Fei-Fei Li

Fei-Fei Li is Stanford’s inaugural Sequoia Professor of Computer Science and cofounder and CEO of World Labs. With collaborators, she built ImageNet, a large, hierarchically organized image dataset, and helped establish a common benchmark that made advances in visual recognition measurable. Her subsequent work includes describing individual regions of images, reassessing the categories used to classify people, and pursuing spatial intelligence: models intended to represent environments and support creation, prediction, and action within them.[1][5][6][10][9][11]

Before World Labs, Li’s academic work centered on visual recognition and the data needed to train and compare recognition systems. That earlier work supplies context for her current agenda.[5][6][11]

Li immigrated from China during high school, earned a physics BA at Princeton in 1999, and completed an electrical engineering PhD at Caltech in 2005. She joined Princeton’s faculty in 2007 and moved to Stanford in 2009. At a 2024 Princeton assembly, she recalled combining university study with work in her family’s dry-cleaning business and learning to seek help from mentors and fellow students. WIRED’s profile also describes math teacher Bob Sabella’s support, including teaching her advanced calculus during lunch breaks.[1][19][8]

According to WIRED’s 2018 profile, Li questioned whether computer vision’s bottleneck was the data rather than another narrowly targeted model. She wanted a collection spanning general and specific concepts, from mammals to star-nosed moles. The 2009 ImageNet paper, by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, reported 3.2 million images across 5,247 categories and 12 subtrees.[8][5]

ImageNet organized images around WordNet synsets: sets of words referring to a particular concept, linked in a semantic hierarchy. This disambiguated labels—a river bank and a financial institution could be separate concepts—and gave annotators a definition against which to judge images. The team expanded search queries with synonyms, contextual terms, and translations into Chinese, Spanish, Dutch, and Italian, then used human verification to filter the candidates.[5]

Building the collection required a change in annotation strategy. WIRED reports that paid Princeton student annotation progressed slowly until a student introduced Li to Amazon Mechanical Turk. Known-label images helped screen unreliable workers. The original paper describes a separate safeguard for ambiguous categories: multiple workers independently labeled each image, with category-specific agreement thresholds. A Burmese cat required more consensus than a generic cat. Confidence tables estimated how much agreement each category needed, allowing the team to scale annotation without treating every judgment as equally dependable.[8][5]

The ImageNet Large Scale Visual Recognition Challenge began in 2010, following PASCAL VOC’s precedent for standardized annual evaluation. Participants trained on shared data and tested their systems on previously unseen images. In 2012, SuperVision won classification and localization with the deep convolutional network later known as AlexNet, authored by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. Li and her collaborators’ contribution was the dataset and benchmark. The challenge paper records that most 2013 entries used deep convolutional networks after SuperVision’s success.[6][7]

WIRED reports that Li, then on maternity leave, bought a last-minute ticket and took an overnight flight to Florence to attend the 2012 award ceremony after recognizing the result’s significance.[8]

Li directed the Stanford Artificial Intelligence Laboratory from 2013 to 2018. DenseCap, presented at CVPR 2016, extended visual description beyond a single image label or whole-image caption. Justin Johnson and Andrej Karpathy received equal-contribution credit alongside Li. Their system jointly located salient image regions and described each in natural language, combining a convolutional network, a dense localization layer, and a recurrent language model. It trained end-to-end without external region proposals.[1][10]

DenseCap also supported text-based region search. For a query such as “head of a giraffe,” it scored the probability that each detected region would generate that description, rather than merely matching words against previously generated captions.[10]

Li mentored Olga Russakovsky, who told WIRED that Li helped develop both her research skills and her self-confidence. When Russakovsky proposed a summer program to introduce girls to AI, Li joined the effort. The program subsequently expanded through AI4ALL to reach underrepresented young people, including girls, people of color, and students from economically disadvantaged backgrounds. Li connected participation to the systems being built, warning that narrow representation among AI’s creators could shape their applications and blind spots.[8]

