Part IThe Story
Around one in the morning on a warm night in June 2018, a computer scientist sat in her pajamas in a Washington, D.C., hotel room and cut a paragraph out of her speech. Not a sentence. A full paragraph. She would have only a few minutes before the House Committee on Science, Space, and Technology, and she wanted to be certain she reached the points that mattered. That meant something she had written, and presumably believed, had to go. Then she went to sleep.
In the morning she put on boots and a black-and-navy knit dress, which, Wired noted, was a departure from her usual T-shirt and jeans. She took an Uber to the Rayburn House Office Building, just south of the Capitol. Before she went into the committee chamber she raised her phone and took a picture of the oversize wooden doors. "As a scientist," she explained, "I feel special about the committee." Then she walked into the cavernous room and sat down at the witness table.
The hearing, held on June 26, was titled "Artificial Intelligence—With Great Power Comes Great Responsibility." The comic-book epigraph was apparently unintentional. Beside her sat Timothy Persons, chief scientist of the Government Accountability Office, and Greg Brockman, cofounder and chief technology officer of a nonprofit then called OpenAI. Fei-Fei Li was the only woman at the table. She was also, as Wired put it, the only one who could lay claim to a groundbreaking accomplishment in the field. She had built ImageNet, the fifteen-million-image dataset that a few years earlier had given machines something close to sight. That put her among a handful of researchers "perhaps small enough to fit around a kitchen table" who were responsible for AI's sudden, astonishing competence.
What surprised the room was what she said. She had come to Capitol Hill to talk about the dangers of the field she loved.
One detail from that morning keeps pulling me back: the photograph of the doors. Li had spent the better part of a decade teaching computers to look at a picture and say what was in it. Door. Wood. Building. On the morning she came to warn Congress about what such machines might do, she took a picture for a reason no classifier could label. She wanted to remember being there. The gap between those two acts of seeing, the machine's and the immigrant's, is the territory she has occupied for her whole career. It is also, not incidentally, the scene she chose to open her memoir.
At the time she was on sabbatical from Stanford, serving as chief scientist of AI and machine learning at Google Cloud. Within three months she would leave Google under a cloud that her own leaked emails had helped create. Within a year she would cofound an institute at Stanford dedicated to the proposition that artificial intelligence should be "human-centered," a phrase that sounds like a slogan until you watch her try to defend it in practice. Within six years she would be running a startup valued at more than a billion dollars. And within seven she would appear in Time's Person of the Year package beside Elon Musk, the man who had once removed her from a corporate board with a single filing.
None of it was obvious in 1992, when a fifteen-year-old girl from Chengdu landed in New Jersey with her mother and, by Li's account, less than twenty dollars between them.
By the Numbers
From Parsippany to World Labs
<$20Money Li says her parents had between them on arriving in America
7 yearsTime Li ran her family's dry-cleaning shop, from age 19 into graduate school
15MLabeled images in ImageNet, across roughly 22,000 categories
48,000+Crowd workers worldwide who labeled ImageNet via Amazon Mechanical Turk
15.3%AlexNet's 2012 top-5 error on the ImageNet Challenge, more than ten points clear of the field
540M yearsTime evolution took to develop vision, the clock Li uses for spatial intelligence
$1.23BWorld Labs' total funding after its early-2026 round, at a reported $5B valuation
A Rice Field, a Bicycle, and a Wall With Nothing on It
Two parents, two temperaments. Li has described them as a binary so clean it would look contrived in a novel.
Her father loved bugs. He loved them then and, she says, he loves them still. He was "a curious mind" who "finds humor and fun in unserious things." In the Chengdu of the 1980s the family lived in apartment complexes at the edge of an expanding city, and on weekends father and daughter walked out past the last buildings into the rice fields, where there were water buffaloes, and where a girl with a puppy could spend a whole Saturday turning over leaves. A reviewer of her memoir, summarizing his character, called him the kind of parent a child might design for herself if left to her own devices. He took her to the neighboring mountains to draw. He showed, as Li tells it, "no interest in my grades or what I'm doing in class. Did I achieve anything? Did I bring back any competition awards? Nothing to do with that."
Her mother was the other strand. She could not do math. Li laughs at her for it. But she had been, as a girl, intensely academic, and the Cultural Revolution "really crushed all her dreams." Her own parents, academically trained, had ended up on the wrong side of that history. What survived the wreckage was discipline and a kind of rebellion that never fully went out. Li's mother did not hover over report cards either. She cared about something stranger and more exacting, which was focus. Don't play while you're doing homework. If it isn't done by six, you don't get to keep working on it; you live with the consequences.
Neither parent hung anything on the walls. Li is precise about this: in their home there were "zero wall hangings of anything." She has kept the habit. Today her own house and her own office display no plaques, no diplomas, no framed covers, and this is a woman who by now holds memberships in three national academies, the Queen Elizabeth Prize for Engineering, and a place in Time's 2025 Person of the Year issue. The walls are bare. Keep that in mind.
Her father left for America when she was twelve. She and her mother followed. Li says she was fifteen; some accounts, including the Carnegie Corporation's profile, say sixteen. As a teenager she had no idea why they were going. "The early teenage Fei-Fei would say, 'I have no idea,'" she told Tim Ferriss. There was a vague sense of freedom, of a different kind of education, and a hunch that she was "not a typical kid": a girl who loved physics and could recite fighter jets from the F-117 to the F-16. Only as an adult did she come to see what the move had cost. Her parents were educated people who spoke no English. In New Jersey her father repaired cameras and her mother worked a cash register. "That's pre-internet, pre-AI age," Li said. "When you are going to a different country, you might as well go to a different planet."
Many years later her mother told her the plan, such as it was:
— Li's mother, as recalled by Fei-Fei Li on The Tim Ferriss Show, December 2025I had no plan coming to New Jersey, but I think I'm going to survive. I just believe I'm going to survive and I'm going to make sure Fei-Fei survives.
Even in Parsippany, poor, her father found treasure. On weekends he went to yard sales. "Yay, let's go to yard sales," Li remembers, "and just use that as a treasure hunt almost." The rice-field instinct had moved to suburban driveways: total attention to whatever happened to be lying in front of him. Her mother, meanwhile, was going to survive.
