describe as a single brave moment. It is not a moment, though. It is a permanent condition. You leave at sixteen, and you are never again fully from the place you were born or fully from the place you went to, and you become very good at reading rooms, because every room is a room you were not raised in.
The phrase she would use a decade later, when she tried to explain what kind of company she wanted to build, was "the messy way we collaborate." Very few people raised in a country that was rewriting its own history in real time come away believing that collaboration is anything but messy.
Everything That Breaks
What followed was an education that refused to choose. She enrolled in a dual-degree program that gave her a Bachelor of Arts from Colby College in Maine in 2011 (Fortune reports the degree was in mathematics) and a Bachelor of Engineering in mechanical engineering from Dartmouth's Thayer School of Engineering in 2012. Dartmouth, El País would later note, is the college where the discipline of artificial intelligence was born. She did not go there for AI. She went there, by the account of Thayer's dean, Alexis Abramson, to build race cars in Allyn Lab with the Dartmouth Formula Racing team.
Race cars are not a metaphor. They are a discipline in failure modes. You learn which welds crack and which suspension geometries fold and how a system that works perfectly on a bench behaves at speed on a wet track. Everything she did for the next six years had that character. Hard things, physical things, things that break when you are careless.
The résumé reads like someone trying on industries the way other people try on cities. A summer analyst internship at Goldman Sachs in Tokyo in 2011. A stint at Zodiac Aerospace, the French aerospace group, as an advanced concepts engineer, the kind of job where ideas meet physics and cost and usually lose. Then, in 2013, Tesla.
At Tesla she joined first to work on the Model S dual motor and then moved to the Model X "from the early days of the initial design," as she told Casado, "and eventually led the whole program to launch." The Model X was the car with falcon-wing doors, a vehicle that was famously difficult to manufacture, a product in which every elegant idea had to survive contact with a supply chain. She called the experience "a very formative moment—going through the whole experience of design and deployment of a whole vehicle." The word to notice is deployment. It would become, at OpenAI, something close to a creed.
During her three years at Tesla, the company released early versions of Autopilot, its AI-assisted driving software. Murati got interested, she said, "in applications of AI, specifically with autopilot," and then, more abstractly, in the question of what happens "when you're using AI and computer vision in a different domain."
So she went to a different domain. In 2016 she joined Leap Motion, a startup building hand- and finger-tracking sensors, as vice president of product and engineering. She wanted, she told Fast Company, to make the experience of interacting with a computer "as intuitive as playing with a ball." It is a lovely sentence. It was also, she concluded, premature. The technology depended on a virtual reality headset, and the world was not ready to strap one on every day. "Obviously, in retrospect, it was too early back then," she said. "But I learned a lot about the limitations of pushing this technology to the practicality of using it every day."
Cars, aircraft components, spatial computing: a pattern emerges only in reverse. Each job was a single modality, a single domain where intelligence had to be engineered by hand for one narrow purpose. And at some point she stopped asking how to make AI competent in one domain after another and started asking the opposite question. "At this point, I started thinking more about what happens if you just focus on the generality," she recalled. "Forget the competence in different domains and just focus on generality."
There were, in her telling, two places in 2018 "laser-focused on this issue": OpenAI and DeepMind. She chose the one with the mission statement.
Making the Investor Believe
She joined OpenAI in June 2018 as vice president of applied AI and partnerships. The company was then a nonprofit research lab, a place that published papers and trained models that played video games, and her first work there was, by her own account, unglamorous: "I began working on our supercomputing strategy and managing a couple of research teams." The person who had launched a car was now in charge of thinking about the physical machines that would make thinking machines possible. Hardware first, again.
"I very quickly believed that AGI would be the last and most important major technology that we built," she told Levy, "and I wanted to be at the heart of it." To Fortune she was more tentative, and more honest: "I wasn't sure it was going to happen at that point, but I knew that even if we just got very close, the things we would build along the way would be incredible."
What she built along the way, in those first years, was mostly relationships. Her title had "partnerships" in it for a reason. OpenAI in this period reorganized so that a capped-profit company lived inside the nonprofit's structure, which was a polite way of saying the mission had become too expensive for donors. "To really understand how to make our models better and safer, you need to deploy them at scale," she explained to Levy. "That costs a lot of money. It requires you to have a business plan, because your generous nonprofit donors aren't going to give billions like investors would."
The investor was Microsoft. Murati helped oversee the partnership, which eventually brought OpenAI a $10 billion investment and, as Fortune put it, "all sorts of new possibilities for distribution." Years later, introducing her onstage in December 2024, Levy volunteered that while reporting a story on Microsoft he had heard from "any number of people," unprompted, how important she had been to the partnership and how much they enjoyed working with her. This is not the kind of thing people say about the executive who wins the negotiation. It is what they say about the one who makes the negotiation bearable.
The most revealing exchange of her OpenAI years, though, is a small one in the Levy interview from July 2023, about whether Microsoft's mission was aligned with OpenAI's.
"In the sense that they believe that this is our mission," she said.
"But that's not their mission," Levy pressed.
"No, that's not their mission. But it was important for the investor to actually believe that it's our mission."
Read it twice. It is a perfectly candid description of how a mission-driven organization survives contact with capital: you cannot make the investor share your goal, but you can make the investor believe that you hold it, and that belief becomes a kind of contract. It is also, in retrospect, a sentence about the precise difference between what people say and what they mean, spoken by someone who would build her later reputation on noticing that difference.
In 2020 she became senior vice president of research, product, and partnerships, an unusually wide scope at a company that still thought of itself primarily as a lab. Then came the thing she remembers. "I do remember GPT-3 being able to translate," she told Levy. "I speak Italian, Albanian, and English. I remember just creating pair prompts of English and Italian. And all of a sudden, even though we never trained it to translate in Italian, it could do it fairly well."
