Dr. Fei-Fei Li on Spatial Intelligence

Language models changed everything. But according to Dr. Fei-Fei Li, they're only part of the story.

At HumanX 2026 San Francisco, the World Labs founder and CEO — and one of the most influential figures in the history of artificial intelligence — laid out a vision for what comes after the language era: spatial intelligence. AI that doesn't just process words, but understands space, physics, and the 3D world we actually live in.

It's a big claim. But Dr. Li has the track record to back it up — from ImageNet to Stanford's AI Lab to a billion-dollar startup focused on making it real.

What Is Spatial Intelligence?

Dr. Li's framing starts with a simple observation: human intelligence isn't just linguistic. "Think about this morning," she said. "You wake up, you hug your kids, you make breakfast, you drive them to school. Everything we do involves the 3D world, involves space, involves movement, involves interaction."

Spatial intelligence is AI that can perceive, reason about, and generate three-dimensional environments. It goes beyond understanding text to understanding the physical structure of the world — geometry, physics, dynamics, and the relationships between objects in space.

Dr. Li described today's language models as "wordsmiths in the dark" — powerful, but fundamentally unaware of the physical world they're describing. Drawing on the evolution of animal nervous systems, she noted that the most ancient forms of intelligence weren't linguistic — they were spatial. Vision and touch came first. Language came much later.

"Seeing the world in a profound way, in a way that you participate in your movement, in your interaction — that is critical for intelligence," she said. "And not having that is intelligence in the dark."

Why World Labs, Why Now

Li started World Labs in early 2024, not on impulse, but as a continuation of her life's work in visual intelligence. The timing came down to a convergence: transformer models and advances in 3D computing and computer graphics were finally making it possible to tackle spatial intelligence at scale. But doing it required resources that academia alone couldn't provide.

"To work in generative AI, you need a lot of compute, a lot of data, and a team of incredible talent," she said. "We can no longer do that in academia."

World Labs, now roughly 40 people and backed by $1 billion in funding, is building what Dr. Li calls a frontier model company for spatial intelligence. Their first model, Marble, generates true 3D worlds — not videos, not flat renders, but persistent, spatially consistent environments that can be used for game development, robotics training, VFX workflows, and more.

Dr. Li was candid about where things stand. "About a year and a half ago, the worlds we created were so small that we felt they were more like closets than worlds," she said. "Right now, they're about the size of a hotel lobby." The ambition is much bigger, but the foundation is being laid.

The Data Problem — Especially for Robotics

If there's one bottleneck Dr. Li kept returning to, it's data. Language models had the internet. Spatial intelligence doesn't have an equivalent — at least not yet.

"If you think that 3D world data is lacking, robotics data is even more lacking," she said. This is where World Labs sees a flywheel opportunity: Marble's generated worlds can serve as training environments for robotics teams, feeding data back into the ecosystem that accelerates progress across the field.

Dr. Li sees spatial intelligence as the "linchpin technology for robotics,” but acknowledged it's still early. Most robotics labs today use a hybrid approach of real-world and synthetic data, and the question of what counts as "real" versus "synthetic" is getting blurrier. "What if it's a reconstruction of the real world that, even though digitized, reflects reality?" she asked. "What we are building is part of that greater technology."

Spatial Intelligence Is Already Here

When asked where spatial intelligence shows up first, Dr. Li pushed back on the premise. It's already here.

"Tesla is building a world model for self-driving cars. Waymo has built a world model for self-driving cars," she said. In healthcare, she described conversations with doctors using spatial intelligence for radiology — understanding 3D structures of the human body from 2D imaging data.

"Sometimes technology is already happening. We're just not realizing it yet."

As for a ChatGPT-scale watershed moment for spatial intelligence, Dr. Li was measured. "It's hard for me to predict, partially because I don't know if there is one consumer behavior using world models that is similar to using language," she said. Chat is ubiquitous. A 3D equivalent might not arrive the same way — but the applications across creative industries, robotics, healthcare, and manufacturing are already building momentum.

The Godmother Question

Ed Ludlow raised it. Dr. Li was honest about it.

"I really did cringe," she said of the title. "That's not my natural personality." But she paused — because rejecting it might close a door not only for her, but for others who may come after her. "The reality is there are so many godfathers of many things, including a lot of appointed godfathers. And there are many brilliant women in history, in the future."

It was one of the most genuine moments of the conversation — and a reminder that the people shaping AI's future don't always fit the narratives built around them.

Builder and Researcher

Dr. Li closed by drawing a line between her two roles. As a researcher, the north star is curiosity. As a founder, it's impact.

"That feeling of starting from ground zero and trying to contribute to this civilizational moment, contribute to a technology that can change people and make people's lives better — that energy is just incredible."

Spatial intelligence may not have its ChatGPT moment yet. But if Dr. Li is right, it won't need one. It'll show up everywhere — in the cars we drive, the games we play, the robots we build, and the way we understand the world around us. The question isn't if. It's when the rest of the industry catches up.

Watch the full session on-demand. 

Register for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027).