Li's career has followed a pattern. In each phase she pairs her father's way of looking with her mother's refusal to quit. She has never said it that baldly, and she doesn't need to.
The Lunch Hour
Parsippany High School was, in Li's own unsentimental phrase, "a middle ranking New Jersey high school." It was also a heavily immigrant one, full of kids from everywhere. Li arrived knowing almost no English, and she spent her free hours working, in Chinese restaurants and elsewhere. Over the years her résumé of odd jobs would include cleaning houses, walking dogs, and working the register at a Chinese restaurant. "I didn't make a lot of friends in high school," she said. "It's a cruel time and I was very geeky."
She once overheard a teacher telling a group of boys it was unacceptable that they were falling behind the girls, because boys were "biologically smarter." She put that in her memoir. She did not editorialize about it there, and I won't either.
Then there was Bob Sabella.
Sabella taught math at Parsippany High. He worked in a building full of children from everywhere, and he had a habit, in Li's telling, of going extra miles that nobody wrote about, for students and for their families. He entered her life around the turn from sophomore to junior year, when she started AP Calculus. Their friendship began not with math but with books. Li had come to him, as the Los Angeles Times recounted from the memoir, to ask for reading recommendations, and she surprised him by naming her own favorites: Mark Twain, Charles Dickens, Ernest Hemingway. She had read them all in Chinese translation. The scene is small and funny and a little heartbreaking. A teacher discovers that the quiet ESL student has already been to Mississippi and London and Pamplona, just in another language.
Parsippany High did not offer a full Calculus BC course. So Sabella gave up his lunch hour, which was his only break, to teach it to her one-on-one. "Later as I became teacher myself, it's exhausting to teach all day long," Li said. "The fact that on top of that, he would use his lunch hours to do that extra class for me is just such a gift that I now appreciate more than I was as a teenager." His family became, in her words, her American family. When she later needed money to buy her parents a business, he was among those who lent it.
Sabella has since died. One reason she wrote the book, she says, was to celebrate a teacher like that, because "these public teachers in America are the unsung heroes of our society."
She graduated in 1995, sixth in her class. Princeton offered her a nearly full scholarship. The letter seemed so improbable that she asked two different high school advisors to read it and confirm it said what she thought it said. Two people had to check the label before she trusted it. Years later, when she was paying strangers on the internet to identify pandas, she would build that same instinct into her method.
Founder, Cashier, Exit
Li tells this part of the story as a joke, and the joke is very good.
"I was 19," she told a room full of founders at Y Combinator's AI Startup School in June 2025. "And that was out of desperation. So I had no means of supporting my family. My parents and I need to go to college to be a physics major at Princeton. So I started a dry cleaning shop and in Silicon Valley language, I fundraised. I was the founder CEO. I was also the cashier and all the other things, and I exited. So after seven years." The audience applauded. "I've never got claps for my Laundromat," she said, "but thank you."
The fundraising was real: loans from friends, and from Bob Sabella. The pressure was real too. Her mother's health was failing just as Li entered college, and the family needed, as she told Bloomberg, "to make some money to survive." As the only English speaker, Li handled "all the business." She answered phones, managed inspections, talked to customers, and did the billing. During the week she was a physics major. On most weekends she went home to New Jersey and ran the shop. When she wasn't doing that, she was washing dishes for extra money, or working at Gest Library on campus. (Tim Ferriss worked in the same library at the same time, in the attic, for six dollars an hour. They never met. "Did you change name or something?" she asked him, a quarter-century later.)
At Princeton she hid, as she puts it, in Jadwin Hall, where the physics department lives. She had come for Einstein. "All I knew," she said in a conversation with Geoffrey Hinton, "was Einstein went there." Speaking to alumni in 2024, she described her younger self as "a nerdy teenager working in a Chinese restaurant during the day and dreaming of special relativity at night." Princeton's president would later quote, from her memoir, the feeling the university libraries gave her: of being "in the presence of something transcendent." She was reading in that building the same weeks she was handling customer complaints about stains.
Physics taught her less about matter than about a certain posture toward questions. "What physics taught me was not just the math and physics," she said. "It was really this passion to ask audacious question." What is the smallest thing? How big is the universe? Where does space-time begin? By her senior year she wanted an audacious question of her own rather than borrowing Einstein's, and she had found it in intelligence. "At that time, I swear I did not know it was called AI." Her senior thesis, supervised by the electrical engineer Bradley Dickinson, proposed a computational model for an auditory illusion called Huggins dichotic pitch, in which listeners hear a tone that doesn't physically exist. Already she was interested in the distance between the signal and what the mind makes of it.
She graduated in 1999 with high honors and a Paul & Daisy Soros Fellowship for New Americans. There was also, according to several accounts of her memoir, a lucrative job offer from the business world, and her parents told her to turn it down. Apple's editors summarized it as her decision to follow "her ailing mother's advice to always live meaningfully." The family that had borrowed money to buy a dry cleaner told their only child to refuse a salary. It is hard to overstate how unusual that is.
The shop stayed with her. She ran it remotely from California until roughly the midpoint of her PhD. During graduate school her mother developed cancer and then had a stroke. "The real existential challenge," Li told CNN in 2016, "is to live up to your fullest potential, live up to your sense of responsibility and to be honest to yourself about your dreams while doing it."
The Hypothesis Nobody Wanted
Caltech in 2000 was, for her, the beginning of formal training as "a budding AI scientist." Her advisors were Pietro Perona, a computer vision researcher, and Christof Koch, a neuroscientist who studies consciousness. The pairing set the shape of everything that followed. Li's early work is less computer science than psychophysics, the measurement of what people actually perceive. Her 2002 paper in the Proceedings of the National Academy of Sciences was titled "Rapid natural scene categorization in the near absence of attention." Its successor, in 2007, asked: "What do we perceive in a glance of a real-world scene?"
The answer, roughly, was a great deal. Humans looking at a photograph for a fraction of a second can tell you it's a beach, a city street, a kitchen, whether there are people. They don't enumerate objects. They take in a world. Li has described the dream she carried out of graduate school as "storytelling of the world." When you open your eyes in a conference room, she says, "you don't just see person, person, person, chair, chair, chair. You actually see a conference room." She thought it would take a hundred years. "I told myself on my deathbed, if I can create an algorithm that can tell the story of a scene, I've succeeded."
