Part IThe Story
On Monday, May 11, 2026, in a federal courthouse in Oakland, a man who had spent most of his adult life teaching machines to predict the next word was asked, under oath, to account for his own. The proceeding was Elon Musk's lawsuit against OpenAI, the company Musk had helped found and now accused of abandoning its nonprofit soul. Musk wanted Sam Altman and Greg Brockman removed and as much as $180 billion moved from the for-profit arm back to the nonprofit parent. Judge Yvonne Gonzalez Rogers presided. OpenAI's attorney, Sarah Eddy, did some of the questioning. A courtroom sketch artist named Vicki Behringer drew the witness in the colored-pencil style that turns every defendant, plaintiff, and bystander into a figure from a Victorian novel: high forehead, close-cropped hair, a face given to stillness.
The witness was Ilya Sutskever, and he confirmed several things in short order. He confirmed that he had turned down a $6 million salary at Google to join OpenAI in its first weeks. He confirmed that he held a stake in OpenAI worth roughly $7 billion, a disclosure that made him, as Forbes put it, the second new billionaire revealed by the trial. He confirmed that he had spent about a year gathering evidence that Altman, his cofounder, displayed "a consistent pattern of lying," including "undermining and pitting executives against one another." He said the behavior was "not conducive to any grand goal," and he meant the grandest goal he could imagine, the safe creation of a mind greater than ours. He said that his abrupt reversal in November 2023, when he helped fire Altman and then, within days, begged for his return, had been a "Hail Mary." "I felt that had I not done this," he said, "the company would be destroyed."
And he said one other thing, almost in passing, that should have been more widely noticed. During the five days in which OpenAI nearly ceased to exist, while much of Silicon Valley refreshed its feeds every ninety seconds, Sutskever had mostly stayed off the internet. He did not know, for a time, that Microsoft was preparing to hire Altman. He was at the center of the most-watched corporate drama of the decade, and he was not watching it.
For two and a half years, the internet had been asking a question of him, half meme and half liturgy: What did Ilya see? The phrase implied a revelation, a glimpse of something in the weights of an unreleased model so terrible that a quiet scientist had been driven to a boardroom coup. The trial suggested a more mundane answer, and a stranger one. What Ilya saw, it turned out, was mostly a man. What he did about it was mostly a mistake. And what he did next, which was to build a company with no product, no revenue, and a valuation larger than most airlines, was either the purest expression of a scientific conviction ever capitalized by venture money or an extraordinary act of faith by everyone who trusted that the person who had been right about the machines would be right again.
He had been right about the machines for a long time. That was the problem, and the asset.
By the Numbers
From a Bedroom to Vera Rubin
2Nvidia GTX 580 GPUs, about $500 each at release, used to train AlexNet in a Toronto bedroom
15.3%AlexNet's top-5 error in 2012, more than 10.8 points ahead of the runner-up
198,000+Citations of the AlexNet paper as of early 2025
$6MGoogle salary he testified he turned down to join OpenAI
52Pages in the memo he wrote documenting Altman's alleged dishonesty
$32BSSI's reported valuation in 2025, with no product and about 20 employees
$5BNvidia's planned investment in SSI, announced July 27, 2026
The Gibberish Was Grammatical
In 2010, a writer from the University of Toronto Magazine found a 24-year-old doctoral student sitting in front of a monitor with his fingers hovering over the keyboard, considering a blinking cursor. The student wanted to demonstrate what he called his most exciting research project yet. "I give it an initial segment of text," he said, typing. "And I say, from this text, keep on producing text that you think looks like Wikipedia."
The machine obliged. It produced, among other things, this: "Akkerma's Alcesia Minor (including) of Hawaiian State Rites of Cassio. Other parish schools were established in 1825, but were relieved on March 3, 1850."
It is nonsense, of course. It is also, if you sit with it, a little eerie. The parentheses close. The dates are plausible dates. The second sentence has the cadence of an actual encyclopedia entry about an actual parish in an actual place. The network had 2,000 artificial neurons and one job, which was to guess the next character in a sequence, one letter at a time, and from that single humble objective it had extracted the architecture of a language. "It spits out one letter at a time, which happen to form words," Sutskever explained. "It discovered that words exist, and it discovered grammar."
What it had not discovered was the truth. No parish schools were relieved on March 3, 1850, at least not by this paragraph's authority. The machine could produce the form of knowledge without its content, the sound of candor without its substance. Twelve years later, a much larger descendant of that network would be released to the public as ChatGPT and gain 100 million users in under two months, and the central complaint about it, the one that would embarrass lawyers who trusted it in court filings, would be precisely this: it was fluent, and it made things up.
The magazine noted two other things about the young man. In June of that year he had become the first Canadian to receive a Google PhD fellowship, which came with $50,000 over two years. And he was strikingly reluctant to talk about applications. His next step, he said, was to train the network on New York Times articles to see whether it could identify authorship; he conceded, without much enthusiasm, that this might one day become plagiarism software. Then he said the sentence that, in retrospect, reads like an epigraph to everything that followed:
"If you know your destination, you will probably get there, but if you don't, there is more of a chance of stumbling upon something really interesting."
He was not, then, a man with a business plan. He was a man with a hunch about prediction, and the patience to keep pressing a single key.
A Field of Desolation and Despair
He was born in 1986 in a city that has since changed its name, in a country that has since ceased to exist. Gorky, now Nizhny Novgorod again, had been renamed for Maxim Gorky, whose pen name means bitter and whose most famous short tale, a staple of Soviet childhood, concerns a young man named Danko who tears his own burning heart from his chest and holds it overhead to lead his people out of a killing forest. Sutskever's family was Jewish. When he was five they emigrated to Israel and settled in Jerusalem. He still speaks Russian and Hebrew in addition to English, and the triple fluency is not incidental to the story; he would spend his life moving between grammars.
He was five, too, when he first saw a computer. "I was utterly enchanted," he told the University of Toronto Magazine in 2022. By eighth grade in Jerusalem he was taking classes at the Open University of Israel. As a teenager, he said, he thought a great deal about questions most teenagers leave to the late-night portion of a party: "I remember thinking a lot about the nature of existence and consciousness … about souls and intelligence. I felt very strongly that learning was this mysterious thing: humans clearly learn, computers clearly don't."
At 16 the family moved again, to Canada. He attended a Toronto high school for exactly one month. The University of Toronto then admitted him, out of Grade 11, into its mathematics program as a third-year undergraduate. This was 2002. He immersed himself in upper-year courses, and after his second year of math he walked into the office of Diane Horton, the associate chair for undergraduate studies, and asked whether there were any research opportunities that involved machine learning. Horton suggested he speak to Geoffrey Hinton.
Hinton deserves his own compression. A British-born cognitive psychologist turned computer scientist, he had spent decades defending the idea that networks of simple artificial neurons, trained by an algorithm called backpropagation, could learn to do what brains do. For most of that time he was regarded by the mainstream of his field with something between politeness and pity. He would leave Google in 2023 to warn that the technology he had helped create might be dangerous, and in 2024 he would win the Nobel Prize in Physics for the work that made it possible. In between, he supervised a run of graduate students who went on to build much of the modern industry. Of all of them, Sutskever was the one who knocked on his door before he was technically old enough to be a graduate student at all.