Her Stanford healthcare collaborations included work with physician Arnold Milstein on camera-based hand-hygiene monitoring intended to alert hospital workers when they missed proper washing. At Princeton’s 2024 assembly, Li described ambient intelligence as extending caregivers’ awareness and linked her understanding of healthcare to caring for her mother. She emphasized that healthcare’s mission remained humans taking care of humans.[8][19]

Li held Google vice-president and Google Cloud AI/ML chief-scientist roles in 2017–2018. WIRED reports that she sought industry experience and contact with customers deploying AI, hoping to broaden her interdisciplinary research. During the 2018 Project Maven controversy, her division administered the Defense Department project, although the profile reports that she was not directly involved in the deal. Leaked emails nevertheless made her a public face of the controversy.[1][8]

Li participated with other Google scientists and executives in developing the company’s public AI principles and advocated bringing philosopher Shannon Vallor into Google Cloud as a consulting ethicist. Of contributing to the principles, she told WIRED: “It was so important for me personally to be involved, to contribute.” She said she returned to Stanford to preserve her academic position as her sabbatical ended.[8]

In a December 2019 preprint, later published in 2020, Kaiyu Yang, Klint Qinami, Li, Jia Deng, and Olga Russakovsky examined three problems in ImageNet’s broader people subtree: inappropriate vocabulary, concepts that could not reliably be inferred from appearance, and demographic imbalance. The study identified 1,593 potentially offensive categories among 2,832 person categories and rated only 158 of the remaining 1,239 as imageable enough for its recommended visual-recognition subset. This reassessment concerned the broader dataset; the three person categories in the commonly used 1,000-category challenge subset were unaffected by the proposed filtering.[9]

The study exposed a limit of annotator consensus: workers could agree on stereotyped search results even when a concept—such as being a philanthropist—could not be established from a photograph. The authors proposed removing inappropriate labels, filtering nonvisual categories, and offering user-controlled demographic balancing. They distinguished perceived gender from gender identity and proposed a balancing interface instead of directly releasing sensitive image-level demographic annotations. These were proposed repairs, rather than a record of completed implementation.[9]

Li became a founding co-director of Stanford HAI in 2019. Stanford’s 2026 leadership transition retained her founding-director and senior-fellow titles while assigning her two advisory roles: co-chair of HAI’s advisory council with John Hennessy, and Special Advisor on AI to university president Jonathan Levin. James Landay leads the institute as Denning Director.[1][2][3]ⁿ¹

Li dates World Labs’ founding to early 2024, with Justin Johnson, Ben Mildenhall, and Christoph Lassner as fellow founders. In her November 2025 essay “From Words to Worlds”, she argues that spatial intelligence must connect perception, imagination, and action. Parking, catching keys, and pouring coffee illustrate her point: each requires understanding geometry and anticipating movement.[11][4]

Li’s proposed world models have three essential capabilities: generate consistent worlds, accept multimodal inputs, and produce next states conditioned on actions. Her research agenda includes training objectives that reflect geometry and physics, extracting spatial information from images and video, supplementing those signals with synthetic and sensor data, and developing spatial representations and memory. She separates near-term creative tools from mid-term robotics ambitions and longer-term scientific applications. Her stated principle is: “AI must augment human capability, not replace it.”[11]

At World Labs, Li’s spatial-intelligence agenda took a concrete form in Marble, which offered creators explorable 3D environments from image or text prompts in its September 2025 limited-access beta. The company subsequently made environment generation programmable through the World API; its Atlas model, announced in September 2026 in select-partner early access, was intended to power future Marble versions.[11][13][14][18]

The company’s July 2026 acquisition of SceniX also extended that agenda toward robotic learning. The SceniX team’s real-to-sim-to-real engine reconstructs physical tasks as interactive simulations, then varies conditions such as clutter and physical properties for robot-policy training and evaluation. World Labs reports that policies trained entirely in simulation transferred to physical robots; reconstructing the tasks still involves real-world capture and demonstrations.[16][17]