The machines were nowhere close. They couldn't reliably tell a cat from a chair. Li's own graduate work pushed on how few examples a model could learn from. A 2004 paper tested its approach on 101 object categories, and a 2006 paper was titled, with an optimism that now reads as poignant, "One-shot learning of object categories." The field "could write papers, no problem," she said, "but we did not have a breakthrough." Researchers would write one algorithm to find dogs and another to find cats. They tried neural networks, which "didn't work," and moved on to Bayesian networks, support vector machines, "whatever it was." One problem haunted her above the others, and it was generalization. A model that cannot generalize has not learned anything.
She took a faculty job at the University of Illinois Urbana-Champaign in 2005. In 2007 Princeton called her back, "one of my happiest moment of my life." She moved into the computer science department and took up her old residential college affiliation at Forbes again. Somewhere in that transition, in the years around 2006 and 2007, she had what she calls an epiphany. She is careful, almost insistent, that it was not hers alone.
"One thing I actually don't like about the telling of scientific history," she told Ferriss, "is there's too much focus on single genius." The epiphany came partly from a psychologist named Irving Biederman, who "was not interested in AI, but he was interested in understanding minds." Biederman had estimated that human beings recognize something on the order of thirty thousand object categories, and that children learn a massive number of them early in life. A child doesn't learn to see from a few hundred carefully posed photographs. A child learns from an unbroken torrent of looking.
The field's datasets were, as she later told Bloomberg, "tiny." They were "toy" databases covering a handful of object types. Researchers tuned algorithms against them and assumed the algorithms were the bottleneck. Li's hypothesis inverted that. Perhaps the models weren't the problem. Perhaps the data was.
"Pre-ImageNet," she has said, "people did not believe in data."
She was a first-year assistant professor, the moment in an academic career when you are supposed to be most careful. She told a senior mentor what she planned, and the mentor, "really out of their good heart," told her: please don't do this. You might not get tenure. Another warned, according to Ars Technica: "I think you've taken this idea way too far." She tried to recruit collaborators. "Nobody in ML or AI wanted to even go close to this project." There was no funding.
Asked later by Stanford's HAI why she kept going, she offered one word and then corrected it into something larger:
Delusion? Maybe it's the same kind of unbelievable conviction that took my parents to America.
Her parents had boarded a plane to a planet they knew nothing about. Her version was a dataset nobody wanted.
Wine Labels, Pandas, and Forty-Eight Thousand Strangers
The ambition was to map, in her words, "the entire world's visual taxonomy." Her collaborator in the computer science department was Kai Li, a systems researcher. They share a surname and are not related, a coincidence that must have confused a great many grant officers. The skeleton was WordNet, a Princeton-built lexical database that organizes English nouns into nested hierarchies: animal, mammal, canine, dog, beagle. The plan was to hang images on every branch, at least a thousand for each concept. The students who did the work included Jia Deng, who would co-author the paper and go on to his own career. For a time the plan amounted to an enormous amount of tedium.
Li did the arithmetic on the back of an envelope. If one graduate student labeled the images, she told the New York Times in 2012, it could take decades. She tried hiring Princeton undergraduates. "As you know, Princeton undergrads are very smart," she said to Ferriss, and he finished the thought: they had a high opinion of the value of their time. "Yes," she said. "And they're expensive." The team considered shortcuts and rejected them. Human labeling was the gold standard, and the whole point was to measure machines against humans. "We were very, very stuck for very, very long."
The solution, she says, "is actually a consequence of desperation." Amazon had recently launched Mechanical Turk, a marketplace where anyone could post a small task and anyone, anywhere, could be paid a little to do it. Li logged into her Amazon account to see what this was. The first task she tried was transcribing wine labels. A photograph of a bottle would come up, and the worker would type something like 1999 Bordeaux. She saw immediately what it meant: "massive parallel processing with online global population."
The labor force that built ImageNet ultimately numbered more than forty-eight thousand people around the world. They sorted, according to Li, through billions of candidate images to distill some fifteen million high-quality ones across roughly twenty-two thousand categories. Quality control had to be designed, not assumed. Pay someone to find pandas and, as Ferriss put it, what stops them from finding a panda in every photograph? Li's answer was a set of what she calls "computational tactics." Workers had to pass upfront quizzes to prove they understood what a panda was. Then, mixed into their queue, the team seeded images whose correct answers were already known, some true pandas and some not. "In a way, we implicitly monitor the quality of the work by knowing where the gold standard answers are." Hours and hours, she says, she and her student spent on those conversations.
Every question that looks simple from a distance turned out to be a scientific one. What counts as quality in an image? High resolution? Clutter, the way real life is cluttered, or clean product shots? How do you trade size against goodness? Why objects, and not colors, or cities? "Coming with the dataset," she insists, "is a scientific hypothesis of what is the question to ask." Pick something too easy and you learn nothing. Pick something too complicated and you drown. Objects sat in between, hard but not hopeless.
The project changed how she saw ordinary things. Steven Levy, writing in Wired about her memoir, described how building ImageNet required her to look at the world the way a neural network might. Walking past a dog or a tree, she began to look past her own instant recognition and ask which visual cues would let software tell a beagle from a greyhound, or an oak from a bamboo, or an Eames chair from a Mission rocker. There is a long passage in the book about trying to collect images of every possible car model. The girl who had followed her father through rice fields, studying everything in sight, was now teaching that habit to machines.
In 2009 they published the result as a paper at CVPR, the field's main computer vision conference. It ran as what she calls "this tiny little CVPR poster." She says: "We barely made it into a poster... and no one paid attention."
That year she moved to Stanford as an assistant professor. ImageNet was finished, freely released, and mostly ignored. "I was a little desperate at that time," she told Hinton. "We open-sourced it. But even with open-source it wasn't really picking up. So my students and I thought, well, let's drive up the competition."
SuperVision
The ImageNet Large Scale Visual Recognition Challenge began in 2010. The design was simple and, in a field that was used to every lab testing on its own private data, nearly radical. Everyone trained on the same public images. Each year the organizers released a held-out test set. Anyone could enter. "We want the whole world's smartest students and researchers to work on this problem," Li said. A USA Today profile later called it the Olympics of AI. It was open, transparent, and competitive. "Science is about openness," she said. "It's about debate, it's about letting the ideas flow."