Hinton gave him a paper to read. Sutskever came back to discuss it. "His immediate reaction to things," Hinton later told The Globe and Mail, "were reactions that had taken experts in the field quite a long time to come up with." The first project was modest, an attempt to improve an algorithm called Stochastic Neighbor Embedding; a paper eventually came out of it, a few years later. The year was 2003.
It is difficult now, when neural networks run on phones and in courtrooms and inside the essays of high school juniors, to convey how unpromising this all was. Sutskever has tried. "The field of AI back then when I was starting out was not a hopeful field," he said on a podcast in 2021. "It was a field of desolation and despair. No one was making any progress at all." From 1990 to 2010, by the sober accounting of the textbooks, neural networks were simply not better than other machine-learning methods — support vector machines, kernel regression, AdaBoost — and computer vision advanced mostly through hand-engineered features with names like SIFT and HoG, built by humans who decided in advance what an edge or a corner should look like. The belief that features could be learned directly from data was, in the field's own retrospective phrase, "a minority position."
He took the minority position and held it for a decade. Asked years later what working with Hinton had meant, he said it gave him "the opportunity to work on some of the most important scientific problems of our time and pursue ideas that were both highly unappreciated by most scientists, yet turned out to be utterly correct." The sentence is notable for its lack of false modesty. Utterly correct. He does not say promising or vindicated in part. He was a young man in a desolate field who believed he could see something nobody else could, and he turned out to be right, and the experience of having been right in exactly that way — alone, early, against the room — is the key to almost every decision he made afterward, including the bad ones.
His bachelor's degree came in 2005, his master's in 2007, his doctorate in 2013, with a thesis titled Training Recurrent Neural Networks. Along the way he published papers with titles that read, to a layperson, like a wine list: The Recurrent Temporal Restricted Boltzmann Machine (2008), Training Recurrent Neural Networks with Hessian-Free Optimization (2011, with James Martens), Generating Text with Recurrent Neural Networks (2011, with Martens and Hinton). The last of these was the cursor and the parish schools. On his faculty-page list of publications, beside one 2010 paper on temporal kernel networks, he appended a small, scrupulous footnote: but note that the idea was invented much earlier. He links the earlier work. It is the kind of thing almost no one does.
Ilya Thought We Should Do It
In 2011, Geoffrey Hinton began asking colleagues a question that sounds, in retrospect, like a man at the end of his patience: "What do I have to do to convince you that neural networks are the future?" Jitendra Malik, a Berkeley computer-vision eminence and a neural-network skeptic, suggested he try the PASCAL Visual Object Classes challenge. Hinton said its dataset was too small. Malik then pointed him toward ImageNet.
ImageNet was the work of Fei-Fei Li, a Stanford computer scientist who had begun assembling it in 2007 on the unfashionable premise that vision would advance not through cleverer algorithms but through vastly more data. Using workers on Amazon Mechanical Turk and the WordNet hierarchy of English nouns, Li and her collaborators labeled more than 14 million images across 22,000 categories. The project was met, initially, with skepticism. It became the foundation of an annual competition, the ImageNet Large Scale Visual Recognition Challenge, in which teams tried to teach software to sort photographs into 1,000 categories. Li, an immigrant who had arrived in the United States as a teenager, had built a test that nobody's methods could yet pass. It was waiting for someone.
What happened next is often told as Hinton's story, and Hinton himself has told it most economically. "Ilya thought we should do it," he said, "Alex made it work, and I got the Nobel Prize."
Alex was Alex Krizhevsky, a fellow graduate student in Hinton's lab, who had a rare and unglamorous gift: he could program graphics processing units for general computation, the chips designed to render video-game explosions bent instead toward linear algebra. Before 2011 he had already written a piece of software called cuda-convnet, which trained small convolutional networks on a modest dataset called CIFAR-10 using a single GPU. Sutskever's contribution, by every account in the record, was persuasion. He convinced Krizhevsky to scale the thing up to ImageNet. Krizhevsky extended cuda-convnet to run across multiple GPUs. Hinton served as principal investigator.
The hardware was two Nvidia GeForce GTX 580 cards, each with three gigabytes of memory and a release price of about $500. The network did not fit on one, so it was split in half, each half living on its own card, communicating only at certain layers, a kludge born of a memory limit that became, in the architecture diagrams reproduced in textbooks for the next decade, a kind of signature. The machine sat in Krizhevsky's bedroom at his parents' house. Training ran for 90 passes through 1.2 million images over five to six days. Through 2012, Krizhevsky tuned and retuned the hyperparameters, nudging the learning rate down by a factor of ten whenever progress stalled, three times in all.
The finished model had 60 million parameters and 650,000 neurons. It did something small and practical with every image it saw: it scaled the picture so its shorter side measured 256 pixels, cut out the central square, and then took random 224-by-224 patches from within it. Why 224? Because 256 minus 16 minus 16 is 224. You frame out sixteen pixels on every side, and you train on what remains in the middle.
On September 30, 2012, the three of them, under the team name SuperVision, submitted an ensemble of seven such networks. Their top-5 error rate was 15.3 percent. The runner-up was more than 10.8 points behind. In a competition where progress had been measured in fractions, this was less a victory than a change of venue. At the European Conference on Computer Vision in Florence that October, Yann LeCun — whose own LeNet, in 1989, had used essentially the same architecture at a fraction of the scale — called it "an unequivocal turning point in the history of computer vision."
— Geoffrey Hinton, on AlexNetIlya thought we should do it, Alex made it work, and I got the Nobel Prize.
Fei-Fei Li, looking back in 2024, located the meaning of the moment not in any single person: "That moment was pretty symbolic to the world of AI because three fundamental elements of modern AI converged for the first time." Data, which she had built. Compute, which Nvidia had sold for $500 a card. And the algorithms, which Hinton had defended through the long winter. AlexNet was, in a sense, LeNet with twenty years of cheap hardware and cheap data behind it. What it needed was someone to insist that the old idea, scaled, would work. Sutskever was that someone.
He later described the logic as almost syllogistic. If you train a large, deep network on a big enough dataset specifying some complicated task people do, he said, "you will succeed necessarily. And the logic for it was irreducible; we know that the human brain can solve these tasks and can solve them quickly. And the human brain is just a neural network with slow neurons."
Slow neurons. Hold onto the phrase. It will come back.
Within months Hinton, Sutskever, and Krizhevsky formed a company called DNNResearch and sold it, along with the AlexNet source code, to Google. The acquisition closed in March 2013. Sutskever, in between, had spent about two months as a postdoc with Andrew Ng at Stanford, then returned to Toronto. The AlexNet paper has since been cited more than 198,000 times. In March 2025, the Computer History Museum released the original 2012 code to the public. It is short.
