Hi everyone, welcome to Radical Talks. This is a podcast from Radical Ventures where we explore the AI era. We dig into the innovation happening at the forefront of AI. That's the cutting edge research, the new product paradigms, and of course, the startups building at the frontier. I'm Molly Welch, and I am a partner here at Radical. Today we're diving into what might be the next major leap and that's world models and spatial intelligence. We are enormously lucky today because we have a very special guest on the podcast and our guest is Dr. Fei-Fei Li. She hardly needs an introduction but I'll introduce her nonetheless. Fei-Fei is an AI pioneer who created ImageNet which was one of the driving forces of course behind the deep learning revolution. Critically, ImageNet was the data that made deep learning really work. And now, of course, data is recognized as more important. So, an incredibly important contribution. Fei-Fei is a professor at Stanford. She's founding co-director of Stanford's Human-Centered AI Institute and she's a former Google vice president. She's also the author of an acclaimed memoir, The Worlds I See, in which she shares her remarkable story as a scientist and as a person. Today, Fei-Fei is CEO and co-founder of World Labs, a spatial intelligence AI company building large world models that can perceive, generate, and interact with 3D worlds. World Labs is one of our portfolio companies here at Radical. We're actually both in the World Labs office right now, both Fei-Fei and myself. And Fei-Fei is a scientific partner at Radical. So, we're really thrilled to have her here on the podcast. Today we're going to talk about Fei-Fei's journey into the field of AI and to ImageNet. Why spatial intelligence represents the next frontier beyond language models. The technical breakthroughs needed to give AI true understanding of the physical world. How this technology will transform creative industries from gaming to architecture. And the path from research breakthroughs to building a company at the cutting edge of AI. Let's jump right in. So you talk a lot about North Star problems. Can you talk about how your north stars have evolved over time and track your intellectual and scientific journey through them?
First of all, great to be here, Molly. That's a great question and hopefully I'll take not too long to answer. I love chasing after a very tough problem, mostly scientifically that excites me and makes me feel it'll make a really big difference if we get it right or we solve it. But you know in the early times when I was a teenager this was a little more nebulous, right? I just loved physics. It's kind of a north star but I didn't have a specific problem to solve in physics. I loved it so much that I majored in it at college at Princeton and I love just learning about these audacious problems in physics like what is matter and special relativity, what's the origin of universe and all that. So in a way I was just excited by someone else's north star, the great physicist's north star. But gradually I realized one north star problem I became extremely interested in is the question of intelligence and the computational side of intelligence as well as what intelligence is. So I pivoted from my undergraduate study to PhD study on intelligence and specifically in this area my north star becomes solving visual intelligence. In the earlier part of my career from PhD to all the way through ImageNet, I wanted to solve the problem of understanding the visual world, to semantically label it, to be able to describe it. And that was north star for quite a few years, probably more than 10, 15 years. And then it brings me to my today that my scientific north star problem is spatial intelligence in a very simple language. It's the reasoning to build AI algorithms and models and applications that can reason, interact, can generate, can understand in 3D space or using 3D space because I think this is a fundamental problem for many of the current human workflows as well as for robots and storytellers and creators.
That's phenomenal and it leads me nicely to my next question which is really around spatial intelligence. So this is the founding concept of World Labs. Tell us more and specifically why are language models not enough? What's lacking in these incredibly capable frontier models that we have today that makes a new approach necessary?
First of all I think language models are great. I think they are actually a very, very essential part of intelligence, is essential part of AI. I just want to make sure that there's no misunderstanding there. Why is spatial intelligence so exciting to me? Well, I'm always inspired by human intelligence, in this case in the evolution of animal intelligence all the way to human intelligence, from a trilobite floating in the ancient water all the way to humans. The ability to be in a 3D space, to navigate it, to understand everything that's going on, to be able to plan your movement, to be able to do things in it, whether you are cooking an omelette or playing a football game or just dancing with your friends. A lot of this is what's in the heart of spatial intelligence, is the deep understanding, deep intelligence about 3D space. And what's even more fascinating is that humans share one 3D world for however long we can remember. But as AI lifts a lot of things into the digital world, helps us to create things in the digital world, we don't have to live and work and have fun in just one 3D world. We can actually create many, many 3D worlds. We can create it for learning, a fun sport game. We can create a new 3D world to tell a story. We can create a 3D world to share a travel experience. We can create a 3D world for socialization. So suddenly this problem of spatial intelligence can lead to many fun and productive applications and one world can become many, many, infinite number of worlds.