The first two years set a baseline that was, she says, "not completely random, but it wasn't that great." Error rates hovered around thirty percent.
Then, near the end of the summer of 2012, as her team was running the submissions through their servers, Li was at home late one night when a message came from a graduate student. We got a result that really, really stands out. You should take a look.
The entry came from Geoffrey Hinton's group in Toronto. They had called their team SuperVision, which Li still admires as a pun on super and supervised learning. The method was a convolutional neural network, an architecture first published in the 1980s, with a few tweaks. Its top-5 error was 15.3 percent, more than ten points ahead of the next-best entry. Nothing that large had ever happened on the benchmark. The surprise, for Li, was partly that the algorithm was old. What was new was everything around it: a dataset large enough to feed it, and, as she always emphasizes, "the first time that two GPUs were put together by Alex and his team" to train it. Data, chips, and a neural network had come together for the first time.
They presented the results that October at the ImageNet workshop at ICCV in Florence. Alex Krizhevsky came, the graduate student whose name would end up on the network. Yann LeCun came. "And now," Li says, "the world knows this moment as the ImageNet Challenge AlexNet moment." Many people date the birth of modern AI to it. Li, who rarely overclaims, says ImageNet was "one third of that, if you count the elements."
A month after Florence, in November 2012, John Markoff profiled her in the Times. The piece identified her, with newspaper precision, as "Dr. Li, 36." It mentioned that Andrew Ng and Jeff Dean at Google had tested against ImageNet that summer and gotten results nearly twice as good as previous neural networks. It counted more than three hundred scientific publications that had used or cited the dataset since 2009. And it quoted her warning that without a new generation of visual search, the web was in danger of "going dark." Facebook had two hundred billion images, she said, and people were uploading seventy-two hours of video to YouTube every minute. The internet was drowning in pictures that machines could not read.
— Fei-Fei Li, Y Combinator AI Startup School, June 16, 2025My entire career is going after problems that are just so hard, bordering delusional.
In the years that followed she got the deathbed goal early. At Stanford, two students arrived in her lab: Andrej Karpathy and, later, Justin Johnson. Around 2015 Li and Karpathy published a series of papers, among the first of their kind, that made a computer write a sentence describing an image. Storytelling of a scene. The goal she had expected to take her whole life arrived before she turned forty. "I almost felt like, what am I going to do with my life?" Karpathy's dissertation was titled Connecting Images and Natural Language. Li joked to him that they should do the reverse: take a sentence and generate an image. "And of course, he knew I was joking, and he said, ha, ha, I'm out of here. The world was just not ready."
In March 2015 she stood on the main stage at TED in Vancouver, explained the fifteen million photos, and showed captions a machine had written. The talk has been played more than three million times. That summer she and her former student Olga Russakovsky, lead author on the challenge's definitive paper, started a summer program at Stanford for ninth-grade girls called SAILORS. She had been the only woman at enough tables to know what that cost.
The field still hadn't quite accepted the name. Li remembers that in 2015 and 2016 some tech companies avoided the word "AI" because "they were not sure if AI was a dirty word." Around 2017, companies started calling themselves AI companies. That was the year she went to work for one.
What the Eyes Are For
In January 2017, on sabbatical from Stanford, where she had directed the Stanford Artificial Intelligence Laboratory since 2013, Li became vice president and chief scientist of AI and machine learning at Google Cloud. The job was about what she calls democratizing the technology: lowering the barrier for businesses and developers, building products like AutoML that let non-specialists train models of their own. For the first time she saw AI from the customer side, and the view changed her. "My whole career up until that point, AI was a personal curiosity," she said. Now she met health systems, farms, factories, energy companies. "Every business I worked with would be or was about to be changed by AI."
The story she tells most often from that period is about cucumbers. A developer in Japan was helping his mother on her cucumber farm. Sorting cucumbers by grade was a slow, fiddly step, so he used Google's tools to automate it. "Here I was in Silicon Valley, working at the biggest company, coming from one of the best universities in the world," Li said, "and on the other side of the world, one person could use AI to change a small business."
That is one story from Google. The other is about drones.
In September 2017, Google won a Pentagon contract called Project Maven, which used AI to interpret imagery captured by drone cameras: automatically identifying vehicles, buildings, and other objects in video footage. It was object recognition. The lineage ran straight back to ImageNet, to a few thousand images per category on a WordNet branch. In internal emails that were later leaked to reporters, Li expressed enthusiasm about Google Cloud's role in the contract, and warned colleagues against mentioning its AI component. Military AI, she wrote, was linked in the public mind with fears of autonomous weapons. In the spring of 2018 thousands of Google employees signed a letter protesting the company's work for what they called "the business of war." Asked by the Times about her September email, Li issued a statement:
I believe in human-centered AI to benefit people in positive and benevolent ways. It is deeply against my principles to work on any project that I think is to weaponize AI.
The statement and the email do not fully reconcile, and I don't think they are meant to. In June 2018 Google told employees it would not renew Maven. That same month, Li sat in the Rayburn building, with the photograph of the doors in her phone, and testified about the grave dangers of the field. That fall she returned to Stanford. Her detractors read the sequence as hypocrisy. Her admirers read it as a scientist learning, in real time and in public, what her field's eyes would be used for. The record supports a less satisfying reading: both things happened in the same year, in the same person.
A sharper critique followed ImageNet itself. Scholars examining the dataset found that its "person" categories, inherited from WordNet, included labels that were biased, offensive, or impossible to verify from a picture. Others pointed to the "invisible" labor of the forty-eight thousand crowd workers whose clicks had produced a field's foundation and whose names appear nowhere in it. One blogger, in a post worth reading for its fury if not its fairness, accused her of founding Big Tech's moral decline on unconsenting data and devalued work. In 2020, Li, Russakovsky, Jia Deng, and colleagues published a paper on filtering and rebalancing the people subtree of ImageNet, removing concepts with low "imageability," words that describe nothing a camera can actually see. She told MIT Technology Review that the database is not perfect. "It takes people to call out the imperfections of ImageNet and to call out fairness issues. This is why we need diverse voices."