Three Bullet Points
On his old university home page, which still exists in the spare HTML of an earlier internet, Sutskever summarizes his career in reverse chronology, ending with: "And in the beginning, I was a student in the Machine Learning group of Toronto, working with Geoffrey Hinton." Just above that line: "I spent three wonderful years as a Research Scientist at the Google Brain Team." The adjective is unguarded. It is one of the few times in the public record that he describes his own feelings about a place without being asked.
At Google Brain, he turned the AlexNet insight from pictures to sequences. Convolutional networks were good at images, which have spatial structure; they were poor at text, speech, and other data that unfold in time. In 2014, with Oriol Vinyals and Quoc Le, he published Sequence to Sequence Learning with Neural Networks, which mapped one sequence (a sentence in English) onto another (its French translation) using deep recurrent networks. It became a foundation of neural machine translation. He also worked on TensorFlow, Google's machine-learning framework, and was one of the many coauthors of the 2016 Nature paper describing AlphaGo.
Ten years later, in December 2024, he stood at the NeurIPS conference to accept a Test of Time award for the sequence-to-sequence paper — his third such award in three consecutive years — and showed a screenshot from the original 2014 talk in Montreal. "It was a much more innocent time," he said. He showed the photograph of the three young coauthors. "This is the 'before' and here's the 'after,' by the way."
Then he summarized the work in its entirety. "It's an autoregressive model trained on text. It's a large neural network. It's a large dataset. That's it."
That's it. The sentence could serve as a description of AlexNet, of his 2010 Wikipedia generator, of GPT-2, GPT-3, and GPT-4. He pulled up an old slide headed with what he called the deep learning hypothesis: a large neural network with ten layers can do anything a human being can do in a fraction of a second. Why that oddly specific time window? Because, he explained, if you believe artificial neurons and biological neurons are not too different, and you believe real neurons are slow, then anything one human can do in a fraction of a second, a ten-layer network should be able to do too. "We focused on 10-layer networks because that's what we knew how to train back in the day."
The slow neurons again. The argument has a peculiar structure, worth noticing because it recurs: it reasons from human limitation to machine possibility. The brain is not magic; it is a slow computer; therefore a fast computer, properly trained, can do what it does. The humility about the brain becomes an audacity about the machine. And he was candid about the audacity, describing the task of translation as "today humble, then incredibly audacious." The 2014 work, he said, was not the first autoregressive neural network, but "it was the first one where we truly believed that if you train it really well, you get whatever you want."
Whatever you want. Spoken by a man who would, a decade later, devote himself to the question of what happens when the machine wants something too.
The Linchpin
He no longer remembers the name of the restaurant. He told the University of Toronto's news office as much in 2017: a dinner somewhere along Sand Hill Road in Menlo Park, the strip of low buildings where much of the world's venture capital is housed. At the table were Elon Musk; Sam Altman, then president of Y Combinator, the most powerful seed fund in the country; Greg Brockman, the former chief technology officer of Stripe; and others. What emerged from that conversation was a plan for a nonprofit artificial intelligence lab that would counterbalance Google, which had acquired DeepMind in 2014 and was, in Musk's telling, alarmingly cavalier about safety under Larry Page.
According to The Globe and Mail, it was Altman who first reached out to Sutskever about the venture. Musk, though, did the hard recruiting. "That was one of the toughest recruiting battles I've ever had, but that was really the linchpin for OpenAI being successful," Musk said in 2023, adding that Sutskever was a "good human" with a "good heart." That Musk would later sue the company, and that Sutskever would testify in the suit, and that the testimony would include the detail of the $6 million he'd left on the table at Google, gives the word linchpin a mechanical irony. A linchpin is the small pin that keeps a wheel from sliding off its axle. Remove it and the wheel keeps turning, for a while.
The micro-biographies of the founders rhyme in an instructive way. Musk, the South African immigrant who believed AI was the gravest threat to humanity and wanted to build it first. Altman, the St. Louis–raised dealmaker who believed the same and was better than anyone alive at raising money to act on the belief. Brockman, the engineer who had built Stripe's infrastructure and would, by most accounts, outwork everyone. And Sutskever, the émigré twice over, the one who actually knew how to make the networks learn. Of the four, he was the only one who had been in the room when it worked.
He was candid about the cost. "It was an extremely difficult decision," he said in 2017. "I was very happy at the Google Brain team. It was moving forward on an extremely strong trajectory — I would be taking no risk by staying and have much to gain." He left at the end of 2015 anyway, initially as research director, later as chief scientist. He joined OpenAI's governing board in 2017. He became close with Brockman — close enough that, by one report, he officiated Brockman's wedding in 2019.
And he became, for a generation of researchers, a kind of oral tradition. Sometime around 2020, John Carmack — the programmer behind Wolfenstein, DOOM, and Quake, a man who had already reshaped one industry and was turning his curiosity to another — asked Sutskever how to get up to speed in AI. Sutskever gave him a reading list of roughly thirty to forty papers and said: "If you really learn all of these, you'll know 90% of what matters today." Carmack later recalled, "And I did. I plowed through all those things, and it all started sorting out [AI] in my head."
The list was never officially published. A reconstructed version went viral in 2024, and it is a revealing document, less for its obvious entries (AlexNet, Attention Is All You Need, the scaling-laws paper by Jared Kaplan and colleagues) than for its strange ones. It includes a tutorial on the Minimum Description Length principle. It includes a textbook on Kolmogorov complexity and algorithmic randomness. It includes Shane Legg's doctoral thesis, Machine Super Intelligence. And it includes two pieces by the complexity theorist Scott Aaronson, one of them a paper titled Quantifying the Rise and Fall of Complexity in Closed Systems: The Coffee Automaton, about why a cup of coffee with cream poured in becomes more interesting before it becomes uniformly brown. By one blogger's tally, information theory accounts for the largest share of the list's content. The deep idea underneath it — that to predict well is to compress well, and to compress well is to understand — is the same idea that animated the cursor in 2010.
Aaronson would later be recruited to OpenAI by Sutskever and Jan Leike to work on safety, and would write, when both men left, of the "tremendous admiration" he had developed for them. The list had found its author's way of thinking. It is a list for people who want to understand why prediction might be enough.
Scale, and the Closing of the Door
He is credited, in the accounts of people who were there, with establishing OpenAI's scaling ethos. The phrase sounds technical; the practice was almost theological. More parameters, more data, more compute, and see what happens. GPT-2, in 2019, was the first public collision between that ethos and its consequences: OpenAI withheld the full model on the grounds that its fluent text could be misused for misinformation and impersonation. Critics called it hype, or a betrayal of the "Open" in the company's name. The staged release established, for better or worse, a norm of treating frontier models as dual-use technology. GPT-3 followed. So did the image generator DALL-E. Then, in November 2022, ChatGPT.
The chatbot that gained 100 million users in under two months was, at its core, the 2010 project with a dozen more years of compute behind it: an autoregressive model trained on text, a large neural network, a large dataset. That's it.
