Fei-Fei Li

Nearly everyone in computer vision was refining algorithms; Fei-Fei Li went and labeled millions of images across roughly 20,000 categories instead. In 2012 a neural network trained on that dataset won her own competition by a margin wide enough to revive the field's interest in deep learning.

Li is a Stanford University computer science professor and the co-founder and CEO of World Labs, a company that builds AI systems designed to understand and generate three-dimensional space rather than text. Before founding World Labs, she was best known for creating ImageNet, the large image dataset that helped trigger the deep learning boom of the 2010s. She argues that spatial intelligence, the ability to perceive, reason about, and act in three-dimensional environments, is the next major frontier for artificial intelligence, beyond the language-centered systems that have dominated the field since large language models rose to prominence.

Background

Li was born in Beijing in 1976 and grew up in Chengdu, in China's Sichuan province. Her father immigrated to the United States first, settling in Parsippany, New Jersey. Li and her mother joined him there when she was sixteen. The family had little money, and Li worked weekends at a Chinese restaurant and later helped run her parents' dry-cleaning shop while attending Parsippany High School, from which she graduated in 1995. She went on to Princeton University, majoring in physics and earning her bachelor's degree in 1999. She earned a PhD in electrical engineering from the California Institute of Technology in 2005, working on computer vision and computational neuroscience. During her graduate years at Caltech she met Silvio Savarese, a fellow computer vision researcher whom she later married.

Getting started

After postdoctoral and early faculty positions, Li joined Princeton's computer science department as an assistant professor in 2007. At the time, most computer vision researchers focused on refining algorithms, on the assumption that smarter models, not more data, would drive progress. Li took the opposite view: she believed models trained on a small number of labeled examples would always be brittle, and that progress required a dataset of unprecedented scale, organized around the tens of thousands of object categories humans can recognize. Beginning around 2006 and continuing after her move to Princeton, she and her collaborators built ImageNet, ultimately labeling millions of images across roughly 20,000 categories using crowdsourced annotation. She then created a public annual competition, the ImageNet Large Scale Visual Recognition Challenge, to benchmark progress on the dataset. In 2012, a neural network called AlexNet, trained on ImageNet, outperformed every rival entry in the competition by a wide margin, a result widely credited with reviving interest in deep learning across the field. Li moved to Stanford in 2009, later led the Stanford Artificial Intelligence Laboratory from 2013 to 2018, took a leave to serve as chief scientist of AI and machine learning at Google Cloud, and co-founded the Stanford Institute for Human-Centered Artificial Intelligence, which she still co-directs. In January 2024, she co-founded World Labs with researchers Justin Johnson, Christoph Lassner, and Ben Mildenhall, moving from studying AI as an academic into building it as a company for the first time.

What she built

World Labs builds on Li's view that understanding language is not the same as understanding the physical world, and that AI systems trained only on text and images will remain limited in their ability to reason about space, physics, and cause and effect. The company's first product, Marble, takes a text description, a photograph, a video, a panorama, or a rough spatial sketch as input and generates an explorable, editable three-dimensional environment as output. Marble produces both a visual rendering, using a technique called Gaussian splatting, and a physically usable representation of the same space, including collision meshes a physics engine or robotics system can act on, so the same generated world can be viewed by a person and operated on by a machine.¹ World Labs emerged from stealth in September 2024 at a reported valuation near one billion dollars. Marble launched in limited beta in November 2025 and moved to a full commercial launch in February 2026, timed to a one-billion-dollar funding round backed by major chipmakers and software firms. By early 2026, World Labs was reportedly in talks to raise additional capital at a valuation near five billion dollars.

How she operates

Li frames her own work through a taxonomy she has laid out publicly: AI systems that generate a world fall into three functions, a Renderer that produces pixels for a person to look at, a Simulator that produces physically and geometrically faithful state, and a Planner that produces actions toward a goal. She argues simulation is the least discussed of the three publicly and the most consequential, since the same knowledge of geometry, physics, and dynamics needed to simulate a space is also what is needed to render it and to plan actions within it. As she has put it, "of the three categories, the simulator gets the least public attention, and is the most consequential of the three."¹ She describes Marble's approach, generating both a rendering and a physically usable structure from one model, as a step toward collapsing the boundary between those functions. She has also spoken about betting on data and scale over clever algorithmic tricks as a recurring theme across ImageNet and World Labs alike, and has acknowledged that the scarcity of annotated three-dimensional data, unlike the abundance of text and video online, is the central bottleneck facing her new company.¹

Where things stand

As of the mid-2020s, Li holds her Stanford professorship and co-director role at the Stanford Institute for Human-Centered Artificial Intelligence alongside running World Labs full time as CEO. World Labs has shipped Marble commercially, raised roughly 1.2 billion dollars in total financing, drawn investment from major chipmakers and software firms, and was reportedly in discussions to raise further capital near a five-billion-dollar valuation. The company describes Marble as an early chapter in a longer roadmap toward a unified world model spanning rendering, simulation, and planning, and has not yet detailed products beyond three-dimensional environment generation.

Key facts

• Created ImageNet, a dataset of millions of labeled images across roughly 20,000 categories, while a professor at Princeton starting around 2007. The dataset trained AlexNet, whose 2012 competition win is widely credited with reigniting deep learning.
• Immigrated from Chengdu, China to Parsippany, New Jersey at age sixteen and worked in her family's dry-cleaning shop while in high school.
• Holds a BA in physics from Princeton (1999) and a PhD in electrical engineering from Caltech (2005).
• Led the Stanford Artificial Intelligence Laboratory (2013 to 2018) and served as chief scientist of AI and machine learning at Google Cloud before co-founding World Labs in January 2024 with Justin Johnson, Christoph Lassner, and Ben Mildenhall.
• World Labs raised roughly 230 million dollars entering stealth in 2024 and a further 1 billion dollars in February 2026 from Nvidia, AMD, Autodesk, Fidelity, Emerson Collective, and Sea, with reported talks of a subsequent round near a 5 billion dollar valuation.
• Marble generates explorable 3D environments from multimodal prompts, outputting both Gaussian splats for visual rendering and collision meshes usable by a physics engine.