Responsible AI · Concepts & context

Fei-Fei LiFrom ImageNet to spatial intelligence

How a pioneering image dataset, a landmark neural network, and a new generation of world models connect—and what they tell us about AI’s ability to understand the physical world.

Fei-Fei Li: the “Godmother of AI”

The note “Lee — God Mother AI” almost certainly refers to Fei-Fei Li, the Stanford computer scientist and cofounder and CEO of World Labs. Frequently called the “Godmother of AI,” she helped shape modern computer vision and deep learning through her work on ImageNet.

Sources: Stanford HAI — Fei-Fei Li · World Labs — About

ImageNet: teaching machines to recognize images

ImageNet is the large image dataset that Li helped create, closely associated with an influential image-recognition competition. Its basic insight was that machine learning needed enormous amounts of labeled visual data: millions of photographs identified with concepts such as dog, car, chair, and bird.

ImageNet became a proving ground where modern deep learning demonstrated how powerful it could be.

Source: IEEE Spectrum — Fei-Fei Li and World Labs

2012: Hinton, AlexNet, and a turning point

In 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton developed AlexNet, a deep convolutional neural network, and entered it in the ImageNet competition. It dramatically outperformed the conventional computer-vision approaches of the time.

The result helped convince the AI community that deep neural networks, large datasets, and GPUs were an extraordinarily powerful combination.

Source: IEEE Spectrum — ImageNet’s role in the deep-learning breakthrough

The historical connection: Fei-Fei Li helps create ImageNet → AlexNet wins the 2012 ImageNet competition → deep learning gains momentum.

TIME’s Person of the Year

The note “TIME Person of the Year” likely refers to TIME’s 2025 Person of the Year: “The Architects of AI.” Fei-Fei Li was one of eight people pictured, alongside Jensen Huang, Sam Altman, Demis Hassabis, Dario Amodei, Lisa Su, Mark Zuckerberg, and Elon Musk.

Source: TIME — The Architects of AI, 2025 Person of the Year

World models: representing how an environment works

A world model is an AI system that attempts to represent an environment and, depending on its design, predict how that environment changes. The term covers systems with different capabilities; generating a convincing scene does not automatically mean understanding its physics.

Describing a kitchen

A language-and-vision model might say: “There is a table, a refrigerator, and two chairs.”

Reasoning about a kitchen

A more capable world model could represent the table’s depth, where a chair blocks movement, what is hidden behind the refrigerator, or what might happen when an object falls.

The kitchen example illustrates the research goal; it is not a claim that every world model can perform all of these tasks.

A useful taxonomy distinguishes world models that act as renderers (produce visual worlds), planners (predict outcomes to guide actions), and simulators (model geometry, physics, and dynamics).

Source: Radical Ventures — A Functional Taxonomy of World Models

Spatial intelligence: understanding the 3-D world

Spatial intelligence means understanding and reasoning about the physical, three-dimensional world. Humans routinely use it to judge:

  • Where objects are and how far away they are.
  • What is behind or hidden by another object.
  • How to move through a room without hitting obstacles.
  • What could happen if an object moves.
  • How a scene would look from a different viewpoint.

Li argues that AI needs these abilities to move from talking about the world to operating within it.

Sources: WIRED — The Godmother of AI Wants Everyone to Be a World Builder · World Labs — Spatial intelligence mission

What can’t LLMs do reliably?

Large language models learn patterns in tokens—units representing words and other data. They can describe physical situations and reason about many of them, but language alone does not guarantee a reliable internal model of three-dimensional reality.

An LLM can explain how to stack dishes. A robot actually stacking dishes must also handle depth, shape, orientation, gravity, force, motion, obstacles, and the consequences of its actions.

The key distinction is reliability in the physical world. Multimodal models can use images and other inputs, but that does not automatically provide consistent geometry, accurate physical prediction, or dependable robotic control.

This is why researchers including Li question whether simply making language models larger will be sufficient for capable physical AI. Spatial intelligence and world models may complement language models; whether they deliver a fundamental advance remains an open research question.

Context for this explanation: World Labs — The case for spatial intelligence · Radical Ventures — World-model capabilities and distinctions

Marble: making explorable 3-D environments

The note “Marbel World Models” almost certainly means Marble, the world-building product from Fei-Fei Li’s World Labs.

Marble takes inputs such as text, images, or video and generates persistent, explorable 3-D environments. World Labs describes its technology as generative 3-D world modeling.

Sources: World Labs — Marble · World Labs — Company and technology

Image generation

“Create a beautiful Victorian living room.”

The result is a picture of a space.

World generation

“Create a Victorian living room I can move around inside.”

The goal is a coherent space that remains consistent as the viewpoint changes.

An explorable 3-D environment is not, by itself, evidence of a complete physics simulator or a robot’s ability to act safely in the real world.

The connected story

These are overlapping strands of AI research, rather than a strict sequence in which one replaces another.

2009–2012 · A breakthrough

Teach AI to see

ImageNet supplies labeled images. AlexNet demonstrates the power of deep neural networks trained with GPUs.

ImageNet and deep learning

The language-model era

Talk and reason

Large datasets and transformer-based models support systems such as ChatGPT, Claude, and Gemini.

Language becomes a powerful interface for working with AI.

An expanding research frontier

Understand physical space

Images, video, 3-D representations, and action data support research into world models and spatial intelligence.

World Labs’ research direction

The larger question: Fei-Fei Li helped create the dataset that catalyzed a computer-vision revolution. Now she is betting that spatial intelligence can drive another major transition. How far will world models take AI beyond the limitations of today’s language-centered systems?