His public statements across this period trace an arc from openness toward caution that he did not bother to disguise. In February 2022, he tweeted: "it may be that today's large neural networks are slightly conscious." The line triggered a predictable cycle of ridicule and defense; it is also exactly what you would expect from a teenager who had once sat in Jerusalem thinking about "souls and intelligence" and then spent twenty years building the thing he'd been thinking about. In October 2022, asked by Spencer Greenberg how a model trained only to predict the next word could seem to understand so many things, he gave a two-part definition of intelligence and then chose the humbler half: find the things people can do; if computers can do them too, then the computers are intelligent, and "the more of those things computers can do, the more intelligent they are."
On March 14, 2023, OpenAI released GPT-4 with a 98-page technical report that disclosed, in its own words, "no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar." The next day, Sutskever explained the decision to The Verge. "On the competitive landscape front — it's competitive out there," he said. "GPT-4 is not easy to develop. It took pretty much all of OpenAI working together for a very long time to produce this thing." On safety, he was more frank than any executive would be: "I would say that the safety side is not yet as salient a reason as the competitive side. But it's going to change." And then: "I fully expect that in a few years it's going to be completely obvious to everyone that open-sourcing AI is just not wise."
The founding premise of OpenAI had been that keeping powerful AI out of the hands of a single company was a safeguard. Its chief scientist now said that keeping powerful AI out of everyone's hands was the safeguard. Whether this was an evolution or a reversal depends on what you think the founders believed in the first place. What is not in dispute is that the man who reversed himself said so, plainly, on the record, the day after the decision.
Two weeks later, Dwarkesh Patel asked him why he had been able to make multiple field-defining breakthroughs when most scientists manage one or none. "Thank you for the kind words," he said. "It's hard to answer that question. I try really hard, I give it everything I've got and that has worked so far. I think that's all there is to it."
Before and After
In October 2023, a reporter from MIT Technology Review went to see him in OpenAI's unmarked office building on an unremarkable street in San Francisco's Mission District. The piece opens with a description of him at the table, head bowed, arms spread, fingers splayed on the tabletop "like a concert pianist about to play his first notes." Then: "We sit in silence."
He did not want to talk about himself. "I lead a very simple life," he said. "I go to work; then I go home. I don't do much else. There are a lot of social activities one could engage in, lots of events one could go to. Which I don't."
Which I don't. The sentence is a perfect little self-portrait: the long list of what other people do, and then the four-word refusal, deadpan, almost a punchline. Elsewhere in the same building, OpenAI was an $80 billion startup whose CEO had spent the summer on a world tour, glad-handing heads of state and filling auditoriums. Sutskever went to work and went home.
But when the conversation turned to where the technology was going, the reporter wrote, "vistas open up." "It's going to be monumental, earth-shattering," Sutskever said. "There will be a before and an after." He spoke of artificial general intelligence the way other people speak of the next iPhone: "At some point we really will have AGI. Maybe OpenAI will build it. Maybe some other company will build it." He said he thought some humans would one day choose to merge with machines. He said ChatGPT had already converted "will never happen" into "will happen faster than you think."
— Ilya Sutskever, to MIT Technology Review, October 2023I lead a very simple life. I go to work; then I go home. I don't do much else. There are a lot of social activities one could engage in, lots of events one could go to. Which I don't.
His work had changed. In July 2023, OpenAI announced a new "Superalignment" project, co-led by Sutskever and the researcher Jan Leike, with the goal of solving the alignment of superintelligent systems within four years. The announcement noted that superintelligence, though it might seem far off, could arrive within the decade. The company promised 20 percent of its computing power to the effort. According to Fortune's later reporting, that promise was never delivered.
The scientist who had spent twenty years arguing that the machines would work was now spending his days arguing that they would work too well. It produced moments that sound, in the retelling, like something from a different genre of story. In 2023, according to Karen Hao's book Empire of AI, he told his team: "We're definitely going to build a bunker before we release AGI." The reasoning was not apocalyptic so much as geopolitical — a technology that powerful would become an object of intense desire for governments, and the people who built it would need protection. "Of course," he added, by the account of people present, "it's going to be optional whether you want to get into the bunker." Other accounts describe stranger rituals still: a wooden effigy representing an unaligned AI, doused in lighter fluid and burned at a firepit in the Sierra Nevada. The man born in the city named for the author of Danko's burning heart was, reportedly, setting fire to a model of his own life's work.
It is tempting to treat these details as eccentricity. That would miss their seriousness. He had seen the 2,000-neuron network discover grammar on its own. He had seen the 60-million-parameter network learn to see. Every time he had predicted that a larger network would do something the field thought impossible, it had. If you held his track record and his logic — the brain is a neural network with slow neurons; what one human can do, a network can do — the bunker was not paranoia. It was extrapolation.
On June 5, 2023, Sutskever and Altman had appeared side by side at Tel Aviv University, in the country where Sutskever grew up, photographed together on a stage. Five months later, Sutskever told Altman he was fired.
Not Consistently Candid
On Thursday, November 16, 2023, Sutskever texted Altman to invite him to a meeting the following day. When Altman joined the Google Meet call on Friday, November 17, Sutskever told him he was fired, according to an account later posted by Brockman. That afternoon OpenAI's board announced that Altman "was not consistently candid in his communications with the board." Brockman, removed as chairman, quit the company the same day. ChatGPT had been launched nearly a year to the day before.
The board that did this had six members: three employees (Altman, Brockman, Sutskever) and three outsiders (the Quora founder Adam D'Angelo, the technologist Tasha McCauley, and the AI policy researcher Helen Toner). Remove Altman and Brockman from the count and the math is simple. Sutskever was the swing vote, and by most accounts the driving force.
What drove him took years to emerge in full. In a deposition taken October 1, 2025, and later in open court, he described a document of 52 pages, written at the request of the board's independent directors, accusing Altman of "a consistent pattern of lying" and of "pitting his executives against one another." (A New Yorker investigation in April 2026 put the dossier at roughly 70 pages; Sutskever's own sworn figure was 52.) The memo relied heavily on information from Mira Murati, OpenAI's chief technology officer, with whom he said he had discussed Altman's behavior "for a long time." In November 2025, The Information published a story about the evidence under a headline that answered the meme at last: "What Ilya Saw: Mira Murati's Screenshots About Sam Altman."
Screenshots. Not a sentient model, not an emergent capability, not a glimpse of the end. Screenshots of messages, assembled by a careful man into a document, sent through disappearing emails because he feared it would leak. "I wanted them to become aware of it," he said of the board in the deposition. "But my opinion was that action was appropriate."
There was a deeper disagreement underneath the personal one, about how fast to commercialize, how to reduce public harm, whether to seek billions from Gulf sovereign wealth funds to build chips that would compete with Nvidia. It came to a head, according to Bloomberg's sources, around OpenAI's first developer day, earlier that month. But the testimony makes clear that the safety question and the trust question were, for Sutskever, the same question. If the mission was to build a mind more capable than any human and to make sure it was honest, then a leader who was not consistently candid was "not conducive to any grand goal." You cannot align a machine through a company you cannot align.