When the Guardian asked her in 2023 whether she took responsibility for how AI was being used, she gave the answer she has refined over the years:
— Fei-Fei Li, interview with The Guardian, November 5, 2023First, to be clear, AI is "promising" nothing. It is people who are promising – or not promising. AI is a piece of software. It is made by people, deployed by people and governed by people.
She went further. "Should Maxwell take responsibility for how electricity is used because he developed a set of equations to describe it?" Then she turned the argument back on herself: "But I am a person who has a voice and I feel I have a responsibility to raise important issues."
She refuses the blame and accepts the duty. She has been arguing with herself along those lines since 2018.
Godmother, Under Protest
In 2018 Li wrote an op-ed in the New York Times calling for "human-centered" AI. In 2019 she and John Etchemendy turned it into an institution: the Stanford Institute for Human-Centered Artificial Intelligence. Etchemendy is a philosopher and a former Stanford provost, which makes him the kind of man who has signed off on more buildings than most architects. He is also someone who believes the humanities belong in the room where the engineering happens, and HAI was designed to put them there. Li ran it, she says, "as a startup for five years in a university." She adds, with the faint grin you can almost hear in the transcript: "Probably some people were not too happy."
Etchemendy also changed her book. During the pandemic a literary agent suggested Li write one, and she spent the lockdown year drafting what she later called a "very nerdy" popular science account of AI. Etchemendy read it and told her to start over. "There are many AI technologists out there who can write a popular science book," she recalls him saying. "But for all the young people, young women, people of diverse backgrounds, and immigrants out there—there isn't much of a voice they identify with."
She is, by her own description, "a pretty shy person." She rewrote it anyway. The result, The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI, published November 7, 2023, is built like a double helix. Her coming of age and AI's coming of age wind around each other. It opens in the hearing room, goes back to Chengdu, and keeps returning. Barack Obama put it on a list of recommended books on AI, and the Financial Times named it a best book of the year. Kirkus was cooler: "Li is a pioneer, but she misses an opportunity to meaningfully weigh in on a pressing current issue," faulting the technical passages as "murky and tedious" and the personal ones as "conveyed with a distance that muffles their impact." That criticism is fair to anyone who has read the book. The distance is not an accident of craft. It is the same reticence that keeps her walls bare. "I still feel uncomfortable," she told Fortune. "But I see that as my responsibility."
The nickname arrived the same way, from outside, and she accepted it reluctantly. The press had been calling Geoffrey Hinton, Yann LeCun, and Yoshua Bengio the "godfathers" of AI since their joint Turing Award in 2018, and sooner or later someone was going to name a godmother. She told the Guardian she would not call herself that, but she accepts it as a way of recognizing women's contributions to the field. When the Queen Elizabeth Prize for Engineering was announced in 2025, the laureates were Bengio, Bill Dally, Hinton, John Hopfield, Jensen Huang, LeCun, and Li. She was the only woman of seven. Again.
Meanwhile she joined boards and commissions. In May 2020 she joined the board of Twitter as an independent director. On October 27, 2022, after Elon Musk closed his purchase of the company, a filing announced that nine directors, including Bret Taylor, Parag Agrawal, Martha Lane Fox, and Fei-Fei Li, "are no longer directors of Twitter." Musk was now the sole director. Three years later, Time named "The Architects of AI" its Person of the Year for 2025. The eight people on that list were Sam Altman, Dario Amodei, Demis Hassabis, Jensen Huang, Fei-Fei Li, Elon Musk, Lisa Su, and Mark Zuckerberg. The magazine did not seat them by who had fired whom.
Her policy positions resist the factions. She served on California's Future of Work Commission and on the White House task force for a National AI Research Resource. In August 2023 she was appointed to the United Nations Secretary-General's Scientific Advisory Board. On September 14, 2023, she testified before the Senate Homeland Security and Governmental Affairs Committee on governing AI through procurement. She has repeatedly called for a "moonshot mentality" of public investment, warning that compute, data, and talent are concentrating in private labs while academia, where she was trained, can no longer afford to train a frontier model. "Where does that leave science for public good?"
But in August 2024 she published an op-ed in Fortune opposing California's SB-1047, a frontier AI safety bill backed by many people who share her worries. She argued that its liability provisions and its mandated "kill switch" would "shackle open-source development" and "deal a death knell to academic AI." She was explicit: "I am not anti-AI governance." Governor Gavin Newsom vetoed the bill at the end of September. At the AI Action Summit in Paris in February 2025, she urged policymakers to ground their rules in "science rather than science fiction." She has said she respects the existential-risk debate "intellectually," and then turns it aside. "I'm much more concerned about the potential catastrophic social risks, like disinformation, impact on democracy, job disruption, workforce disruption, bias, and privacy infringement."
Asked at a Bloomberg conference in 2024 which AI company she trusts most, she declined to rank them. "My trust is in the collective solutions we create together. The founding fathers [of the U.S.] did not put trust in a single person." This was an immigrant who had once needed two advisors to confirm her own acceptance letter, citing the Constitution as an argument against trusting any single person.
Five Hundred Forty Million Years
In November 2022, ChatGPT "blasted open the door," as Li puts it, of generative models that could plausibly pass the Turing test. For a computer vision scientist this was both thrilling and a little humbling. The headlines now belonged to language. Li was inspired, "even for someone as old as me," and she went looking for her next north star the way she usually does, by asking what evolution had done.
Human language took perhaps three to five hundred thousand years to evolve. Vision took far longer. The first trilobites developed eyes underwater some 540 million years ago, and Li, like many evolutionary biologists, credits sight with setting off the arms race that produced complex animal life. Before eyes, simple animals. After eyes, everything else. Understanding a three-dimensional world, navigating it, acting in it: "That journey took evolution 540 million years." Language, she points out, is "purely generative." There is no language in nature. "Language literally comes out of everybody's head." The physical world is different. It is three-dimensional, four with time. Every eye and every camera collapses it into two dimensions, which makes reconstructing it "mathematically ill posed." And the world isn't purely invented; it has to obey physics. She called large language models "wordsmiths in the dark; eloquent but inexperienced, knowledgeable but ungrounded."
