Fei-Fei Li’s new bet: AI is incomplete without spatial intelligence

Fei-Fei Li needs little introduction on a show about AI. She created ImageNet, the dataset that helped ignite the deep-learning era, is a Stanford professor and co-director of the Stanford Institute for Human-Centered AI, and previously led AI at Google Cloud. Sarah Guo and Elad Gil have her on No Priors because of the newest line on that résumé: she recently co-founded World Labs, a company dedicated to spatially intelligent AI. Guo’s first question is the blunt one. After two decades of science and policy, why start a company now?

Because in her heart, Li says, she wants to build. She sees this as a critical, fun, exciting moment to make technology that everybody can use, and she believes in spatial intelligence — 3D world models that can empower people and use cases — enough to do it with an extraordinarily brilliant group of young technologists. Gil has a footnote. He had been desperately trying to convince some of her co-founders to start a company. Their answer was that they now had a bigger mission, with Fei-Fei.

She defines spatial intelligence as the ability to understand, reason about, interact with, and generate 3D worlds. The physical world is 3D no matter how we project it, and a true 3D digital representation would make design, creation, navigation, simulation, and AR/VR experience an order of magnitude easier. Humans have spatial intelligence as a core capability. Animals have it. The entire journey of evolution is deeply intertwined with it. So it is fundamental: without spatial intelligence, AI would be incomplete. World Labs, she says, is the first company they know of that is trying to solve the 3D-generation foundation-model problem. Crack that, and you unlock a lot of spatial-intelligence problems.

Guo asks whether those world models will someday be realistically accurate — physics included. Li’s standard is slightly different: realistically accurate, or plausible. You can create a fantastical world, but the geometry and physics of it need to be plausible. That, she says, is fundamental to spatial intelligence. From a neural and cognitive-science point of view, she thinks spatial intelligence is one of the hard problems evolution had to solve for animals. The job is brutal. Evolve an organ that collects light — eyes, mostly — then reconstruct a 3D world in the mind so the animal can navigate, act, and interact. Humans are the most capable animals at manipulation, and all of that is spatial intelligence, rooted in how we think.

What is interesting, she says, is that animals have solved it only to an extent. It is not fully solved even in us. Close your eyes right now and build a 3D model of the room around you: it is not that easy. Most of us cannot generate an extremely complicated 3D model until we get trained. Architects, designers, and people with a lot of training and talent can do it, and it is still hard. Imagine doing it at your fingertips, with much more fluid interactivity and editability. That would be a whole different world for people — no pun intended.

Gil asks what other big gaps remain as we build the AI future. Li’s three buckets are language, 3D, and emotion. Language, she would say, is solved to a huge extent. 3D is as critical and as difficult. The entire space of emotional intelligence is something she does not even know how to begin to solve. She knows a lot of people who have not solved it either. When AGI is achieved, she jokes, the training data for that is not going to come from Silicon Valley people. Guo tells her not to underestimate Silicon Valley. Li puts herself in the bucket and still thinks you need a broader set of people. Gil adds physics simulation at macro and micro scales — the micro end becoming materials science and molecular modeling — and notes that robotics is as much a systems-integration problem as a model problem. Animal intelligence, he says, looks more distributed than a CNS story implies.

Li’s Stanford lab also works in robotics. She has no doubt humanity will move into an age of cohabiting with robots, and she wants the word itself unbound from humanoids. Robots will take all kinds of forms and shapes. A few years ago her lab wrote a paper on morphological intelligence: the body of an agent can itself change by optimizing the task it is trying to achieve. We should be a little more imaginative, she says, than just humanoids. Guo lays out a data hierarchy for robotics foundation models — off-the-shelf video, simulation, teleoperation, then expensive embodied collection — and asks whether people underestimate simulation and world models. Li likes the pyramid. Training will be a hybrid of many forms of data. Simulation is underrated, she says, then immediately qualifies: it is not underrated by experts. Plenty of robotics companies are already working on simulation and synthetic data. Outsiders are the ones who undersell it.

What is truly underappreciated, in her opinion, is haptics. Unlike language models or even spatial-intelligence foundation models, robotics is a highly multimodal system. If you want manipulation and not just navigation, integrating haptic data into vision, perception, and spatial data is absolutely critical. Gil then poses the form-factor fight. One camp says supply chain and manufacturing scale will collapse robots into a few shapes. The other says the economic value of specialization is so high that there will be thousands of form factors. Li expects gradient descent toward productivity and efficiency, and she lands with the second camp. Task requirements are so vast that sticking with one form is energy-inefficient.

Her trivial example is not trivial if you care about watts. If we put robots underwater, they should not be in the shape of humans. They had better be in the shape of fish. The same with flying: she does not think human form is right, and she notes that airplanes are becoming more and more like robots. Diversity of morphology, in her telling, is the destination, not a rounding error.

Gil asks for near-term commercial applications of generated 3D worlds. Li’s first answer is creativity. She draws a straight analogy to software engineering. Today’s LLM success in that field, including applications like Cursor and Windsurf, is collaboration between AI and humans at different skill levels. Creativity will look similar. Designers, 3D artists, VFX artists, marketing talent, and game developers all need to design and create 3D space. It is a fundamentally hard problem even for trained people. A collaborator that does it right would be extremely fun. That is the exciting near-term surface. The other line is XR. A lot of what the industry has been waiting for in the metaverse, AR, and VR is content creation. Hardware still has to evolve, but software is looking for content, and that lends itself naturally to 3D modeling and generative spatial models.