At an all-hands meeting shortly after the board vote, Sutskever told employees that firing Altman was "the board doing its duty." Then the floor gave way. Microsoft, which owned 49 percent of the for-profit entity, moved to hire Altman and Brockman. The board appointed Emmett Shear, the former Twitch chief, as interim CEO. More than 500 of OpenAI's roughly 700 employees — eventually nearly all of its 770 — signed a letter threatening to resign and follow Altman to Microsoft unless the board stepped down. "Your conduct has made it clear you do not have the competence to oversee OpenAI," the letter said. "We are unable to work for or with people that lack competence, judgement and care for our mission and employees."
On Monday, November 20, Sutskever's name appeared among the signatories. He was demanding the resignation of a board of which he was a member, for a decision he had led. That morning he posted on X: "I deeply regret my participation in the board's actions. I never intended to harm OpenAI. I love everything we've built together and I will do everything I can to reunite the company."
In the days between, the board had held talks with Anthropic — the rival lab founded by former OpenAI researchers — about a possible merger. Sutskever recalled a call with its founders, Dario and Daniela Amodei, and testified that he was "very unhappy" about the idea. "I really did not want OpenAI to merge with Anthropic," he said. In court, years later, he put it more mildly: he had been "not excited." He acknowledged that the whole process "seemed rushed."
Satya Nadella, Microsoft's chief executive, testified in the same 2026 trial that he was never given a clear explanation for the firing, and summarized the episode in two words: "amateur city." Hao's reporting reached a similar verdict with more sympathy: Sutskever and Murati had acted too slowly, failed to consolidate power, and never won over the key executives. They had the evidence. They did not have the coalition, or a story they were willing to tell.
This was the cruelest symmetry of the affair. The board's single public sentence accused Altman of not being consistently candid. The board itself then declined, for days, to say anything more — to employees, to investors, to Microsoft, to the press. A body that had removed a man for withholding information withheld it from everyone else. The most fluent communicator in Silicon Valley was defeated, briefly, by people who would not explain themselves, and then he returned, because in the absence of an explanation his fluency filled the vacuum. On November 22, OpenAI announced that Altman would return as CEO. Sutskever left the board.
A film adaptation, Artificial, directed by Luca Guadagnino, has since been fast-tracked by Amazon MGM Studios. Andrew Garfield plays Altman. Monica Barbaro plays Murati. The Russian actor Yura Borisov plays Sutskever. It is safe to assume the movie will invent dialogue for the five days in which, by his own sworn account, the real Sutskever mostly stayed off the internet.
What Did Ilya See?
Then he vanished, or nearly. He stopped appearing at OpenAI's office. Some reports said he was leading the Superalignment team remotely; others said he no longer had access to its work. The phrase "What did Ilya see?" metastasized into a meme, a merchandising opportunity, a punchline, a genuine question asked by serious people. Musk mused about it publicly. Online, enthusiasts went treasure-hunting for the Carmack reading list, as though the answer might be encoded in a bibliography; a reconstructed version gained close to a million views. A book publisher would eventually build an entire volume around it, titled Sutskever's List, its first chapter called "What Did Ilya See?" A small shelf of paperback biographies followed, with titles like Ilya Sutskever: Architect of Intelligence and Ilya Sutskever: The Mind That Taught Machines to Think — a strange fate for a man who does not go to events, which he doesn't.
On May 14, 2024, he announced his departure from OpenAI to work on a project that was "very personally meaningful" to him. The company named Jakub Pachocki its new chief scientist. Hours later, Jan Leike, his co-lead on Superalignment, announced his own resignation, citing an erosion of safety and trust in the company's leadership. The Superalignment team was split up and folded into other groups. Scott Aaronson, who had been recruited by both men, wrote that he was "sad," and that in two years at the company he had "never once found a smoke-filled room where they laugh at all the rubes who take the talk about 'safety' and 'alignment' seriously." It was a defense of OpenAI that read, between the lines, as an elegy for the people who had left it.
At Vox, Sigal Samuel offered the most precise answer to the meme that anyone published that spring: the real explanation "may have less to do with pessimism about technology and more to do with pessimism about humans — and one human in particular." A person with inside knowledge of the company described it as "trust collapsing bit by bit, like dominoes falling one by one."
In October 2024, the week Geoffrey Hinton won the Nobel Prize in Physics, he told reporters that he was proud his student had fired Altman. It is difficult to think of another case in which a laureate used the occasion of his prize to endorse a former student's failed boardroom coup. It said something about Hinton, who had spent his own career being right too early. It said something, too, about the intimacy of the lineage — the knock on the door in 2003, the paper handed across the desk, the old man who still saw the young one the way he had first seen him: as someone whose immediate reactions took experts years.
Sutskever, for his part, explained his departure most simply under oath. "Ultimately, I had a big new vision," he said in the October 2025 deposition. "And it felt more suitable for a new company."
One Goal and One Product
On June 19, 2024, he posted the announcement on X. The company's website, preserved in the archive in its first plain form, read in its entirety like a manifesto typed into a blank document:
"Superintelligence is within reach. Building safe superintelligence (SSI) is the most important technical problem of our time. We have started the world's first straight-shot SSI lab, with one goal and one product: a safe superintelligence. It's called Safe Superintelligence Inc."
It went on. The company would approach safety and capabilities "in tandem," advancing capabilities "as fast as possible while making sure our safety always remains ahead." Its singular focus meant "no distraction by management overhead or product cycles," its business model insulated from "short-term commercial pressures." It would assemble "a lean, cracked team" — the last word Silicon Valley slang for extraordinarily capable — and it would have offices in Palo Alto and Tel Aviv, "where we have deep roots." Then the sentence that sounds, depending on your temperament, either serene or chilling: "This way, we can scale in peace."
He told Bloomberg's Ashlee Vance that the company's "first product will be the safe superintelligence, and it will not do anything else up until then."
His cofounders were Daniel Gross, a former head of Apple's AI efforts turned investor, and Daniel Levy, a former OpenAI researcher. The American headquarters sits at 317 University Avenue in downtown Palo Alto. In February 2025 the company leased offices at 144 Menachem Begin Road in midtown Tel Aviv, a street named for an Israeli prime minister who was born in the Russian Empire. Sutskever had come back, by a long route, to the country of his childhood, though only partway; SSI describes itself as an American company.
The money came fast and in sums that strained comparison. In September 2024, SSI announced $1 billion from Andreessen Horowitz, Sequoia Capital, DST Global, SV Angel, and NFDG, the investment partnership of Gross and Nat Friedman, at a reported $5 billion valuation. In March 2025, a round led by Greenoaks Capital raised $2 billion more and valued the company at $30 billion, later reported as $32 billion — a sixfold increase in six months, for a company with about twenty employees and no revenue. The Wall Street Journal attributed investor interest largely to Sutskever's reputation. There is no other honest way to describe it. They were buying the man who had told Krizhevsky to try ImageNet.
