So, you published this essay in November titled "From Words to Worlds: Spatial Intelligence is AI's Next Frontier." And you basically present that language models in the past or last wave of AI were dominant, but the next era comes with spatial intelligence. So I thought to start to define spatial intelligence, please.
Yeah. First of all, I think language intelligence is critical. So this is not by any means to say language is the past. In fact, look at the economy and all the exciting work and the future—we will use large language models and language intelligence in more and more ways. But just as we all know that human intelligence is not just linguistic, right? Just think about this morning: you wake up, you hug your kids, you make breakfast, you drive them to school, you go to work, you make coffee. Everything we do in our daily life as well as in our work involves the 3D world, involves space, involves movements, involves interaction, involves the ability for us to see, to reason, to make sense, to act—and that is spatial intelligence. It goes beyond language, and it goes beyond linguistics.
In the essay, you describe language models or today's systems—probably a better phrase—as "wordsmiths in the dark." Very poetic, but would you expand a little on what you meant by that?
Yeah, I was using this metaphor because I love reading about evolution, and a lot of my own research is inspired by the evolution of the nervous system. If you look at animal intelligence, it began half a billion years ago when animals started to sense the world. In fact, the most ancient sensory systems are haptics and vision. It's the ability to touch your environment as well as see your environment. So, anything before that was profound darkness. And it's not the darkness of contrasting light from dark; it's actually the unaware situation of what the world is. But the flip side is when animals became aware of what this world is, became aware of their own body and its relationship with this world, evolution's arms race began, animal speciation began. Nervous systems began to evolve in a rapid way till we have very sophisticated animals like us. So seeing the world in a profound way, in a way that you participate in your movement, in your interaction, in your communication, is critical for intelligence. And not having that is intelligence in the dark.
It's difficult for industry, even in academic fields, to distill it to its most basic. But the goal is to build systems that can perceive, reason, and act in the world as we do, as humans do—
—and generate. And so there is a fundamental difference between what you call a world model and a language model, and I think probably helpful then to explain the key differences.
Yeah. So, it's 2026. There are many definitions of world model. It's actually a funny moment because about two years ago, when I was actually preparing my TED Talk in 2024, it was the first time I shared with the world about my vision for spatial intelligence and world models, and no one was talking about this. It was actually an easy time for me to define spatial intelligence and world models. Come 2026, I'm keenly aware that different groups have different definitions of world model at this point. But for me, a world model involves creating intelligent machines or intelligent models that have the ability to understand space, what's in it, to reason about the geometry, the interactivity, and eventually the physics and dynamics of this space, as well as the ability to generate 3D/4D space just like our computers today generate words. And when we endow computers or software with the ability to generate space and interact in this space, we really unlock a lot of applications—from experiences, gaming, art, and creativity, design, to robotics, to education, to healthcare applications, to manufacturing. There's just a lot more that can be unlocked.
In the course of the essay, it's almost like an academic framework: argument, hypothesis, and at the end, conclusion. And in between, there's a pitch in there for your company, World Labs. You've had this career at the forefront of research for 25 years, and I think it's probably time to ask why you felt that to move spatial intelligence forward, it had to be done within the company setting and not at Stanford or in the academic field.
Yeah. So starting World Labs was a decision I made at the beginning of 2024. I know there are many startups in the space of AI being started almost every day—you cover all of them. For me, this was not an impulsive project. It really is a continuation of my career; in a way, it's my life's work. My entire career has been focusing on visual intelligence and trying to make computers see and understand. And by the time the generative AI model era arrived, around 2022, one of the most important revolutions that I saw, that the public didn't see yet around 2023, is that not only did we have a generative AI revolution happening because of the transformer models—we saw ChatGPT get rolled out in late 2022—we also had a quieter revolution in computer vision, 3D computer vision, and computer graphics in our ability to model the 3D world. And these two areas of technology were actually converging. So for the first time, it gives us the opportunity to really conquer some of the most fundamental problems in AI, which is spatial intelligence, as we just described. At that time, I was working on robotics at Stanford, and I knew that's a lynchpin technology for robotics and creativity, so I was very excited. The truth is, to work in GenAI, you need a lot of resources: you need compute, you need data, and you also need a focused team of incredible talents. To do that at that scale, we can no longer do it in academia. So I decided, with my former student and a bunch of founding members, to start World Labs.
I think it's fair to say that the vast majority of your focus now is World Labs, as opposed to your prior affiliations with different parts of Stanford. You and I—
I did just teach the first lecture of CS231n last week.
Yeah, some of you probably have taken that class, right?
You know, you and I have discussed this in the past, though. The other way of looking at it is: could this work be done in an academic institution? But as you outlined it, it's a constraint of resources.
So let's be very clear. One thing that's very special about our country, and in fact, I think the modern world, is the ecosystem of innovation, and America is really the embodiment of that. Especially in the past decades, especially since World War II, this country really did invest in the ecosystem of innovation—not only on the industry side, but as well as on the public sector, academia side, as well as private resources. So I still believe it's critical to have academia. It's critical to have public sector research and innovation, just because this particular spatial intelligence technology that I believe in can only be done in industry. In fact, I believe it can only be done in a place like World Labs because we're so focused. It doesn't mean there isn't a place for academia. I actually think this is the moment that we need to resource academia so that the researchers and students who are not tethered to a revenue-driven business model can go after the blue-sky, curiosity-driven research that the next generation needs. So they're not mutually exclusive.
World Labs is focused. It's about 40 people. Is that right? Yeah. You just raised $1 billion. What was that like?
I thought my friends in Nvidia and AMD would be very happy.
Well, a better question is: how much of that went to Jensen?
And Lisa, right? No, look, AI is very expensive. We are a frontier model company. Being a frontier model company means that we need a lot of compute resources, a lot of data, and a lot of expensive talent. I'm keenly aware of that. I'm also keenly aware that in human history, startups before we have really proven our market success in a way that traditional companies have, traditionally don't have this level of resources. So it keeps me awake every day, but it also does excite me because we have a historical opportunity to do something that nobody could do before.
In the earlier wave of AI, there were three unlocks: ImageNet, which you and your team should get great credit for, neural networks as a broad term, and then GPUs. And so if we're entering this next era going towards spatial intelligence, what are the constraints on progress? And I suppose, what will the unlocks be that your team is working on?
Yeah. So I do think these three key elements of modern AI are still very much true for our work in spatial intelligence. Compute, we already talked about; and model, we need to be really innovating new models. I do think the hardest piece is data. Compared to language, there isn't an internet of spatial data out there. There are unique datasets, for example, Google Earth data or some more unique data, but not in a way that language data is so easily and readily available. So we have to do harder work.
Your colleagues in industry, particularly those in physical AI, would put a lot of emphasis on synthetic data, simulated data. How is that relevant to the ambitions of World Labs?