An emigrant, testing the machine in a language that was not her first and not her current one, and watching it cross a border nobody had shown it how to cross. Among all the milestones of those years, that is the one she chose.
In May 2022 she became chief technology officer. Six months later, almost by accident, she launched the fastest-growing consumer application in history.
Fifty-Five Days
ChatGPT was not supposed to be a product. By Murati's own account to Casado, the conversational interface was originally conceived as a research tool, a way to gather feedback for aligning and safety-tuning what would become GPT-4. The technique underneath was reinforcement learning from human feedback, the method by which a model's outputs are ranked by people and the model is nudged toward the rankings. OpenAI had already exposed similar capabilities through its API, to modest effect among developers. The difference, she observed later with some wonder, was the interface. Same technology, wildly different adoption. A chat window turned out to be the ball you could just pick up and play with.
It launched in late November 2022 and reached more than 100 million monthly active users in two months. OpenAI's revenue in 2022, according to documents that surfaced during a fundraising round, had been just under $30 million. The following year it was on track for more than $1 billion.
There is a scene, reported by Fast Company, from a Monday morning at the end of January 2023, fifty-five days after the launch. OpenAI's headquarters was then a former factory in San Francisco housing 375 researchers, engineers, and policy specialists. Past an imposing gate, through a courtyard edged with palm trees, down a hall from a pebble fountain that made what the reporter called "a spalike burble," the six-person leadership team behind ChatGPT assembled in a snug conference room for its weekly meeting. That same morning Microsoft had announced it would pour a reported further $10 billion into the company. The largest school district in America had banned the product. Podcasters were comparing it to the iPhone.
Murati, her hair in what the magazine called a "haphazard ponytail," dressed in head-to-toe black, sipped green tea and turned to a fresh page of a yellow legal pad. The first item on the agenda was, to be precise about it, bullshit, the industry's polite word being "hallucination." When she spoke for the first time, it was to ask a question that sounds like jargon and is in fact one of the deepest problems in the field.
"Do we know the relationship between the factuality balance and the creativity evals?"
What she was asking was whether the machine's capacity to make things up, the thing that made it charming, was the same capacity that made it lie. Whether you could have imagination without fabrication. Whether, in a system that generates a fresh answer every time, truth and invention could be pulled apart at all. A woman who had fled the humanities because "the source of information and truthfulness was hard" was now responsible for the most widely used generator of plausible-sounding text in human history. She took notes on a legal pad.
— Mira Murati, to Fortune, October 2023I think the downside is a race to the bottom on safety. That is the downside for sure.
The months that followed made her a public figure, which by temperament she had never seemed to want to be. When Fortune came to interview her in 2023, Sam Altman darted out of the conference room just before she entered, gathered his belongings (a laptop, the magazine noted, underwhelmingly) and shuffled away, "clearing the path for Murati to take center stage." The image is almost too neat. He was the frontman, freshly returned from talks with Congress about regulation. She was the one who, in Fortune's phrase, "tinkers just outside of the spotlight," dressed that day in jeans and a T-shirt, answering questions "slowly and thoughtfully, with a tone that's both relaxed and ardent."
She told the magazine that she spent her time thinking not only about features but about whether AI would end humanity. She put the chance at "small." She described asking her teams, for each new product iteration, "how bad are the bad things." And she said something that would later sound less like philosophy than like prophecy: "What's particularly difficult is being able to predict the emerging capabilities and to get ahead of some of the deployment risks. Because at the end of the day, you need to institutionalize and operationalize these things, and they can't just be policies and ideas."
They can't just be policies and ideas. A safety board is only a safety board if the models actually go through it.
Her reputation in these years was for calm, for being, in the description of one research dossier, an "engineer-diplomat," allergic to both hype and doom. She described the current models, in September 2023, as roughly "intern level" in reliability. Capable of glimpses. Not yet trustworthy on their own. It is the most careful possible way to describe a product you are selling to a hundred million people, and it was, for a while, exactly the voice OpenAI needed.
In October 2023, Fortune ranked her 57th on its list of the 100 Most Powerful Women in Business and put her on the cover. In October 2023, according to later reporting in the New York Times, she also began telling members of OpenAI's board what she thought of the man who had left the room with his laptop.
Three Days in November
The board fired Sam Altman on Friday, November 17, 2023, with a statement that he "was not consistently candid in his communications with the board, hindering its ability to exercise its responsibilities." It named Murati interim chief executive. She was thirty-four.
What happened next has been told so many times that it has hardened into folk history: the employee revolt, the open letter, the Microsoft offer to hire everyone, the return of the king. Her role in it remains the most contested part of the story, and the most interesting, because she occupied every side of it at once.
Here is what has been reported. In October 2023, according to people with knowledge of the board's discussions cited by the New York Times, Murati approached some board members and described what some considered Altman's playbook: he would tell people what they wanted to hear to win their support, and if they did not go along, or took too long to decide, he would try to undermine the credibility of whoever challenged him. She told the board she had previously sent Altman a private memo outlining her concerns and shared some of its details. Around the same time, Ilya Sutskever, OpenAI's cofounder and chief scientist, made similar complaints. Some board members feared both executives would leave if nothing changed. Sutskever's 52-page memo to the board, which would later be read aloud in a deposition, said Altman "exhibits a consistent pattern of lying, undermining his execs, and pitting his execs against one another." According to The Information, that memo relied heavily on screenshots and information Murati had provided.
