Fei-Fei Li

In the [YOU] on AI Field Guide

The cycle that [YOU] on AI opens describes the present as a moment when the interface between human intention and machine capability collapsed from formality to conversation—when a transformation arrived faster than anyone, including the people closest to it, anticipated. Li belongs to the cycle not only as one of the architects of that transformation but as one of its clearest-eyed witnesses. She has been candid about how much the speed of the last decade surprised her. She has been equally candid about what the surprise does not license: either triumphalism or despair.

Her concept of human-centered AI is the cycle's closest institutional analogue to the orange pill itself. The orange pill is the refusal of both naive optimism and paralytic fear—the decision to see the machine clearly and to take seriously both what it can do and what it costs. Human-centered AI is the engineering expression of that refusal: the insistence that the questions of who the technology serves, whose faces are in the data, and what the system is optimized for are not afterthoughts to be handled once the technical work is done but are constitutive of the technical work itself. Augmentation over replacement is a design choice, not a destiny, and Li has built institutions to make that choice visible and repeatable.

She occupies a position in the cycle's gallery that no other thinker quite fills: the builder who stayed inside the question of consequence. Where some architects of the AI transition moved entirely into capability research and others moved entirely into criticism, Li has maintained the double commitment that the cycle argues is actually required—to understand what the machine does well enough to build with it, and to understand what it costs well enough to constrain it. Her refusal to choose between the engineer's role and the humanist's role is not a compromise. It is the thesis.

Origin

Born in Beijing in 1976 and raised in Chengdu, Li immigrated to Parsippany, New Jersey at sixteen, entering an American high school in a language she barely spoke. Her family was not comfortable; when her mother's health declined as Li reached college age, they opened a dry-cleaning shop and Li ran it—answering phones, handling billing, managing the frictions of a small business on weekends while solving physics problem sets during the week at Princeton. She has described this doubled life not as a hardship that cost her but as formative to how she understands science: as a practice of resilience and orientation, of fixing a North Star and steering toward it through fog without any guarantee of arrival.

The North Star she fixed on was vision: the conviction, formed early, that understanding how seeing works would be a key to understanding intelligence itself. After Princeton she earned her doctorate at Caltech and joined Stanford, where she began assembling the argument that would produce ImageNet. The field consensus held that the bottleneck in computer vision was algorithmic; better mathematics would produce better object recognition. Li's dissent came from studying how children learn. A toddler can be shown one cat and thereafter recognize cats everywhere—fat, thin, shadowed, from behind—because the child has been flooded with an unimaginable quantity of visual data since birth, not because the child has a cleverer algorithm. She drew on Irving Biederman's estimate that humans can distinguish tens of thousands of object categories and set out to build a dataset that spanned the whole range, borrowing the hierarchical structure from WordNet and the labor from Amazon Mechanical Turk. The result, after years of effort, was millions of images sorted into thousands of categories, each verified by human eyes.

The 2012 ImageNet Challenge result—AlexNet's error rate so far below the competition that it initially looked like a mistake—proved the principle that had driven the project: capability scales with data. Li did not build AlexNet; she is careful to credit Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton for the architecture and the training. What she built was the condition of possibility. The subsequent decade—deep learning sweeping through vision, speech, language, domain after domain—is, in this precise sense, the consequence of the principle Li proved in the teeth of a field that thought the principle was wrong.

Key Ideas

The data bottleneck. Li's foundational insight is that capability scales with data, not only with architectural cleverness. The principle sounds simple and was heretical in context: it meant that a workable model fed an enormous diet could outperform a brilliant model fed a thin one. The deep-learning era has confirmed this at every scale, and the large language models that now reshape the economy are, in this precise lineage, ImageNet's grandchildren—workable architectures fed an almost incomprehensible quantity of human-generated data. But the same insight carries its shadow: a model that learns from data that contains everything also learns the data's gaps, skews, and silent assumptions about whose faces are worth labeling. The promise and the problem come from the same root.

Human-centered AI. Li's concept rests on three commitments that function as corrections to three observed failures. First, AI development must be guided by concern for human impact from the beginning, not as an afterthought. Second, AI should augment human beings rather than replace them—a commitment she holds with unusual force, arguing that the word "replace" should itself be replaced in the field's self-description. Third, AI should be inspired by human intelligence: the path forward runs through deeper understanding of how human cognition actually works, not away from it into pure engineering abstraction. The Stanford Institute for Human-Centered AI is this conviction made institutional, deliberately bringing computer scientists into sustained contact with physicians, philosophers, legal scholars, and social scientists on the premise that the human questions are not a supplement to the technical work but part of it.

The faces that were missing. The systematic absence of certain faces from training data—faces darker, more female, less Western than the populations most photographed and labeled—is not, for Li, a technical bug to be patched. It is a structural failure to see the humanity the technology claims to serve. The AI4ALL nonprofit she co-founded with Olga Russakovsky and others embodies her diagnosis: that the exclusions in the technology and the exclusions in the profession are the same problem viewed from two angles, and that widening who builds AI is not a side project to the technical work but a part of it. A more diverse field will build less biased systems not because diversity is a virtue in the abstract but because a wider range of people sees a wider range of gaps.

Spatial intelligence as the deeper frontier. Li has argued that spatial intelligence—the capacity to perceive, understand, and act within three-dimensional physical space—is more foundational to intelligence than language. Language models are, in her memorable phrase, wordsmiths in the dark: eloquent but inexperienced, knowledgeable but ungrounded, having ingested enormous quantities of human text without ever closing the perception-action loop that, in living creatures, is the root of intelligence. World Labs, the company she co-founded in 2024, pursues world models—generative systems that can produce and reason about consistent three-dimensional worlds. The wager is that the systems now dazzling us with words will come to look like a brilliant but partial achievement, resting on a missing foundation that is spatial, embodied, and older than language itself.