Fei-Fei Li, known as the godmother of AIjust closed a $1 billion round for World Labs, his startup dedicated to teaching machines to understand the world in three dimensions. Behind this bet are large companies such as NVIDIA, AMD, Autodesk or the Andreessen Horowitz fund, among others. Li, like other important figures in the field of AI, believes that world models are the way to go, instead of the AGI.
Who you are and why what you do matters. Li is one of the people who made it possible for the Generative AI as we know it today existed. He was part of the team that developed ImageNet, a database of millions of images that allowed computers to learn to recognize objects in photos. That academic work was the trigger for the leap towards deep learning that gave rise to everything that came after: from voice assistants to generative models of text and images.
Now, from Stanford University, where he directs the Institute for Human-Centered Artificial Intelligence, and from World Labs, the startup he founded in 2024, Li points to what he considers the next big unsolved problem in AI: that machines understand the physical world, not just text or flat images.
The problem you want to solve. The great language models like GPT either Claude They are extraordinarily good at processing text. But the real world is not text, or at least it is not only text: it is three-dimensional, it has physics, it has geometry, it has objects that move and relate to each other. “If AI is to be truly useful, it must understand worlds, not just words,” counted Li in his statement.
That is what so-called spatial intelligence, the central focus of World Labs, pursues. Unlike working with two-dimensional data, the models the startup works on are designed to perceive, generate and interact with three-dimensional environments. The idea is that an AI with spatial intelligence can reason about how things work in space, where an object is, how it moves, what will happen if it is pushed, how it fits into a larger environment, etc.
What already exists and what is coming. In November of last year released Marbleits first commercial product. It is a model that generates editable and downloadable 3D environments from text, images, videos or panoramas. The user can create a virtual world, modify it, expand it and export it in different formats. The startup positions it mainly for video games, visual effects and virtual reality, or sectors with a huge demand for 3D content in which there are few tools to put them into operation.
With this new round of financing, the focus also expands to robotics. And in this field, spatial intelligence is especially critical, since a robot that understands the space around it can plan actions before executing them, process different ways of completing a task or adapt to changing environments without needing to be reprogrammed for each situation.
Autodesk has put 200 million at your table. It really makes perfect sense. It is the company that makes the design software used by architects, engineers, animation studios and manufacturers around the world. Your business is, by definition, thinking in three dimensions. And if Li’s models can generate and reason about 3D environments, Autodesk tools can also benefit from what the startup aims to offer.
Daron Green, chief scientist at Autodesk, explained to TechCrunch that the collaboration between both companies will initially focus on entertainment and audiovisual production. The idea is that design workflows can be combined with AI-generated worlds. In this way, a user designs an object in Autodesk and places it in an environment created by World Labs, or the other way around. “You might anticipate that we will consume their models or that they will consume ours in different contexts,” Green said.
You are not alone in this race. World Labs is not the only commitment to world models. Google DeepMind works on your family of Genie modelscapable of generating and simulating 3D environments. Yann LeCun, who was chief AI scientist at Meta, just founded AMI Labs with the same approach. Startups like Decart and Odyssey They also move in this spacealthough with products still in the demo or research phase.
However, there are differences in their respective approaches. LeCun, for example, defends that to build true world models a completely new AI architecture will be needednot generative. Li, from World Labs, is committed to advancing with current generative models and improving from there.
Cover image | World Labs and Andria Lo
In Xataka | We’d love to tell you that ‘Her’ hasn’t come true and there aren’t people dating an AI, but we can’t.
GIPHY App Key not set. Please check settings