Ilya Sutskever is the other figure in this story whose biography rhymes with hers. He was born in the Soviet Union, emigrated as a child, became a researcher of almost mystical intensity, and was perhaps the purest believer in the mission at a company that was rapidly becoming a business. He supported the firing, then signed the letter demanding its reversal, then effectively vanished from his own company before leaving it in May 2024. Two immigrants from collapsed communist states, both of them technical, both of them convinced the founder could not be trusted. Both reversed in public within days.
By mid-November, the board planned to name Murati interim CEO while it searched for a replacement. And so it did. She held the job for about three days. Seventy-two hours, more or less, depending on whose clock you use: the Times says she rejected the role after two; Wikipedia counts three. In that window, the board that had elevated her could not, or would not, explain its reasoning, either to her or to the 770 employees she suddenly led. "I realized the board had not followed a process that could be trusted with regards to firing Sam," she would testify in 2026. "As we can see in retrospect, the way they handled it caused complete and utter chaos."
What she did with her three days was, mostly, hold the building up. Text messages later shown in court reveal her in close contact with Microsoft's CEO, Satya Nadella, asking him to call her to coordinate "how to talk with sam here." She called Nadella "a voice of reason." She worried, with the cold practicality of someone who had run a technical organization, about the talent walking out the door. "Very important we don't lose researches [sic] to Demis or Elon," she texted him, meaning Demis Hassabis of Google DeepMind and Musk, both of whom were circling.
— Mira Murati, deposition testimony shown in Musk v. Altman, May 6, 2026OpenAI was at catastrophic risk of falling apart. I was concerned about the company completely blowing up.
Then she joined the letter. Ninety-five percent of OpenAI's employees, by the company's later count, signed an open letter demanding Altman's reinstatement and threatening to follow him to Microsoft. Her name was on it. The senior leadership team she led as interim CEO, OpenAI would later say, unanimously asked for his return. The board replaced her with Emmett Shear, the cofounder of Twitch. Shear lasted barely longer than she had. Five days after the firing, Altman was back. Murati returned to being CTO.
Was it a betrayal of what she had told the board in October? Was it loyalty to Altman, as El País framed it, describing her standing by "her friend"? Or was it something less personal and more engineered: the conclusion that the institution's survival came before the question of who ran it, and that an organization which cannot survive its own governance is not one whose governance questions matter? She has never fully said. In June 2026, onstage at Bloomberg Tech with Emily Chang, she described November 2023 as a failure of governance, and named one regret: not pausing on the transition plan. Whether removing Altman had been right, she left unanswered. She seems to have decided it was the wrong question.
Inside OpenAI, people started calling the whole episode "The Blip."
The Blip
On March 7, 2024, the New York Times published the first detailed account of Murati's role in the firing, the October conversations and the private memo. OpenAI's response was immediate and unanimous in tone. A spokeswoman, Hannah Wong, said Altman was "an effective C.E.O. who is open to different points of view." Murati's lawyer, Marc H. Axelbaum, said the claims that she had approached the board to get Altman fired or supported its actions were "flat wrong," adding that she had been "perplexed at the board's decision then, but is not surprised that some former board members are now attempting to shift the blame to her."
Her own statement to employees, sent after publication, is a small masterpiece of the precise denial. She said she and Altman "have a strong and productive partnership and I have not been shy about sharing feedback with him directly." She did not reach out to the board, she said, but "when individual board members reached out directly to me for feedback about Sam, I provided it — all feedback Sam already knew." That did not mean, she wrote, that she was "responsible for or supported the old board's actions."
Every clause holds weight. Did not reach out: a statement about who initiated contact. All feedback Sam already knew: a statement that she had told him first, to his face, in the memo. It is the denial of a person who understands that the difference between "I approached the board" and "the board approached me and I answered honestly" is the difference between a coup and a conversation, and who is not willing to pretend the conversation did not happen.
A week later the Times appended an editors' note. An earlier version of the article had said Murati did not respond to a request for comment. In fact, she said, two incorrect phone numbers had been used to try to reach her, and she never got the request. It is a tiny correction, buried at the bottom of a story about manipulation and candor, and it is the most Murati detail in the entire affair: even the press, it turned out, had been calling the wrong source.
The new board that emerged from the Blip included Larry Summers, the former Treasury Secretary who had said in 2005 that innate differences might explain why fewer women succeed in science and engineering. El País noted, dryly, that Murati was by then the only high-ranking woman at OpenAI. She went on doing the job. In May 2024 she demonstrated GPT-4o, the model that brought humanlike spoken conversation to ChatGPT, and weathered the subsequent dispute with Scarlett Johansson over a voice in the product that sounded uncannily like the actress. She presented Sora, the video generator. She sat for interviews.
Some of those interviews went badly. On June 8, 2024, at an "AI Everywhere" event at Dartmouth's engineering school, moderated by trustee Jeffrey Blackburn before an audience of about 200, she argued that smarter AI would also be safer AI. Then she said something about the creative industries that would follow her for years: "Some creative jobs maybe will go away, but maybe they shouldn't have been there in the first place if the content that comes out of it is not very high quality." The Guardian read it as "a veiled threat against artists and writers" at a moment when OpenAI was being sued by writers and newspapers. Dartmouth students criticized her. The next day, Dartmouth gave her an honorary Doctor of Science at commencement. Honored and denounced on the same campus, on consecutive days, by the same college that had once watched her build race cars.
The remark is worth examining rather than dismissing, because it reveals the one place where her habitual precision failed her. The qualification — if the content that comes out of it is not very high quality — was there, but it came last, and it is the first half of a sentence that people remember. An engineer's instinct is to state the general case and then constrain it. A public figure's survival depends on doing the opposite. For a woman who had built a reputation on saying exactly what she meant, it was a lesson in how other people hear.
Leaving on the Upswing
On the morning of Wednesday, September 25, 2024, Murati told Altman she was leaving. A few hours later she told the world, in a memo to employees that she also posted to X.