Guo asks whether world models are an interesting answer to scalable RL for more generalizable agents. Li thinks they are, because AI is not complete without spatial intelligence and humans interact in 3D worlds. Take design: in the mind’s eye you are simultaneously optimizing for beauty, efficiency, constraints, whatever it is, and that process lends itself pretty naturally to RL settings. The hard parts of the path are not a mystery. Data is absolutely a challenge. Images and video exist; large numbers of 3D worlds in the format you need to train on do not. Creating 3D foundation models, she says, requires more and more sophisticated data engineering, acquisition, processing, and synthesis. She is envious of her NLP colleagues. Language data is abundant on the internet. 3D does not have that luxury.

The other challenge is ironic. Every one of us uses 3D every day. Open your eyes and the life you experience is 3D, even when you are typing at a screen. Yet it is still not as easy a form factor to deliver into people’s hands as language. Language is easy, and it is also an active form rather than passive viewing. Nobody wakes up and says they are just going to sit here and watch 3D. That creates a productization problem. Li is not a gamer, though her kids love Minecraft. Asked what world she wants to experience, she wants to see worlds she does not see: zooming into microscopic worlds, going inside an engine and watching it work, even — she knows people will laugh — being inside a dishwasher. All of that can be done virtually if the world models exist.

Before the taping, Gil asked Andrej Karpathy what to ask. Karpathy said Fei-Fei is magic about ambition and thinking about data, and that the story of her PhD and the 101 dataset with Pietro Perona is instructive. Li is delighted her student is more well known than she is; it makes her proud. The work goes back to about 2003. The internet was barely a thing, data was not much of a thing, and her PhD problem was object recognition: call out cats, dogs, microwaves, and chairs from a picture. They were beginning to hypothesize that data matters. There was no scaling law. They had no idea how far data could go. Every dataset out there was tiny, and she was not convinced. Perona told her to curate one. They talked about 15 categories, or 30. Then, she says, God forbid, the advisor named a three-digit number: 100.

It was a lot of work. Deep in her heart she knew he was right. From a mathematical point of view, pushing a model to generalize — neural net, Bayes net, SVM, whatever was popular — needed enough data. She writes about the process in her book *The Worlds I See*. She stumbled on a Webster’s dictionary she was using to study English, with seemingly random visual depictions of words: flowers, bicycles, dogs. She grabbed 101 of those words, one more than asked, which made Perona chuckle. Google Image Search was new and terrible. Cleaning crashed her. She wrote a tiny click-only interface and handed it to her mother, who did not know computers but knew how to click. That dataset became Caltech 101. ImageNet was the same idea at a different scale: 15 million labeled images across thousands of categories, not 101. Amazon Mechanical Turk came to the rescue. Early on, she was told she would not get tenure for doing it.

Then AlexNet won, and a couple of years ago at an event in Toronto, Geoff Hinton said publicly how defining that was, almost apologetic that ImageNet was not as recognized as neural networks. For scientists, Li says, validation is not about recognition or awards. It is that you made a difference — the conjecture nobody believed, the hypothesis nobody believed, made to happen. Gil remembers that in 2016 or 2017 his history-of-AI slide had a very small number of events; ImageNet and AlexNet were among them. Transformers and maybe diffusion models have joined the slide since. Another moment she is proud of is the dissertations of Karpathy and Justin Johnson, where language and images converged in captioning and storytelling. At the end of her own PhD she had literally thought that if she lived to 100, storytelling from pictures might be the problem they could solve. ImageNet would do object recognition; the rest of the career would go to stories. Around 2013 and 2014, LSTM plus CNN blasted image captioning open. Her group’s work, together with Google’s, was first out the door. She was so proud she almost had a crisis: what was she going to do for the rest of her 65 or 70 years. The field had moved that fast.

Guo asks whether there is still room for moonshots and creativity in AI research outside $10 billion corporate training runs. Li’s singular advice, the one she still says in her company and her lab, is be fearless. Scientists, technologists, and entrepreneurs have to be. Eventually you have to figure out whether you need a $10 billion run, and then you come to Sarah for funding — probably a lot of both, Guo jokes — or you have to figure out data. Fearless, for Li, sits in a very specific place: somewhat delusional and crazy, and somewhat just rationally bold. Too rational and you are not courageous enough; you do not identify problems that are big enough. Completely crazy and many things can go wrong. Even as old as she is, that is how she feels she started World Labs: she wants to be fearless and solve spatial intelligence.

The company has open roles. They want the best engineers, AI researchers, and product thinkers. Under the hood it is not a homogeneous “AI company.” They hire computer-graphics experts, computer-vision experts, data people, generative-AI people, machine-learning infrastructure, optimization. A problem as hard as spatial intelligence is not a single-background problem. Then she looks for fearlessness. How do you identify it? You talk to them. You can sense what drives someone, and you can sense the questions they ask. Of course you have to ask how to get the thing done. If that questioning comes from being scared of solving it, that is not fearlessness. Fearless people, she says, are creative and ambitious. They are not afraid of uncertainty or the unknown. She really loves that.

The last question is the other pride of her career: Stanford’s Human-Centered AI Institute, and the movement toward that way of thinking. The world she wants is one where AI collaborates with and superpowers people, and where the human world stays human-centered — love, relationship, prosperity across communities, justice. No piece of machinery, whether AI or airplane or biotech, should take those away. With those values in mind, having AI superpower us still matters, because there are so many unsolved problems. Healthcare is the example she has worked on at Stanford: drug discovery, diagnosis and treatment that can reach all people, delivery, aging, chronic disease, mental health. The bottleneck is not excess humans. It is lacking help — lacking scientific discovery, diagnosis, precision medicine, safer delivery. Guo says she and Gil have invested in companies along that spectrum, from Abridge to OpenEvidence to Latent, and that after 15 years of pessimism about healthcare-tech adoption, this time it feels different. Li started a digital-health company before this. Her hope is that things people have talked about for decades will finally come to fruition, and that AI is a great delivery mechanism for them.