During that first ChatGPT winter she and Justin Johnson, her former student, were having long, excited phone calls about what would come next. Johnson had done his PhD in her lab, built real-time neural style transfer, and become a professor at the University of Michigan. Diana Hu of Y Combinator later described him as having "this super system engineering mind." Early in 2023, Li had dinner with Martin Casado. Casado is a pioneer of virtual networking. He left Stanford to start his own company around the time Li arrived, built a career on infrastructure, and became a general partner at Andreessen Horowitz. He had known her for nearly two decades, and he thought AI was retracing the path of computer games, from text to 2D to immersive 3D. By the end of the meal they agreed the next step was "large world models."
Casado suggested two more founders. Ben Mildenhall had created neural radiance fields, NeRF, a technique that spawned an entire subfield of 3D reconstruction. Christoph Lassner had worked at Amazon, Meta's Reality Labs, and Epic Games, and had built a renderer called Pulsar that anticipated the technique now known as Gaussian splatting. "The luckiest thing in my life," Li said, "is no matter how old I am, I always get to work with the best young people." Like ImageNet, the team was an argument that the right question needs the right people.
She took partial leave from Stanford at the start of 2024. Reuters reported the seed round that May; the Financial Times reported in July that the company had reached a valuation above a billion dollars in four months. On September 13, 2024, World Labs came out of stealth with $230 million from Andreessen Horowitz, NEA, and Radical Ventures, where Li had become a scientific partner in 2023. The angel list included Geoffrey Hinton, which completed a circle of some kind: the man whose team had won on her dataset was now investing in her company. In December the company showed its first results, interactive 3D worlds generated from single 2D images, including Edward Hopper's Nighthawks. In November 2025 it released Marble, a product that takes a text prompt or a photo and, after a few minutes, returns a world you can walk around in.
Li talks about Marble's uses with the same delight her father brought to yard sales. A high school drama department that can't afford sets could generate a medieval French town as a backdrop. A psychiatric researcher studying obsessive-compulsive triggers could vary a scene without limit: a strawberry field in summer, then at night, then with more strawberries, at almost no cost. Robots could rehearse in a kind of flight simulator. ("We are still early," she cautions, "so the flight simulator is not complete yet.") In February 2026 World Labs announced a further billion dollars from investors including Nvidia, AMD, Autodesk, Emerson Collective, and Fidelity, bringing total funding to $1.23 billion at a reported $5 billion valuation.
She says she is happiest at what she calls ground zero. "Forget about what you have done in the past, forget about what others think of you. Just hunker down and build. That is my comfort zone." She chose, more than once, to join departments where she was the first computer vision professor, "against a lot of advice." She hires, she says, for one quality above all, which is intellectual fearlessness. As of 2025 she would not hire a software engineer who refused AI coding tools. That isn't because the tools are perfect. It's because refusing them suggests someone who has stopped learning. When a young founder at Y Combinator asked how she handles the moments of feeling like the only one in the room, she said she had learned "to not over index on that," and then offered advice that only an engineer could deliver with a straight face: "Gradient descend yourself to the optimized solution."
When Tim Ferriss asked her what she would put on a billboard in front of a billion people, she gave him a question: What is your north star?
Near the end of that conversation, almost as an afterthought, he asked why her parents had named her Fei-Fei.
Her mother was in labor in a Beijing hospital. Her father was late, characteristically. He was bicycling across the city to get there, and somewhere along the way he caught a bird. She doesn't know how. Maybe he was just distracted. He held it for a moment.
Then he let it go, and rode on to the hospital with a name. Fei. To fly.
Part IIThe Playbook
Li's career is often summarized as a single good bet that the field needed data, followed by a long, distinguished aftermath. That reading misses the method. The same moves recur across a dry cleaner, a dataset, an institute, and a startup. She picks a question everyone else considers unserious or impractical. She defines it narrowly enough to be answerable and broadly enough to matter. She builds infrastructure that turns her bet into everyone's problem. And she accepts, sometimes awkwardly and sometimes after the fact, responsibility for what that infrastructure enables. What follows are the principles beneath those moves, drawn from the record above. Some will transfer cleanly to operators and founders. Others are uncomfortable, which is a reason to take them seriously.
The principles
- 1.Bet on the neglected input, not the fashionable one.
- 2.Get the question right before you scale the answer.
- 3.Build the arena, not just the entry.
- 4.Let desperation choose your tools.
- 5.Plant gold standards inside your pipeline.
- 6.Go where you will be the first one.
- 7.Borrow your roadmap from a field that isn't yours.
- 8.Hire for intellectual fearlessness.
- 9.Refuse the blame, accept the duty.
- 10.Tell the personal story when someone else needs it.
- 11.Raise the bar instead of policing the tool.
- 12.Keep the walls bare.
Principle 1
Bet on the neglected input, not the fashionable one
In the mid-2000s the prestige in computer vision lay in algorithms: clever architectures, better optimizers, elegant Bayesian models. Data was background. It was the fixed, modest stage on which the real performers competed. Li's insight was not that data mattered. Everyone knew that. Her insight was that data was the binding constraint, and that because nobody respected it, nobody was going to fix it. "Pre-ImageNet, people did not believe in data."
The tell was in her own work. Her graduate papers had pushed toward learning from fewer and fewer examples, "one-shot" learning. She had been optimizing in the wrong direction alongside everyone else, and she recognized the pattern from the inside. The neural network that won in 2012 was from the 1980s. The algorithm was never the missing piece. Whoever supplied the missing piece would shape what the field could do.
The pattern generalizes. In most mature fields the visible competition happens on one axis while the real constraint sits quietly on another, unglamorous and ignored. The neglected input is usually tedious to build. That is exactly why it is still available to whoever is willing to build it.
Tactic: List every input your industry needs, mark the one people complain about but nobody wants to work on, and ask what would happen if it were a hundred times larger.
Principle 2
Get the question right before you scale the answer
It is tempting to read ImageNet as a brute-force story: scrape a billion images, hire the crowd, done. Li rejects that reading. "Coming with the dataset is a scientific hypothesis of what is the question to ask." A dataset of colors would have been easy and useless. A dataset of cities would have been impossibly hard. Objects, organized on WordNet's hierarchy, were hard but tractable, and grounded in Irving Biederman's estimate of how many categories humans actually use.
Inside that framing were dozens of smaller decisions that each changed the outcome. Cluttered everyday photos or clean product shots? How many examples per category? What counts as "good enough"? Most failed datasets, she argues, failed here, not in collection. The same discipline appears at World Labs, where she resists defining the company by a single application. Instead she defines a continuum "between generation and reconstruction," with gaming at one end and robotics at the other.