"After much reflection, I have made the difficult decision to leave OpenAI," it began. "My six-and-a-half years with the OpenAI team have been an extraordinary privilege." She thanked Altman and Greg Brockman "for their trust in me to lead the technical organization." She celebrated the recent releases of speech-to-speech and the o1 reasoning model as "the beginning of a new era in interaction and intelligence." And then, the only line anyone remembered: "I'm stepping away because I want to create the time and space to do my own exploration."
The note is generous, and it is diplomatic, and it says nothing. That was the point. Levy would later call it "very generous and diplomatic" to her face, and she did not disagree.
Altman's reply on X was warm. "It's hard to overstate how much Mira has meant to OpenAI, our mission, and to us all personally," he wrote, adding that he felt "personal gratitude towards her for the support and love during all the hard times." His note to employees that afternoon was more revealing, because it relayed her reasons, or at least his version of them.
— Sam Altman, note to OpenAI employees, September 25, 2024I obviously won't pretend it's natural for this one to be so abrupt, but we are not a normal company, and I think the reasons Mira explained to me (there is never a good time, anything not abrupt would have leaked, and she wanted to do this while OpenAI was in an upswing) make sense.
There is never a good time. Anything not abrupt would have leaked. She wanted to go while things were going well. Three reasons, each of them an engineer's reason, each about timing and information control. Within hours, Bob McGrew, the chief research officer, and Barret Zoph, the vice president of research for post-training, announced their departures too. Altman said they had made their decisions "independently of each other and amicably," and that it "made sense to now do this all at once."
The context was not subtle. OpenAI was collecting more than $3 billion a year in revenue while spending about $7 billion, according to the Times, and was in talks to raise money at a valuation of up to $150 billion, nearly double its last round. That same day, Reuters reported that OpenAI planned to restructure into a for-profit benefit corporation no longer controlled by its nonprofit board, and that Altman would receive equity for the first time; the Verge put the stake at a reported 7 percent. The capped-profit structure she had once defended to Levy as "not something that was done lightly," the one whose key purpose was "protecting the mission of the nonprofit," was being dismantled. The investor no longer had to believe it was the mission.
For several months she said almost nothing. At Wired's Big Interview in San Francisco in December 2024, Levy asked what she was doing next. "I'm not going to share much about what I'm doing next because I am figuring out what that looks like," she said. "I'm in the midst of it." She asked the audience to "totally ignore the noise and the speculation," adding that there was "too much noise and obsession about who is leaving the labs." On AGI she was serene: "Right now, it feels quite achievable." On the technology itself, she sounded like the daughter of literature teachers after all: "This technology is not intrinsically good or bad. It comes with both sides."
Somewhere in this period a slim, quickly assembled book appeared under the title MIRA MURATI: The Biography and Unveiled Story Behind Her Interim CEO Appointment at OpenAI. Its subtitle tells you which seventy-two hours the market assumed would define her. She was, at that moment, about to make the subtitle obsolete.
Control, Weighted One Hundred to One
On February 18, 2025, she announced Thinking Machines Lab. The name was borrowed, knowingly or not, from a distinguished lineage of failure: Thinking Machines Corporation, the supercomputer maker that went under in 1994, and before that the "thinking machines" Frank Herbert's Dune imagined humanity outlawing. The company itself was a public benefit corporation based in San Francisco, with a mission to make AI systems "more widely understood, customizable, and generally capable." Its launch post diagnosed the industry with unusual bluntness: "Knowledge of how these systems are trained is concentrated within the top research labs, limiting both the public discourse on AI and people's abilities to use AI effectively."
The team was the announcement. Twenty-nine employees were named, about two-thirds of them from OpenAI. John Schulman, an OpenAI cofounder who had helped build ChatGPT and had decamped to Anthropic only five months earlier, became chief scientist. Barret Zoph, who had walked out of OpenAI the same afternoon she did, became chief technology officer. Lilian Weng, a former OpenAI vice president who had worked on safety and robotics, joined, as did Luke Metz, a ChatGPT co-creator; Jonathan Lachman, OpenAI's former head of special projects; and researchers from Meta, Google DeepMind, Character.AI, and Mistral. Bob McGrew and Alec Radford, the lead researcher behind OpenAI's original GPT papers, signed on as advisers.
Schulman is the useful figure here, the clearest rhyme. He was a cofounder of the most famous AI company in the world who left it, tried a rival, and left that too within months, in search of a place where the research would be shared and the work would be the thing. He was not the only senior person in this story who would discover that the institution mattered more than the logo.
The money came quickly and in absurd amounts. In October 2024, Reuters had reported she was raising more than $100 million. By April 2025 the target was $2 billion, with a minimum check reported to be $50 million; Fortune's Allie Garfinkle pointed out the oxymoron of a seed round that size. On July 15, 2025, Thinking Machines closed it: $2 billion at a $12 billion valuation, led by Andreessen Horowitz, with Nvidia, AMD, Cisco, Accel, Jane Street, and ServiceNow participating. It was the largest seed round in history. The company had no product.
One of the investors was a country. The government of Albania put in roughly $10 million, an investment that required amending the national budget for 2025. The state broadcaster's English-language headline announced that Albania was investing in "Machine Thinking Lab," transposing the name in the way a proud relative might at a wedding. One account reported that the money came through the Albanian Investment Corporation, from revenue on confiscated assets. A small, poor, formerly sealed-off nation that had once made the humanities "a bit questionable," buying a sliver of a frontier AI lab run by a woman who left at sixteen. There is no clean way to read that, and nobody should try.
The term sheet was the real tell. The Information reported, under the headline "Thinking Machines Lab CEO Has Unusual Control in Andreessen-Led Deal," that Murati held a deciding vote on board matters, weighted to give her a majority on her own. Founding shareholders' votes carried one hundred times the weight of regular shares.