In April 2025, Google Cloud announced it would provide SSI with Tensor Processing Units. That spring, Meta tried to buy the company outright. Sutskever said no. So Meta hired his CEO instead: in July 2025, Daniel Gross left for Meta Superintelligence Labs, offering on his way out a benediction — "I expect miracles" — and Sutskever became chief executive of SSI. His message to staff, as reported in Israel, was four words long: "We know what to do." The company had about fifty employees by then. They were discouraged from listing SSI on their LinkedIn profiles, ostensibly to make them harder to poach. A man who does not go to events had built a company that does not exist on the professional networking site.
Then he did something unusual for him. He gave a long interview.
On November 25, 2025, Dwarkesh Patel released a ninety-minute conversation with him, his first deep public discussion since leaving OpenAI. It opened without preamble. "You know what's crazy?" Sutskever said. "That all of this is real." When Patel asked what he meant, he gestured at the whole of it: "All this AI stuff and all this Bay Area… that it's happening. Isn't it straight out of science fiction?"
What followed was, in effect, a recantation of the ethos he was credited with creating, or at least an amendment to it. The years from 2020 to 2025, he said, had been the age of scaling. "Scaling sucked out all the air in the room." Because the recipe worked so reliably — more data, more compute, predictable gains — every company did the same thing, and pre-training data was finite. Now the field was returning to an age of research, "just with bigger computers." "We got to the point where we are in a world where there are more companies than ideas," he said. "By quite a bit."
The flaw he kept returning to was generalization. "These models somehow just generalize dramatically worse than people," he said. "It's a very fundamental thing." He described the experience every programmer using an AI assistant will recognize: you ask the model to fix a bug, it says, "Oh my God, you're so right. I have a bug. Let me go fix that," and introduces a second bug; you point that out, and it tells you, "Oh my God, how could I have done it? You're so right again," and brings back the first. "How is that possible? I'm not sure." The models aced hard evaluations and stumbled in real work, partly, he suspected, because human researchers unconsciously trained them toward the benchmarks they would be judged on — reward hacking, practiced by the people rather than the machines.
A teenager learns to drive in about ten hours. The best models need oceans of data and still drive into the ditch. Sutskever's speculation about the difference turned, of all places, to feeling: "Maybe it suggests that the value function of humans is modulated by emotions in some important way that's hardcoded by evolution." The man who had spent his career reasoning from the slowness of human neurons now reasoned from the richness of human emotion. The brain was still a neural network. It just had something the machines did not, and he was trying to find out what.
The current approach, he said, would "go some distance and then peter out. It will continue to improve, but it will also not be 'it.'" He gave a timeline of five to twenty years for a system that could learn the way humans do. He said the first such systems should be built to care about sentient life. He described SSI as "squarely an age of research company." When the interview was widely summarized as declaring the scaling era dead, he posted a clarification on X, precise to the point of pedantry: "Scaling the current thing will keep leading to improvements. In particular, it won't stall. But something important will continue to be missing."
Something important will continue to be missing. It is the most Sutskever sentence in the public record: technically exact, emotionally enormous, and declining to say what the missing thing is.
— Ilya Sutskever, Dwarkesh Podcast, November 2025These models somehow just generalize dramatically worse than people. It's a very fundamental thing.
The next summer brought the trial in Oakland and, eleven weeks later, the deal. On July 27, 2026, SSI and Nvidia announced a long-term strategic partnership. Nvidia would invest — a reported $5 billion — and give SSI access to its next-generation Vera Rubin platform, increasing the company's compute "by an order of magnitude." The press release said Nvidia had entered the partnership "after obtaining rare access into the company's closely guarded research," which it described as "a new research direction to unlock a powerful and robustly aligned artificial intelligence." Jensen Huang, Nvidia's founder, said: "Ilya has pioneered fundamental breakthroughs at the foundation of modern AI, beginning with AlexNet." Sutskever said: "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so."
Worthy of scaling up. The age of research, it turned out, ends where every age of research ends, in a bigger computer. Bloomberg's billionaires index that month estimated his net worth at $13.8 billion, roughly half of it in OpenAI, the company he had tried to save by firing its leader and then saved again by asking for him back, and roughly half in SSI, the company that does one thing and will not do anything else until then.
Fourteen years earlier, the computer had been two Nvidia cards that cost $500 apiece, so short of memory that the network had to be cut in half to fit, humming in a bedroom in a house in Toronto that belonged to someone's parents. Before that, it had been a single monitor in a university office, a 24-year-old's fingers hovering over the keys, a cursor blinking, and a machine that had taught itself grammar producing, one character at a time, a sentence no one had ever written:
Other parish schools were established in 1825, but were relieved on March 3, 1850.
The parentheses in the paragraph above it closed. The subject agreed with the verb. And none of it had happened.
Part IIThe Playbook
Sutskever's career resists the usual founder template. He has not built a product in years and has said, plainly, that his current company will not build one until it builds the only one that matters. He has been spectacularly right about technology and publicly, painfully wrong about power. The principles below are drawn from both columns of that ledger. They are most useful to operators working on long-horizon problems where the core asset is a technical thesis rather than a go-to-market motion — and to anyone who must one day decide whether to act on what they have seen.
The principles
- 1.Reason from the irreducible argument, not the consensus.
- 2.Be the one who insists; find the one who can build.
- 3.Hold the unappreciated idea until the hardware arrives.
- 4.Repeat the trick that works — across every domain it fits.
- 5.Choose a simple objective and let the capabilities emerge.
- 6.Compress your worldview into a canon.
- 7.Leave when staying is riskless.
- 8.Declare the end of your own era.
- 9.Evidence is not a coalition.
- 10.Reverse fast when being right would destroy the thing.
- 11.Make the mission structural, not rhetorical.
- 12.Spend reputation like capital — and know it must eventually be repaid.
I
A Career in Twelve Dates
The decisions that compounded, and the one that nearly unwound them.
2002
Admitted to the University of Toronto's math program out of Grade 11, as a third-year undergraduate.
2003
Walks into Diane Horton's office; is sent to Geoffrey Hinton; begins work on Stochastic Neighbor Embedding.
2010
First Canadian awarded a Google PhD fellowship; demonstrates a character-level network that generates Wikipedia-like text.
2012
Persuades Alex Krizhevsky to train a CNN on ImageNet; AlexNet wins with 15.3% top-5 error.
2013
Google acquires DNNResearch; joins Google Brain.
2014
Publishes sequence-to-sequence learning with Oriol Vinyals and Quoc Le.
2015
Turns down a $6 million Google salary to cofound OpenAI.
2023
Co-leads Superalignment; votes to remove Sam Altman; publicly regrets it four days later.
2024
Leaves OpenAI in May; founds Safe Superintelligence Inc. on June 19; raises $1 billion.
2025
Rejects Meta's acquisition attempt; becomes CEO; declares a return to the age of research.
2026
Testifies in Musk v. Altman; SSI announces Nvidia partnership and planned $5 billion investment.
