Who is Fei-Fei Li?
Fei-Fei Li is an artificial intelligence researcher, Stanford computer science professor and World Labs cofounder. Her work spans computer vision, robotic learning and spatial intelligence. She directed the Stanford AI Lab from 2013 to 2018. [1]
Her background includes immigrating from China to the United States as a teenager. She received a Paul & Daisy Soros Fellowship for graduate study. In recounting that experience, she has connected her interest in diversity with the need for technology developers to represent the society that uses their work. [8]
Education and early university appointments
She graduated in physics from Princeton in 1999 and earned an electrical engineering doctorate at Caltech in 2005. [1]
At Caltech she worked with computer vision specialist Pietro Perona and neuroscientist Christof Koch. The opportunity to combine computational models with the study of human perception was central to her choice. This training placed her questions about visual recognition at the intersection of engineering and brain science. [2]
She taught at the University of Illinois Urbana-Champaign in 2005–2006 and Princeton in 2007–2009, joining Stanford in 2009. These appointments preceded her leadership responsibilities and place ImageNet’s early development within her university research career. [9]
Learning object recognition from few examples
With Rob Fergus and Pietro Perona, she presented a Bayesian method for learning visual categories from few examples in 2004. They evaluated it on 101 object categories. It reused previously learned information and represented shape and appearance through sets of features, investigating how to reduce the examples required for a new category. [10]
ImageNet and comparing vision systems
ImageNet was presented at CVPR 2009 in a paper by Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li and Li Fei-Fei. It organized millions of images using WordNet’s semantic hierarchy. The contribution involved building and cleaning labelled data at scale, alongside defining how visual concepts were grouped. [11]
The ImageNet Large Scale Visual Recognition Challenge began in 2010, providing a common reference for object classification and detection. Its methodological analysis described annotation collection, result comparisons and the difficulty of comparing human and computational performance. A shared test made progress across teams and years observable. [12]
Li also participated in critical reassessment of the data. A 2019 paper, accepted at FAT* 2020, examined problematic person categories, concepts difficult to illustrate and imbalances in representation. The research showed how vocabulary and image distributions can introduce problems before a model is trained. [13]
From recognizing objects to describing scenes
With Andrej Karpathy, she investigated generating descriptions of images and specific regions. Initially released in 2014 and revised in 2015, the work aligned visual and linguistic representations through neural networks. It went beyond assigning categories to producing sentences related to visible elements and actions in a scene. [15]
For Visual Genome, presented in 2016 with Ranjay Krishna and other collaborators, the team annotated objects, attributes and relationships in over 100,000 images. The dataset enabled questions about who does what and how elements relate. Its approach expanded the unit of analysis from individual objects to scene structure. [14]
AI4ALL and artificial intelligence education
In 2015 Li, Olga Russakovsky and Rick Sommer established a Stanford summer programme introducing high-school girls to AI. This preceded AI4ALL’s creation in 2017. The organization subsequently expanded towards university students and research opportunities, with attention to historically underrepresented groups. [5]
Li also teaches Stanford’s CS231n, Deep Learning for Computer Vision. Its spring 2026 edition combined neural network architectures, implementation assignments and a final project. This teaching connects mathematical foundations with practical decisions about training, fine-tuning and evaluating visual systems. [16]
Google Cloud and corporate roles
During her Google Cloud period, Li and Jia Li introduced Cloud AutoML in January 2018. The initiative aimed to make building custom models easier for organizations with limited machine learning expertise. Its announcement described techniques such as transfer learning as a way to lower technical barriers to entry. [17]
In September 2018 Google announced her return to Stanford and transition into an AI and machine learning advisory role. Andrew Moore would lead Google Cloud AI. This was a change of responsibilities, placing her chief scientist position in that historical period. [18]
In May 2020 Twitter appointed her an independent director. The company linked the appointment to her engineering and AI expertise. This corporate governance position was distinct from directing the everyday development of every platform system. [27]
Stanford HAI and human-centered AI
Stanford launched HAI in March 2019 with Li and philosopher John Etchemendy as codirectors. The institute brought together research, education and analysis of AI’s social consequences. Its approach included disciplines beyond computing to investigate how technology affects people and how to guide its development. [19]
In May 2026 Stanford announced the integration of HAI and its data science initiative under James Landay’s leadership. Li moved into a university-wide AI advisory role and cochairing the advisory council with John Hennessy. The reorganization expanded coordination across academic disciplines. [3]
World Labs, Marble and world models
In 2024 she cofounded World Labs with Justin Johnson, Ben Mildenhall and Christoph Lassner. The company focused on spatial intelligence: models that represent, generate and interact with three-dimensional environments. This continues her visual perception research while adding the need to preserve spatial relationships and produce environments that can be explored. [20]
Marble became publicly available in November 2025. According to World Labs’ presentation, it creates, edits and exports 3D worlds from text, images, video or initial 3D layouts. The model fills in details absent from its inputs, so a generated scene does not automatically constitute an exact documentary reconstruction. [21]
In September 2026 the company introduced Atlas, a multimodal model using a shared spatial context to generate images, video and geometry. The team demonstrated reconstruction and simulation tasks and explained that unseen areas are filled with plausible content. Its announcement distinguished early-access research from future products based on it. [22]
On 28 September 2026 AMD announced an agreement to acquire World Labs. It expected completion before year-end, subject to approvals and other conditions, after which Li would become executive vice president and chief scientist. The announcement documents the agreement and plans, rather than itself establishing final completion. [4]
Memoir and recognition
In 2023 she published The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI. The book combines memoir and scientific explanation, from her immigrant experience to AI’s development. In the publisher’s excerpt, Li connects technological capabilities with the responsibility of those deciding what to build them for. [23]
She belongs to the National Academy of Engineering, National Academy of Medicine and American Academy of Arts and Sciences. She received Intel’s lifetime achievement recognition in 2023 and Princeton’s Woodrow Wilson Award in 2024, associated with service to society. [24]
She shared the 2024 VinFuture Grand Prize with Yoshua Bengio, Geoffrey Hinton, Jensen Huang and Yann LeCun. [25]
In 2025 she jointly received the Queen Elizabeth Prize for Engineering for the foundations of modern machine learning. Stanford highlighted ImageNet’s role as a shared resource and benchmark. The award places her contribution to data alongside the advances in algorithms and computing power that enabled this period of AI. [26]
Hand hygiene and observation of healthcare practices
Fei-Fei Li coauthored a JAMIA study that installed depth sensors in a hospital unit to detect use of hand-hygiene dispensers. The algorithm was checked against human observation and image annotation. The result measures detection of a practice, rather than infections prevented. [6]
In this research area, Stanford investigated visual reminders beside hospital rooms. The choice of visual signals incorporated healthcare staff’s views: they advised against adding more audible warnings to an environment already filled with alarms. [7]
Its contribution to prevention in healthcare work
The specific connection is supporting an infection-control practice within clinical work and designing alerts compatible with the task. For OSH, the case allows the observed procedure and the burden introduced by its monitoring to be considered together.
Detecting dispenser activation does not verify the entire hand-hygiene technique or every biological exposure at the workstation. Nor does it establish fewer occupational infections. A preventive application needs to define which event it recognizes, what it may miss and how the team responds without turning each detection error into another interruption.