Set that structure beside the one she had just left. At OpenAI, she had watched a six-person board fire a CEO without, in her later words, "a process that could be trusted," and then watched that board dissolve under employee and investor pressure in five days. She had watched a nonprofit's control over a for-profit be argued away. She had, by her own testimony, spent years asking a CEO "to lead with clarity, and not undermine my ability to do my job." Her answer, when it was her turn, was not a more elaborate system of checks. It was the elimination of ambiguity about who decides.
It is possible to read this as hypocrisy: the executive who complained about an unaccountable CEO making herself unaccountable. It is also possible to read it as a lesson learned in the most expensive classroom in Silicon Valley, where what destroyed the company was not a bad leader but contested legitimacy, with nobody certain who was in charge or whether the people in charge were telling the same story to everyone. "Morality is not everything," she said at Bloomberg Tech in 2026, arguing for decision-making structures and transparency over trust in good intentions, "because even people who are well-intentioned can make mistakes." She was describing OpenAI. She was also, unavoidably, describing herself.
The structure was tested almost immediately. Mark Zuckerberg, building Meta Superintelligence Labs on a budget of nine-figure pay packages, reportedly offered to buy the company. She declined. He then went after her people, especially Andrew Tulloch, a cofounder who had worked at both Meta and OpenAI. In October 2025, Tulloch left for Meta. Control of the boardroom, it turned out, did not extend to control of the hallway.
What the Founders Took With Them
On October 1, 2025, Thinking Machines shipped its first product. It was not a chatbot. It was Tinker, an API that let researchers fine-tune open-weight models with their own data and their own training algorithms, while the company handled the scheduling, the distributed infrastructure, and the failure recovery. It became generally available in December. By the summer of 2026, a public GitHub repository of recipes for using it, the "tinker-cookbook," had been starred more than four thousand times.
Tinker was a curious first product for a company valued at $12 billion. It was infrastructure for other people's ambitions, a middle layer in a market that rewards owning either the bottom or the top. It was also the most direct possible expression of the launch post's complaint: if knowledge of how models are trained is concentrated in a few labs, then build the tool that moves the training into the hands of people outside them.
Then, in January 2026, the founding team started to come apart in public. On January 15, TechCrunch reported that Barret Zoph and Luke Metz were leaving to return to OpenAI, the company all three of them had walked out of together sixteen months earlier. Within the week the New York Times had a headline for it: "The A.I. Start-Up Soap Opera Riveting Silicon Valley." Other researchers followed. Reports that the company had been in talks to raise at roughly $50 billion said those talks had collapsed. A LinkedIn post from a small consultancy, the kind of thing that circulates when founders want to feel smarter than other founders, called it "the Mira Murati problem": a company whose public philosophy of broadly useful, understandable AI was "structurally incompatible" with the winner-take-all math of its valuation.
She answered in the only currency she seems to trust, which is output and time. In March 2026, Thinking Machines announced a strategic partnership with Nvidia: an undisclosed investment and a multiyear agreement to deploy a gigawatt of Nvidia's Vera Rubin computing capacity. The woman who had begun at OpenAI on supercomputing strategy was now negotiating power at the scale of power plants. By mid-2026 the staff numbered more than 150, four times its size at launch, and the company moved its headquarters to 2300 Harrison Street in the Mission District. Asked by TIME about the departures, she said success isn't measured by how much of the original team remains, but by what the company delivers.
It is an unsentimental answer, and it is consistent. At OpenAI she had told Levy there was "too much obsession about who is leaving the labs." She had been the subject of that obsession. Now she was its object from the other side, and she said the same thing.
What They Think About
On July 15, 2026, a year to the day after the seed round closed, Thinking Machines released Inkling: a 975-billion-parameter multimodal model, open weights, under the Apache License, downloadable by anyone. Sixteen days later it released Inkling Small, a 276-billion-parameter distillation with comparable performance.
The details of how Inkling was made are more interesting than its size. It drew on DeepSeek-V3, a Chinese open-weights model, for its architecture, and on Moonshot AI's Kimi K2.5 for synthetic data in post-training. Media outlets described it as the strongest open-weights model released outside China, ahead of Nvidia's Nemotron, Mistral's models, and Google's Gemma, while trailing the open releases from Chinese labs such as DeepSeek, Z.ai, and MiniMax. Rest of World ran the release under a headline about how Silicon Valley keeps using the Chinese AI that Washington wants to contain. The company itself did something almost unheard of in its industry. It said, plainly, that Inkling was not the best model on the market. It called it a broad, balanced generalist designed for customization.
One observer, who admitted he had written off her departure from OpenAI as just another routine exit, wrote that it was this admission, more than any benchmark, that made him trust everything else the company said. The minister of truth, applying the job description to herself.
The deeper bet had been previewed two months earlier. In May 2026 Thinking Machines unveiled what it called interaction models: systems that take in audio, text, and video continuously, in twenty-millisecond increments, rather than in the turn-based rhythm of every chatbot since ChatGPT, prompt, wait, answer, wait. A model that can keep listening while you talk. At Bloomberg Tech in June, Murati described the company's whole thesis through the image of a tandem bicycle, both riders pedaling, both with hands on the handlebars. She conceded, without being asked, that the autonomy-first path most labs were racing down is probably the faster way to advance AI. She was choosing the slower one on purpose.
— Mira Murati, to TIME, 2026If humans want to delegate everything in their lives to AIs ... there will be options. But if they don't want that, they shouldn't feel like [they] have no choice.