Principle 1
Reason from the irreducible argument, not the consensus
Sutskever's conviction about AlexNet did not rest on a leaderboard or a funding trend. It rested on an argument he called "irreducible": the human brain can classify images quickly; the brain is "just a neural network with slow neurons"; therefore a large enough artificial network, trained on enough data, "will succeed necessarily." The same structure produced the "deep learning hypothesis" of the 2014 talk — anything a human can do in a fraction of a second, a ten-layer network can do.
The power of this kind of reasoning is that it is immune to social proof. In a field he described as one of "desolation and despair," where learning features from data was "a minority position," a consensus-driven researcher would have migrated to support vector machines. An argument that runs from first principles to a necessary conclusion lets you hold a position the room rejects, because the room's opinion was never an input.
The risk, which Part I makes vivid, is that irreducible arguments about machines do not transfer cleanly to people. The same certainty that made him right about scaling made the board's case against Altman feel self-evident to him — so self-evident he did not prepare to explain it.
Tactic: Write down the shortest chain of reasoning that makes your bet necessary rather than merely plausible, and if you cannot get it under four steps, you do not yet have conviction — you have a preference.
Principle 2
Be the one who insists; find the one who can build
Hinton's summary — "Ilya thought we should do it, Alex made it work" — is a precise division of labor. Krizhevsky had the scarce skill of programming GPUs and had already written cuda-convnet. Sutskever had the thesis and the persistence. Neither alone produced AlexNet.
This is a rare and underrated founding pattern: the person whose contribution is conviction applied to someone else's capability. It is easy to undervalue because it leaves fewer fingerprints in the code. But the record is unambiguous that the decision to point Krizhevsky's tool at ImageNet, rather than at another small dataset, was the decisive act.
The pattern repeated at Google with Vinyals and Le, and arguably at OpenAI, where his role was less to write the models than to establish "the scaling ethos" under which others wrote them.
Tactic: Identify the one person on your team whose tool is more powerful than the problem they are currently pointing it at, and make the case — in specifics — for the larger target.
Principle 3
Hold the unappreciated idea until the hardware arrives
AlexNet and LeNet, the textbooks note, "share essentially the same design." The gap between 1989 and 2012 was not an idea; it was twenty years of cheaper data and cheaper compute. Fei-Fei Li's ImageNet and Nvidia's $500 consumer cards were the missing pieces. Sutskever spent nine years in Hinton's orbit working on ideas that were, in his own words, "highly unappreciated by most scientists, yet turned out to be utterly correct."
The lesson is not patience as a virtue. It is that some correct ideas are gated by external inputs, and the job during the waiting period is to keep the idea in working condition — to publish, refine, and stay close enough to the inputs to notice the moment they mature. Hinton's 2011 question to colleagues and Malik's suggestion of ImageNet are the moment a waiting idea met its dataset.
Tactic: For each long-held contrarian bet, name the specific external input — a dataset, a cost curve, a regulation — whose arrival would make it work, and assign someone to watch it.
Principle 4
Repeat the trick that works — across every domain it fits
"It's an autoregressive model trained on text. It's a large neural network. It's a large dataset. That's it." Sutskever's description of the 2014 sequence-to-sequence paper could describe his 2010 Wikipedia generator, AlexNet's logic transferred from images to text, and the GPT series. A single recipe — big network, big data, simple objective — carried across vision, translation, and language.
Most researchers and founders feel pressure to be novel each time. Sutskever's record suggests the opposite discipline: when a method works, the highest-leverage move is often to find the next domain where it has not yet been tried, rather than to invent a new method. Fortune's summary of his Google years captures it exactly: he "helped show that the same kind of pattern recognition displayed by AlexNet for images could also work for words and sentences."
Tactic: Take your single most successful method and list three adjacent domains where no one has applied it at full scale; start with the one where the incumbents are most certain it won't work.
Principle 5
Choose a simple objective and let the capabilities emerge
The 2010 network had one job: predict the next character. From that it "discovered that words exist, and it discovered grammar." GPT-3, as Spencer Greenberg put it to him, "was trained to do a 'single' task," yet seemed to understand many things. The insight is that a well-chosen, dumb-sounding objective can elicit capabilities you never specified — and that specifying them in advance would have constrained the result.
This is what Sutskever meant when he said, at 24, "if you don't [know your destination], there is more of a chance of stumbling upon something really interesting." It is not an argument against goals. It is an argument for goals that are simple enough to optimize relentlessly and general enough to leave room for surprise.
His 2025 critique is the mirror image: reinforcement learning, by forcing researchers to choose specific environments, invites them to optimize toward the evaluations they will be judged on, producing models that are "too single-minded."
Tactic: Replace your team's five specific success metrics with one general objective that, if pushed hard enough, would produce all five as side effects — and watch for what else it produces.
Principle 6
Compress your worldview into a canon
Asked by John Carmack how to learn AI, Sutskever did not offer a course or a mentorship. He offered a list of roughly thirty to forty papers and a claim: "If you really learn all of these, you'll know 90% of what matters today." The list's heavy weighting toward information theory — Kolmogorov complexity, Minimum Description Length, Aaronson's coffee automaton — encoded not just facts but a way of thinking: prediction as compression, compression as understanding.
A canon is leadership at scale. It lets a person who gives few interviews and does not go to events shape the thinking of people he will never meet. The list reportedly became part of onboarding at OpenAI and, after it leaked, a cultural shibboleth across the field.
Tactic: Write down the twenty documents — internal or external — that you believe contain 90 percent of what matters in your business, and give the list to your next three hires on day one.
⇄
The Sutskever Inversions
Where his choices departed from standard founder practice.
| Conventional wisdom | Sutskever's approach |
|---|---|
| Build a public profile to attract capital and talent. | "Which I don't." Few interviews, no events; let the work and the reputation do the recruiting. |
| Ship early, iterate on customer feedback. | "Our first product will be the safe superintelligence, and it will not do anything else up until then." |
| Novelty is the mark of a great researcher. | Reapply one recipe — big network, big data, simple objective — across every domain it fits. |
| Defend your past positions to protect credibility. | Publicly revise them: on openness (2023), on the board vote (2023), on scaling (2025). |
| Accept acquisition offers at peak valuation. | Rebuff Meta; become CEO when the cofounder leaves. |
| Let employees advertise affiliation to build the brand. | Discourage listing SSI on LinkedIn to reduce poaching. |
Principle 7
Leave when staying is riskless
Sutskever's description of leaving Google Brain is a small masterpiece of honest accounting: "I was very happy at the Google Brain team. It was moving forward on an extremely strong trajectory — I would be taking no risk by staying and have much to gain." He left anyway, and turned down $6 million to do it.
The counterintuitive point is that "no risk by staying" was itself the signal. When a position offers only upside within a known trajectory, its ceiling is set by that trajectory. OpenAI offered an unknown trajectory — and a mission he cared about more. He repeated the move in 2024, leaving the most valuable AI company in the world for a startup with no product, because, as he testified, "I had a big new vision. And it felt more suitable for a new company."
