Fei-Fei Li

Co-founder & CEO — World Labs Sequoia Professor, HAI Co-Director — Stanford University

Listen — profile

0:00 / 2:35

Profile

Fei-Fei Li is the person who put the fuel in the deep learning fire. In 2006, while most of the field was obsessing over smarter algorithms, she bet that the real bottleneck was data. Five years and a lot of Amazon Mechanical Turk hours later, she and her students shipped ImageNet — 14 million hand-labeled images organized against the WordNet hierarchy — and started running the ILSVRC competition to see who could classify them best. For two years, entrants inched forward. Then in 2012, Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky showed up with AlexNet and cut the error rate nearly in half. The deep learning era had a starting gun, and Li had built the track.

She’s been a Stanford professor since 2009, ran the Stanford AI Lab, and co-founded Stanford HAI in 2019 with philosopher John Etchemendy — an explicit bet that AI needs humanists and policy people in the room, not just ML PhDs. Between 2017 and 2018 she took a sabbatical to serve as Chief Scientist for AI/ML at Google Cloud, which ended messily (see Controversies). Her lab trained a generation of people now running parts of the field — Andrej Karpathy, Jia Deng, Justin Johnson, and Jim Fan all came through there.

Her current project is World Labs, co-founded in 2024 with Justin Johnson, Ben Mildenhall, and Christoph Lassner. The thesis: language models plateau on problems that require reasoning about 3D space, and the next frontier is “spatial intelligence” — AI that perceives, imagines, and acts in three dimensions. In November 2025 they released Marble, which generates editable 3D worlds from images, video, or text. In February 2026 the company raised $1 billion from AMD, Nvidia, Autodesk, and others, putting real money behind the spatial-AI bet.

For developers, Li’s career is a useful reminder that the field’s biggest unlocks often come from infrastructure, not models. ImageNet wasn’t an algorithm — it was a dataset and a benchmark, and that was enough to change everything. If you’re trying to figure out where to put your attention, watch what she’s doing with world models; the 2D-to-3D shift is her next call on where the scaling curve bends.

Books

The Worlds I See: Curiosity, Exploration, and Discovery at the Dawn of AI

2023

Part memoir, part history of ImageNet and the deep learning breakthrough told from the inside — honest about the grind and the doubt.

Key Articles & Papers

ImageNet: A Large-Scale Hierarchical Image Database 2009 — The CVPR paper that introduced ImageNet. Unglamorous poster session, field-shifting consequences. ImageNet Large Scale Visual Recognition Challenge 2014 — The ILSVRC retrospective — how the benchmark worked, who won, and why it mattered. Deep Visual-Semantic Alignments for Generating Image Descriptions 2014 — With Karpathy — one of the first systems that could caption images with natural language sentences. Visual Genome: Connecting Language and Vision Using Crowdsourced Dense Image Annotations 2016 — The follow-up dataset to ImageNet, aimed at relationships and scene understanding rather than just object labels. From Words to Worlds: Spatial Intelligence is AI's Next Frontier 2025 — Her manifesto for why 3D world models are the next scaling story after LLMs. How to Make A.I. That's Good for People 2018 — NYT op-ed laying out the human-centered AI argument that became Stanford HAI's founding thesis.

Videos

Controversies

Project Maven (2018). While running Google Cloud AI, Li was tied to Project Maven, a Pentagon contract to apply computer vision to drone footage. Leaked internal emails showed her cautioning colleagues to “avoid at ALL COSTS any mention or implication of AI” in public communications about the contract — a tone that read to critics as PR management rather than ethical engagement, and sat awkwardly next to her public advocacy for AI-for-good. She left Google Cloud in late 2018 and returned to Stanford. Coverage: The Intercept, NYT.