The bet is not without its own contradiction, and TIME named it: the more completely a system absorbs a person's expertise through customization, the better positioned it becomes to replace that person. Murati's answer is not to deny the risk but to keep humans continuously in the loop, through interfaces so rich and so immediate that stepping out of the loop is never the default. Many domains, she argues, depend on "human participation, on the messiness of the real world." She expects math and coding to be automated first. The fields that resisted her as a child, history and sociology and everything else where the sources were ambiguous, she expects to resist the machine too.
Her company's biography, posted on Dartmouth's AI website, now describes the mission in language that sounds like someone who has spent time thinking about who owns what: "Most AI is trained by a handful of companies, frozen at release, and then rented by the people who use it. Thinking Machines lets people own their intelligence." It notes that she served on the United Nations High-Level Advisory Body on Artificial Intelligence. It also notes she was born in Albania, in a single clause, near the end, the way emigrants tend to mention it.
Two days before her deposition was played in Oakland, she walked up the steps of the Met in a gown. Two months after it, her company gave away a model it had built partly on the work of a Chinese lab and said, in writing, that it was not the best. These are not contradictions, exactly. They are what it looks like when someone has decided that being consistent is more important than being consistent-looking.
And on GitHub, among the Thinking Machines repositories, between the cookbook and a fork of a Redis library for Rust, there is a small project with more than a thousand stars, released under the MIT license in the fall of 2025. It is called
batch_invariant_ops. Its purpose, as the name says to anyone who knows how inference servers work, is to make a model's output stop depending on what else the machine happens to be processing at the same moment. To make the machine, asked the same question twice, give the same answer twice. Not to one person one thing and to another person the opposite. The same thing, to everyone, every time.Part IIThe Playbook
What follows is an attempt to extract operating principles from a career that has mostly been narrated as a sequence of headlines: the ChatGPT launch, the three days in November, the record seed round, the defections. Read as a pattern rather than a plot, Murati's decisions are unusually coherent. They return again and again to verification, timing, structure, and the patient building of trust with people who have no particular reason to extend it. None of these principles require being a frontier-AI executive. Most of them are about what to do when the information around you is unreliable, which is to say, most of the time.
The principles
- 1.Verify through a second channel, then act quietly.
- 2.Stand on ground that can't be argued with.
- 3.Collect domains before you chase generality.
- 4.Learn to recognize "too early."
- 5.Treat deployment as the experiment.
- 6.Make the counterparty believe your mission, not share it.
- 7.Ask the question that sits between two metrics.
- 8.In a crisis, keep the institution alive before you litigate it.
- 9.Make the denial precise enough to survive scrutiny.
- 10.Leave on the upswing, and leave abruptly.
- 11.Write the governance you wish you'd had.
- 12.Say out loud that you are not the best.
M
The Arc in Dates
Each role added a domain; each exit was timed.
1988
Born December 16 in Vlorë, in the final years of Albania's communist regime.
2005
Leaves Albania at sixteen for Pearson College UWC on Vancouver Island.
2012
Completes BEng in mechanical engineering at Dartmouth's Thayer School, after a BA from Colby in 2011.
2013
Joins Tesla; works on Model S dual motor, then leads Model X program to launch.
2016
VP of product and engineering at Leap Motion; concludes spatial computing is "too early."
2018
Joins OpenAI in June as VP of applied AI and partnerships.
2022
Becomes CTO in May; ChatGPT launches in November.
2023
Interim CEO for about three days after Altman's November 17 firing.
2024
Leaves OpenAI September 25; McGrew and Zoph follow within hours.
2025
Launches Thinking Machines Lab (Feb.); raises $2B at $12B (July); ships Tinker (Oct.).
2026
Zoph and Metz return to OpenAI (Jan.); Nvidia gigawatt deal (Mar.); deposition aired (May); Inkling released (July).
Principle 1
Verify through a second channel, then act quietly
The most consequential thing Murati described in her 2026 deposition was not an accusation. It was a procedure. Told by her CEO that legal had cleared a model to skip the deployment safety board, she asked the general counsel the same question, found the answers didn't match, and routed the model through the board anyway. She didn't escalate the discrepancy into a confrontation. She didn't broadcast it. She corrected the outcome and kept the evidence.
This is the behavior of someone who grew up in a place where "the source of information and truthfulness was hard," and who drew the practical conclusion: never let a single source carry weight it can't bear. The quiet part matters as much as the verification. Public confrontation over a discrepancy forces the other party to defend their version; a quiet correction simply makes their version irrelevant. The screenshots she reportedly provided to Sutskever suggest the same instinct. Her complaints about Altman were documented, not felt.
The risk is obvious: a habit of silent correction can look, from outside, like complicity, and it delays the moment when a pattern becomes undeniable. Murati paid for that delay in November 2023, when the board acted on concerns she had raised but could not explain them to anyone.
Tactic: When a critical claim reaches you through one person, confirm it with a second person who has independent knowledge before you act on it, and keep a written record of any mismatch.
Principle 2
Stand on ground that can't be argued with
Murati chose mathematics as a teenager because, in post-communist Albania, history and sociology were "a bit questionable." It was a choice about epistemic footing: build your competence where results can be checked independently of who holds power. The choice compounded. Olympiads led to a UWC scholarship, which led to Dartmouth, which led to an engineering career where a car either launches or it doesn't.
Operators in politically noisy environments (fast-growing startups, reorganizing enterprises, companies in crisis) face a scaled-down version of her childhood. Narratives shift weekly. What survives is work whose quality doesn't depend on anyone's narrative: a shipped product, a working system, a verifiable metric. Murati's authority at OpenAI rested on precisely this. She "led the whole program to launch" at Tesla, and she ran the teams that shipped ChatGPT. Nobody could revise that.
It also explains her later blind spot. In the domains she distrusted, the ambiguous, interpretive, human ones, she was sometimes careless, as the "creative jobs" remark at Dartmouth demonstrated.
