In recent years, we’ve seen generative AI move quickly through different eras: chatbots, image-generation, voice, video-generation, and more. But Dr. Fei-Fei Li, longtime AI pioneer and co-director of Stanford’s Institute for Human-Centered Artificial Intelligence (HAI), is staking out what she thinks is the next frontier: spatial intelligence, a nascent field that she believes is the “defining challenge of the next decade,” as she wrote in a Substack post this week.
That’s why Li co-founded World Labs in 2024 — and raised $230 million last fall — to build world models, or generative AI models that can “perceive, generate, reason, and interact with the 3D world,” per the company. And this week, World Labs released its first commercial product, Marble, which allows users to generate their own downloadable 3D worlds from text, image, or video prompts. Li wrote that she foresees spatial intelligence potentially transforming sectors from storytelling and filmmaking to architecture, robotics, and scientific discovery.
“We see that [the] world model is just as big and exciting, if not more [than the previous eras],” Li, who is also CEO of the company, told The Verge in an interview.
“Bringing 3D to life, and understanding the richness of spatial and 3D stuff, is just a whole next level beyond the baseline of most of these other single modes,” Ben Mildenhall, co-founder of World Labs, told The Verge. He added that, for solely human teams, “It’s such a massive problem to build these worlds. It requires such a large team and so many pieces of software and so much time and effort … Think about the radical change there, that can come if you empower people to build stuff much more rapidly — ideate, iterate, and edit things in a much tighter loop.”
Marble offers four subscription tiers: Free, which allows for up to four world generations; Standard ($20 per month), which allows for up to 12 generations and more editing options; Pro ($35 per month), which allows for 25 generations and commercial rights; and Max ($95 per month), which allows for up to 75 generations, as well as everything the pro tier offers. The Verge was able to generate an open-air castle with waterfalls and, in other users’ generations, explore ruined structures reclaimed by nature and Hobbit-like spherical homes. It was possible to take a few steps into such environments before essentially running into a wall in the 3D generation, and in the non-free tiers, the downloaded files are compatible with tools like Unreal Engine and Unity. Mildenhall said that he’d seen some people who are willing to put in hours of work are able to stage out “fairly large environments” using Marble.
Mildenhall said he could imagine authors using it to build out their imagined world, or people working on VFX or location scouts in the filmmaking industry. At the enterprise level, he said, he could imagine companies using Marble or one of World Labs’s future products to analyze and visualize their wide swaths of data.
“Even given the limitations of this model, we are seeing the light beyond where we are in some emerging behaviors,” Li said, adding that people can put together spaces in a way that’s “beyond human imagination.”