Tactic: When you catch yourself describing your current role as "no risk and much to gain," treat that sentence as an alarm, and spend one week seriously specifying what you would build if you left.
Principle 8
Declare the end of your own era
The man credited with OpenAI's scaling ethos stood up in November 2025 and said the age of scaling was ending: "Scaling sucked out all the air in the room," and the industry had reached "a world where there are more companies than ideas. By quite a bit." He did this while running a company that would, eight months later, take a reported $5 billion from the world's largest seller of the hardware that scaling consumes.
Founders rarely declare the obsolescence of the paradigm that made them. Sutskever's willingness to do so — and his care in clarifying that scaling "won't stall" but "something important will continue to be missing" — is a form of positioning: if everyone is running the same recipe, the only differentiated bet is on the next recipe. It is also a form of intellectual honesty that cost him nothing he valued.
Tactic: Write a short memo arguing that your company's founding insight has stopped being a differentiator, and if you cannot rebut it convincingly, start allocating research time to what replaces it.
Principle 9
Evidence is not a coalition
Sutskever spent about a year assembling a 52-page memo. He sent it through disappearing emails. He secured the votes of a six-person board. And he lost within five days, because more than 500 employees, the company's largest investor, and its most effective communicator lined up on the other side. Hao's reporting concluded that he and Murati "failed to consolidate their power and win key executives over to their cause." Nadella's verdict was "amateur city."
The board's error was a category error: it treated governance as a proof to be verified rather than a coalition to be built. It had the legal authority to remove a CEO and none of the narrative, relational, or financial support required to make the removal stick. And by explaining nothing beyond "not consistently candid," it surrendered the story to the man it had accused of controlling stories.
This is the most transferable lesson in the profile, precisely because it is a failure. Technical people often believe that being right is sufficient. In organizations, being right is the entry fee.
Tactic: Before taking any action against a powerful incumbent, list by name the five people whose reaction will determine whether it holds, and do not act until you know what each of them will hear, from whom, and when.
Principle 10
Reverse fast when being right would destroy the thing
On Monday, November 20, 2023, Sutskever signed a letter demanding the resignation of the board on which he sat, and posted: "I deeply regret my participation in the board's actions. I never intended to harm OpenAI." In 2026 he called it a "Hail Mary" and explained it simply: "I felt that had I not done this, the company would be destroyed."
The reversal was humiliating and widely mocked. It was also, by his own lights, correct: whatever he believed about Altman, he believed more strongly in what OpenAI had built, and a destroyed OpenAI — or one merged into Anthropic, which he "really did not want" — served no one's mission. Most people in his position would have held the line to protect their credibility. He spent his credibility to preserve the institution, then left it six months later on his own terms.
Tactic: For any high-stakes decision you've already made, define in advance the specific outcome that would cause you to reverse it publicly, so that when it arrives you are executing a plan rather than suffering a defeat.
Principle 11
Make the mission structural, not rhetorical
OpenAI's mission was safe, broadly beneficial AGI. Its structure — a nonprofit controlling a capped-profit subsidiary with billions from Microsoft — proved unable to hold that mission against commercial gravity, and the promised 20 percent of compute for Superalignment, by Fortune's reporting, never materialized. SSI was designed in response. Its name is its mission, its mission is its product roadmap, and its website promises "no distraction by management overhead or product cycles."
The lesson Sutskever drew is that missions survive only when the structure makes deviation costly or impossible. A company with no product cannot be pressured to ship one early. A company whose employees do not list it on LinkedIn is harder to raid. A founder who rebuffs Meta and becomes CEO himself removes the most obvious exit.
◎
The Ledger of Bets
Sutskever's major wagers, and how each has resolved so far.
| Bet | Period | Status |
|---|---|---|
| Features can be learned from data, not hand-engineered | 2003–2012 | Vindicated |
| Next-token prediction at scale yields broad capability | 2010–2022 | Vindicated |
| Open research as a safeguard | 2015–2023 | Revised |
| Removing Altman to protect the mission | November 2023 | Reversed |
| Scaling alone reaches the goal | 2020–2025 | Revised |
| A straight-shot lab with one product | 2024– | Pending |
Tactic: For each sentence in your mission statement, identify the structural mechanism — ownership, contract, product decision, hiring rule — that would make it expensive to abandon, and if none exists, build one.
Principle 12
Spend reputation like capital — and know it must eventually be repaid
SSI reached a $32 billion valuation with roughly twenty employees, no revenue, and almost no public technical output. The Wall Street Journal attributed the valuation largely to Sutskever's reputation. Nvidia's 2026 press release cited AlexNet, fourteen years old, as the credential. Investors were not buying a product; they were buying a track record of seeing things early.
This is a legitimate and powerful form of capital, and Sutskever has deployed it with discipline: no product distractions, no premature releases, a single clarifying interview, then "rare access" for one strategic partner. But reputational capital is a loan against future evidence. As one skeptical analysis put it, "Mission is not evidence of safety." The Vox reporting on OpenAI's departures is a reminder of how quickly trust in a leader can collapse "bit by bit, like dominoes." At some point, the research that is "worthy of scaling up" must be shown.
Tactic: If your company's value rests primarily on your reputation, set a private date by which you will have produced the evidence that justifies it, and treat that date with the seriousness of a debt maturity.
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
— Ilya Sutskever, University of Toronto Magazine, 2010If you know your destination, you will probably get there, but if you don't, there is more of a chance of stumbling upon something really interesting.
— Ilya Sutskever, The Robot Brains Podcast, 2021The field of AI back then when I was starting out was not a hopeful field. It was a field of desolation and despair. No one was making any progress at all.
— Ilya Sutskever, on leaving Google Brain for OpenAI, 2017It was an extremely difficult decision. I was very happy at the Google Brain team. It was moving forward on an extremely strong trajectory — I would be taking no risk by staying and have much to gain.
— Ilya Sutskever, Dwarkesh Podcast, November 2025We got to the point where we are in a world where there are more companies than ideas. By quite a bit.
— Ilya Sutskever, on SSI's partnership with Nvidia, July 2026We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so.
Maxims
- Argue from necessity. A thesis that follows irreducibly from first principles can survive a decade of consensus against it.
- Insist, then delegate. The decisive contribution is sometimes persuading the right builder to aim at a larger target.
- Wait for the inputs. Some correct ideas are gated by data and compute; keep them alive until the world catches up.
- Reuse the recipe. When a method works, the next breakthrough is often the same method in a new domain.
- Keep the objective dumb. Predict the next character, and grammar emerges; specify grammar, and you get less.
- Distrust riskless comfort. "No risk by staying" is the sentence that should make you leave.
- Retire your own paradigm. The person who built the era is best positioned to announce its end.
- Evidence does not govern. Being right without a coalition and an explanation is the fastest way to lose.
- Reverse before ruin. Credibility spent to save the institution is better spent than credibility spent to save face.
- Repay the reputation. Trust extended on a track record is a loan; the evidence is the repayment.