Tactic: Anchor your reputation in at least one output that can be independently verified, so that your standing survives any change in who controls the story.
Principle 3
Collect domains before you chase generality
Before OpenAI, Murati worked in finance (Goldman Sachs, Tokyo), aerospace (Zodiac), automotive (Tesla), and spatial computing (Leap Motion). Each was a domain where intelligence had to be engineered by hand for a narrow purpose: Autopilot for driving, hand-tracking for VR. She arrived at the idea of generality by exhaustion, having seen the same problem in enough costumes to recognize it. "Forget the competence in different domains," she said of her pivot to OpenAI, "and just focus on generality."
This is the opposite of the conventional advice to specialize early. Her breadth was not dilettantism. Each move was deliberately orthogonal to the last ("I just wanted to get experience with different domains"), and together they gave her something a pure researcher lacks: an intuition for where systems actually break when they meet users, supply chains, and physics.
Tactic: Before committing to a general bet, deliberately work in two or three narrow domains that share an underlying problem, so you can see the problem's shape instead of its local disguise.
Principle 4
Learn to recognize "too early"
At Leap Motion, Murati wanted interaction with a computer to feel "as intuitive as playing with a ball." She concluded the technology was "too early," because it depended on headsets people would not wear every day. What's notable is that she said so plainly and moved on, after two years, rather than spending a decade waiting for the market.
The lesson she extracted, about "the limitations of pushing this technology to the practicality of using it every day," came back with interest. ChatGPT succeeded where OpenAI's earlier API offerings had made smaller waves not because the model was different but because the interface removed the headset, so to speak. Her later bet on interaction models at Thinking Machines is the Leap Motion idea, intuitive and continuous human-machine contact, returned to when the substrate finally exists.
"Too early" is not "wrong." It is a timing judgment, and the skill is keeping the idea alive while abandoning the vehicle.
Tactic: When a product stalls, separate "the idea is wrong" from "the enabling layer doesn't exist yet," and write down which dependency would have to change for you to come back.
Principle 5
Treat deployment as the experiment
ChatGPT began as a research instrument for gathering feedback to align GPT-4. Murati's defense of OpenAI's for-profit structure rested on the same premise: "To really understand how to make our models better and safer, you need to deploy them at scale." Her consistent position, across years of interviews, is that safety and capability are learned through contact with real users, not deduced in a lab, and that putting models in many hands accelerates the discovery of both their uses and their failures.
This is a Tesla lesson. She called Model X "going through the whole experience of design and deployment of a whole vehicle." Deployment, in that world, is where assumptions die.
The principle has costs she has acknowledged and some she hasn't: deployment at scale is also how a "small" risk meets a hundred million people. The corollary she insisted on is that iterative deployment demands real, operational safety processes. Hence "they can't just be policies and ideas."
Tactic: Ship the earliest version that generates the feedback you most need, and decide in advance which signal from real users would cause you to pull it back.
Her most candid statement about the Microsoft partnership was that Microsoft did not share OpenAI's mission, "but it was important for the investor to actually believe that it's our mission." This is a precise and underrated model of partnership. You will rarely find capital that shares your goals. You can find capital that believes you hold yours, and whose behavior is constrained by that belief.
Murati made that belief credible through relationship rather than rhetoric. People at Microsoft told Levy, unprompted, how much better she made the partnership work. During the November 2023 crisis, the relationship became load-bearing: she was texting Nadella to coordinate "how to talk with sam here," and he was, in her words, "a voice of reason."
At Thinking Machines she reapplied the principle structurally. Investors bought into a public benefit corporation where she holds the deciding vote, so their belief in her mission is enforced by the cap table rather than by goodwill.
Tactic: In any partnership, identify the one belief about your intentions that the partner must hold for the deal to work, and invest disproportionately in the relationships and structures that keep that belief true.
Principle 7
Ask the question that sits between two metrics
Fifty-five days after ChatGPT launched, with Microsoft announcing billions and school districts banning the product, Murati's first contribution to her leadership meeting was this: "Do we know the relationship between the factuality balance and the creativity evals?" She was not asking how the product was doing on either measure. She was asking about the trade-off between them, which is where the real decision lived.
Most dashboards show metrics independently. The hard problems sit in the correlations: speed versus quality, engagement versus trust, imagination versus accuracy. A leader's distinct contribution, especially a technical one, is to notice where two teams are optimizing quantities that pull against each other and make the tension explicit.
Her habit of asking "how bad are the bad things" for each product iteration is the same move pointed at risk.
Tactic: In your next metrics review, pick two numbers that different teams own and ask what happens to one when the other improves.
⚖
Two Ways to Run a Frontier Lab
Murati's choices at Thinking Machines read as point-by-point responses to what she saw at OpenAI.
| Prevailing lab approach | Murati's choice | Type |
|---|---|---|
| Board-centered governance; contested authority | CEO deciding vote; founding shares weighted 100× | Structure |
| Proprietary models rented by the token | Tinker for fine-tuning; Inkling released open-weights under Apache | Product |
| Race toward fully autonomous systems | Interaction models; humans "in the loop" by design | Bet |
| Superlative launch claims | States openly that Inkling is not the strongest model overall | Credibility |
| Team retention as a proxy for health | "Success isn't measured by how much of the original team remains" | Risk |
| Training knowledge held inside top labs | Publishes research and code; open-source repositories | Structure |
Principle 8
In a crisis, keep the institution alive before you litigate it
Murati's behavior during her three days as interim CEO is the most debated sequence of her career, and the most instructive. She had raised concerns about Altman. The board acted on concerns like hers. Then, finding that the board "had not followed a process that could be trusted," she spent the crisis keeping researchers from defecting to Google DeepMind or Musk, coordinating with Microsoft, and finally signing the letter demanding Altman's return.