Scale amplifies whatever you point it at. A badly framed question at fifteen million examples is just an expensive mistake.
Tactic: Before committing resources to scale, write down the specific question your effort answers and test whether a much smaller version produces signal on that question.
Principle 3
Build the arena, not just the entry
ImageNet was released in 2009 and "no one paid attention." Open-sourcing it wasn't enough. What changed the field was the decision in 2010 to turn the dataset into a recurring public competition with a held-out test set, fixed rules, and an open invitation. In one move Li turned her private bet into a shared scoreboard that every ambitious lab in the world wanted to top. Two years later, a team in Toronto topped it by a margin that ended an era.
The lesson is about distribution as much as science. A resource that sits on a server is a contribution. A resource with a leaderboard is a market. The challenge gave ImageNet a heartbeat, an annual moment when progress was measured in public, and that cadence pulled talent and compute toward the problem far faster than any grant could have.
◎
From Poster to Paradigm
How a neglected dataset became the field's benchmark, then its foundation.
2006–07
Li conceives ImageNet; senior mentors warn it could cost her tenure.
2009
ImageNet published as a CVPR poster and open-sourced. Little attention follows.
2010
First ImageNet Large Scale Visual Recognition Challenge. Error rates near 30%.
2012
Hinton's SuperVision team (AlexNet) posts 15.3% top-5 error; results presented in Florence.
2015
Li and Karpathy publish early image-captioning work; Li's TED talk reaches a mass audience.
2017
Challenge concludes its run; Li joins Google Cloud and co-founds AI4ALL.
2024
World Labs emerges with $230M to pursue spatial intelligence.
Tactic: When your product or idea isn't spreading, design a public, repeatable measure of success that others can compete on, and host it yourself.
Principle 4
Let desperation choose your tools
Li describes two of her defining moves with the same word. The dry cleaner was "out of desperation." Mechanical Turk was "a consequence of desperation." Neither was the obvious choice for someone in her position. A Princeton physics major doesn't usually buy a dry-cleaning business, and in 2007 a respectable computer vision lab didn't usually outsource its ground truth to anonymous strangers on a year-old Amazon marketplace that mostly transcribed wine labels.
Desperation, used well, clarifies things. When the respectable options are exhausted, because undergraduates cost too much and one graduate student would need decades, attention moves to options that respectability had ruled out. The key is that Li tried the new tool herself. She logged in, did a task, and recognized what it actually was: "massive parallel processing with online global population."
This principle has a shadow side, and the critics of ImageNet's labor practices have named it. Tools adopted out of urgency carry costs that urgency doesn't stop to count. The lesson isn't to stop being desperate. It's to come back afterward and count.
Tactic: When you're stuck, spend an hour personally using the newest, least prestigious tool adjacent to your problem before you rule it out.
Principle 5
Plant gold standards inside your pipeline
The panda problem is a management problem. If you pay people to find pandas, some will find pandas everywhere. Li's solution had two layers. Workers first passed a qualifying quiz. Then images with known answers were seeded invisibly into their work queues, so quality could be monitored continuously without anyone knowing which images were the test. "We implicitly monitor the quality of the work by knowing where the gold standard answers are."
This is how you get reliable output from a system of strangers at scale, whether those strangers are crowd workers, contractors, a distributed sales team, or AI agents. Inspection after the fact is slow and adversarial. Embedded ground truth is cheap and constant, and it changes behavior because workers know it exists even when they can't see it.
There is a personal echo here that may or may not be coincidence. A teenager who asked two advisors to verify her Princeton acceptance letter grew into a scientist who built verification into everything.
Tactic: Seed a small percentage of known-answer cases into any high-volume process you depend on, and track accuracy on those cases as a standing metric.
Principle 6
Go where you will be the first one
Standard advice to a young professor is to join a department with a strong community and senior mentors in your area. Li did the opposite more than once. She chose departments where she was the first computer vision professor, "against a lot of advice." "Of course I would love to have senior mentors," she said, "but if they're not there, I still have to trailblaze my way." She later ran HAI "as a startup" inside a university, and in her late forties she left the comfort of tenure to found a company.
Being first means having no one to imitate, which forces you to form your own judgment. It also means no one is positioned to tell you your idea is too big. The respected mentor who told her not to build ImageNet was being kind, and acting on that kindness would have been the mistake of her career. Li's phrase for the feeling is "standing on ground zero," and she calls it, without irony, her comfort zone.
Tactic: When choosing between two roles, give real weight to the one where you'd be the first person doing your kind of work, especially early in your career.
Principle 7
Borrow your roadmap from a field that isn't yours
Li's two biggest bets both came from outside computer science. ImageNet came from Biederman's psychology of object recognition and the observation that children learn from torrents of examples. World Labs came from evolutionary biology: the 540 million years since trilobites first saw, against the few hundred thousand years of language. Her doctoral training sat between electrical engineering and neuroscience, and her early papers measured what people see in a glance.
She is insistent that "science is a lineage and science is actually a non-linear lineage," and that the lone-genius story distorts how discoveries happen. The practical version of that humility is a reading habit. When she needs a new north star, she says, "I ask myself what evolution has done or what brain development has done." Mature fields have already solved, or at least characterized, versions of the problem you are stuck on. They just used different words.
Tactic: For your hardest current problem, find the field that has studied its closest natural or historical analogue, and read one foundational paper from it this month.
Principle 8
Hire for intellectual fearlessness
Li has advised an unusual concentration of people who went on to shape the field, including Andrej Karpathy, Olga Russakovsky, Jia Deng, Justin Johnson, and Jim Fan. She says they differed in almost every way. Some were "pure scientists," some industrial leaders, some "the greatest disseminator of AI knowledge." The common thread was "intellectual fearlessness": the willingness to take on something hard, go all in, and solve it "in however way you want." It is now, explicitly, her hiring criterion at World Labs.
The 2025 rule about AI coding tools is the same filter. She is not asking whether engineers like the tools. She is checking whether they have stopped learning. "The degree they have matters less to us now. It's more about what have you learned?" For the founding team itself she chose people who had each created a subfield, not people with the most impressive résumés: NeRF, Pulsar, real-time style transfer.