The logic, reconstructed from her 2026 testimony, is sequential rather than contradictory. First, the company must survive; it was "at catastrophic risk of falling apart." Second, and only later, the question of leadership can be addressed. Her grievances about Altman were real and enduring; she repeated them under oath two and a half years afterward. But she did not let them become the reason 770 people lost their jobs.
The cost was reputational. She has been read as disloyal by both camps. Her later regret was procedural: not pausing the transition plan.
Tactic: When a crisis forces a choice between being right about a person and preserving an organization, stabilize the organization first and record your concerns for the moment it can afford to hear them.
Principle 9
Make the denial precise enough to survive scrutiny
When the Times reported in March 2024 that she had played a pivotal role in Altman's ouster, Murati's response did not deny that she had criticized him. It specified who initiated the contact ("when individual board members reached out directly to me"), the content of what she said ("all feedback Sam already knew"), and what that did not imply ("responsible for or supported the old board's actions"). Every clause was defensible and none was a lie, which is why the statement held up even after her 2026 testimony made her criticisms public.
Most denials fail because they deny too much. They contest the whole story to escape one implication, and the next revelation destroys them. A precise denial concedes what's true and contests only what isn't.
Tactic: When responding to an unflattering account, list every factual claim in it, concede the ones that are accurate, and dispute only the specific inference you can prove is wrong.
Principle 10
Leave on the upswing, and leave abruptly
Murati's departure reasons, as Altman relayed them, were three engineering judgments about timing: "there is never a good time, anything not abrupt would have leaked, and she wanted to do this while OpenAI was in an upswing." She informed the CEO in the morning and the world hours later. Two senior research leaders announced their exits the same day.
Leaving on an upswing protects both parties. The institution isn't destabilized during a crisis, and the person leaving isn't read as fleeing failure. The abruptness protects the narrative: a long, leaky departure becomes a story told by others. And the diplomatic memo, generous and essentially content-free, removed any grievance that could be quoted against her or the company.
She then said almost nothing for months, refusing to share plans "because I am figuring out what that looks like."
Tactic: Plan exits for a moment of strength rather than conflict, tell the most senior affected person first and the wider world within hours, and resist explaining your next move before you know it.
Principle 11
Write the governance you wish you'd had
At OpenAI, Murati watched governance fail in both directions: a board that acted without a process people could trust, and then a board that collapsed under pressure from employees and investors. Her structure at Thinking Machines eliminates the ambiguity. She has a deciding vote weighted to give her a majority, and founding shares vote at one hundred times the weight of regular ones. The Information called it "unusual control."
Paired with that concentration is an argument for structure over virtue: "Morality is not everything," she said in 2026, because "even people who are well-intentioned can make mistakes." The tension is real. She concentrated power while arguing that good intentions are insufficient protection. Her implicit answer is that clarity of authority is itself a safeguard, and that the catastrophe at OpenAI came not from one person having too much power but from nobody knowing who had it.
The structure does not protect against everything. It could not stop Tulloch, Zoph, or Metz from leaving.
Tactic: Before raising outside capital, decide in writing who holds the final vote on what, and make sure your structure answers the question "who decides?" before a crisis forces someone to ask it.
Principle 12
Say out loud that you are not the best
When Thinking Machines released Inkling, it did not claim the frontier. It said the model was not the strongest overall, described it as a balanced generalist for customization, and was open about building on Chinese open-weights work from DeepSeek and Moonshot. In an industry that runs on superlatives, the admission became, for at least one skeptical observer, the reason to believe everything else the company said.
This is the external version of the trait that runs through her whole career, a refusal to let any source, including her own company, overclaim. It is also strategically shrewd. A company valued at $12 billion with founding-team departures has a credibility problem, and nothing repairs credibility like a voluntary concession that turns out to be accurate. Pragmatism about inputs (using the best available open weights regardless of origin) and candor about outputs reinforce each other.
Tactic: In your next launch, state one specific way your product is not the best available, and let that honesty earn credibility for the claims you do make.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Mira Murati, to Steven Levy, WIRED, July 2023I very quickly believed that AGI would be the last and most important major technology that we built, and I wanted to be at the heart of it.
— Mira Murati, to Fortune, 2023At the end of the day, you need to institutionalize and operationalize these things, and they can't just be policies and ideas.
— Mira Murati, deposition testimony, Musk v. Altman, May 2026I had an incredibly hard job to do in an organization that was very complex. I was asking Sam to lead, and lead with clarity, and not undermine my ability to do my job.
— Mira Murati, memo to OpenAI employees, September 25, 2024There's never an ideal time to step away from a place one cherishes, yet this moment feels right.
— Mira Murati, to TIME, 2026We have machines that can think; they already exist today. But what they think about should still be up to us.
Maxims
- Ask twice. A claim confirmed by one source is a rumor; a claim contradicted by a second is a decision.
- Fix it before you name it. Correct the outcome quietly, and keep the record for when the pattern needs proving.
- Build where results can't be revised. Shipped work outlasts every change in who tells the story.
- Wander deliberately. Several narrow domains will teach you the shape of the general problem faster than one deep one.
- Too early is not wrong. Abandon the vehicle, keep the idea, and note what has to change before you return.
- Deployment is the lab. Real users reveal both the uses and the failures that no internal test anticipates.
- Partners needn't share your mission. They need to believe you hold it, and your structure should keep that belief honest.
- Save the institution, then settle the score. An organization that collapses can't benefit from your being right.
- Decide who decides. Contested authority does more damage than concentrated authority.
- Concede what's true. The admission that you are not the best is often what makes the rest of your claims credible.