Tactic: In interviews, ask candidates to describe a problem they took on that people told them was too hard, and probe whether they would choose it again.
Principle 9
Refuse the blame, accept the duty
Li's position on responsibility is precise and, to some, maddening. She refuses moral ownership of every downstream use of her work. "Should Maxwell take responsibility for how electricity is used?" In the same breath she claims an obligation to speak, to build institutions, and to testify. Project Maven tested that position under pressure. Her leaked enthusiasm and her public statement do not fully align, and she returned to Stanford months later to build the infrastructure of "human-centered AI."
The useful lesson for operators is that the two halves have to come together. Refusing blame without accepting duty is just deflection. Accepting blame for everything is paralysis. What Li models, imperfectly, is a third position: separate the tool from the hand, then put your own hand to work on governance, education, and public argument. Even her SB-1047 op-ed opened with "I am not anti-AI governance," and then went on to argue against a specific governance proposal.
⚖
Conventional Wisdom vs. Li's Approach
Where her choices diverged from the expected path, and how each turned out.
| Conventional wisdom | Li's move | Outcome |
|---|---|---|
| Improve the algorithm; data is fixed. | Build 15M labeled images. | Scale |
| Join a department with senior mentors. | Be the first vision professor. | Try |
| Keep the benchmark in-house. | Open-source it and run a public challenge. | Scale |
| Pay experts for ground truth. | Crowdsource via Mechanical Turk. | Risk |
| Stay quiet on industry contracts. | Maven emails, then public principle statement. | Risk |
| Follow the language-model wave. | Bet on 3D world models. | Scan |
Tactic: For anything you build, write down both what you are not responsible for and the specific actions you will take anyway to shape how it is used.
Principle 10
Tell the personal story when someone else needs it
Li drafted a "very nerdy" book, and John Etchemendy told her to start over. His argument wasn't that her story would sell. It was that young immigrants, women, and people of diverse backgrounds had almost no voice in AI they could identify with. She calls herself shy and says the rewrite still makes her uncomfortable, "but I see that as my responsibility."
That framing matters. Personal narrative offered as self-promotion is noise. Offered as evidence that a particular kind of person can do a particular kind of work, it is infrastructure, the same kind she built with SAILORS and AI4ALL. The Kirkus critique, that her personal passages are "conveyed with a distance," is fair, and it may be the cost of telling a story you feel obligated to tell rather than one you want to. The book still reached the readers it was meant for. Princeton made it the pre-read for the class of 2028.
Tactic: If you find yourself consistently being the only one of something in the room, decide which specific audience would benefit from hearing how you got there, and tell it to them in plain terms.
Principle 11
Raise the bar instead of policing the tool
Li tells a story about a high school English teacher who, on the first day of class, ran an essay prompt through the best available AI, graded the result in front of the students, and gave it a B-minus. Use the tool all you want, the teacher said. That is the floor. To get an A, you have to do the thinking on top of it. "If the school evaluation is structured in a way that whatever AI gives and whatever the student gives is the same," Li said, "there's something wrong with the structure of the evaluation."
This is the same logic as her hiring rule, reversed. You don't win by banning a powerful tool or pretending it doesn't exist. You win by redefining excellence to start where the tool stops. Benchmarks, which she has built her career on, are exactly this: a published bar that shows everyone where "good" now begins.
Tactic: Run your team's typical deliverable through the best available AI tool, share the output as the baseline, and define excellent work explicitly as what exceeds it.
Principle 12
Keep the walls bare
In her parents' home there were "zero wall hangings of anything." In her own house and office there are still none, despite the national academies, the Queen Elizabeth Prize, and the Time cover. Her parents never asked whether she had won anything. Li treats the habit as an inheritance, and it shows up in her working method. At Y Combinator she told the audience: "Forget about what you have done in the past, forget about what others think of you."
Displayed achievements quietly anchor identity to past wins, and past wins make the next audacious question more frightening, because now there is something to lose. A bare wall is a standing reminder that the current problem is the only one that counts. It is also why a professor with a career's worth of honors could walk into a startup at nearly fifty and describe it as ground zero.
Tactic: Take down or put away the visible markers of past success in your workspace for ninety days and notice what problems you become willing to take on.
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
— Fei-Fei Li, interview with Stanford HAI, 2023Delusion? Maybe it's the same kind of unbelievable conviction that took my parents to America.
— Fei-Fei Li, Y Combinator AI Startup School, June 16, 2025I love the feeling of ground zero, like standing on ground zero. Forget about what you have done in the past, forget about what others think of you. Just hunker down and build. That is my comfort zone and I just love that.
— Fei-Fei Li, interview with The Guardian, November 5, 2023Should Maxwell take responsibility for how electricity is used because he developed a set of equations to describe it? But I am a person who has a voice and I feel I have a responsibility to raise important issues.
— Fei-Fei Li, Bloomberg Technology Summit, May 2024My trust is in the collective solutions we create together. The founding fathers [of the U.S.] did not put trust in a single person, and so my hope is not in a single A.I., it is in people.
— Fei-Fei Li, to Bloomberg, 2025Science is a nonlinear journey. Nobody has all the solutions. You have to go through such a challenge to find an answer.
Maxims
- The bottleneck is rarely where the prestige is. Look for the input everyone needs and nobody respects, then build it at a scale that embarrasses the consensus.
- A dataset is a hypothesis. Whatever you choose to collect, measure, or benchmark is a claim about which question matters.
- Make your bet everyone's problem. An open resource with a public scoreboard spreads faster than any argument for it.
- Desperation is a search strategy. When the respectable options run out, try the unrespectable tool yourself before dismissing it, and count its costs afterward.
- Trust is engineered, not assumed. Seed known answers into every pipeline you rely on.
- Being first is a feature. With no one to imitate and no one positioned to say "too big," your own judgment is all you have.
- Read outside your field for your next north star. Evolution, psychology, and history have usually studied your problem under another name.
- Hire the fearless, not the credentialed. The willingness to take on something hard and go all in predicts more than a degree.
- Decline the blame, keep the duty. You don't own every use of what you build, but you owe a voice in how it is governed.
- Keep the walls bare. Past achievements make the next audacious question harder to ask.