It's very evocative and as someone who's played with some of the early access models myself, I have had a lot of fun creating new kinds of 3D worlds and we'll talk about Marble a little later and some of the new things that you're seeing and seeing people create with the models. So, I'm excited to talk more about that. But before we do, it's been just over a year since the launch of World Labs, and I'm curious, you know, you've had this incredible career in academia. You've spent some time in industry. What convinced you that this was really the right time to make the leap and create a startup?
Yeah, great question. I'm sure people start startups for many different reasons. For me, fundamentally is the combination of how exciting the technology is as well as the potential of using it to help people and make products that can really empower people. I believe so much about AI is not to replace people, not to replace creators, is here to empower and augment people and augment creativity, augment different workflows, augment different use cases. And I think 3D spatial intelligence is an extremely powerful technology. If you look at just the trajectory of Gen AI in the past less than 10 years, really past five years I would say, we've had many breakthroughs. We've had breakthroughs in language models and with that we got breakthrough in speech generation, we got breakthrough in music, we now have, we're also looking at a lot of great applications and technology coming from multimodal LLMs. And then diffusion models for video generation as well as of course image generation. And if you look at that there is a missing piece, right? The missing piece is the world. And as we said earlier, not only it's a modality that's fairly underrepresented, it's also a huge business opportunity because so many people, so many industries from design to creativity to robotic simulation to AR, VR, mixed reality use 3D as a fundamental representation or need 3D for their workflows and products. So the combination of this need as well as just the technological opportunity to crack this problem of spatial intelligence is just so exciting. And my co-founders and I in a way just couldn't wait and we feel building a startup that focuses on this colossally big and huge problem would be so rewarding and so much fun.
Makes a lot of sense and I want to talk a little bit more about the technology. So, World Labs is a very particular kind of startup. World Labs is a frontier lab. You're working on deep science. You're not just recreating or scaling what's been done already, what's already worked. What are the scientific advancements necessary to make world models a reality?
Well, there are many and at this point we are focusing on being able to have a model to generate 3D worlds that respects geometry, that users can traverse and navigate in this world and we can interact with it. That requires 3D representation which brings in deep problems of where is the data and what's the best model to represent this and how do you render this? It's actually a whole family of problems that has not been solved. And every day we feel like we're at the very edge of what has been known and we have to push that envelope in order to make a product or application that's usable for users.
And you just shared a blog post yesterday sharing an update on this very thing, the thing that you're putting in front of users. So I want to hear you talk a little bit about Marble. So for those who haven't been following, the World Labs team just shared a blog post yesterday sharing an update on their work and released limited access beta to their model which is called Marble. I also want to hear the story about how you picked the name Marble too because I'm just personally curious and it's a nice name. It's incredibly impressive and lucky to be able to play with it a little bit. There's a stylistic coherence and a geometric consistency that just isn't possible with generative video. So I'd love to hear you talk more about Marble. Tell us more about it.
Yeah. So, you're referring to yesterday we opened a blog that we want to share our technological progress with the world because I know that it's such a fun period where companies and teams are just so eager to share the science and technology and we want to be part of the ecosystem. And in the past year we have been working so hard to make the generative 3D model work that it's still not perfect, but we want to share with the world. So, we wrote this blog and you can find it on our website worldlabs.ai, but we don't want to just share a blog and cherry-pick some examples. We actually want to invite people to experience what the model does. In order to do that, we actually have to build this viewer for our users to be able to experience using this model, which is you come here and you prompt whether you use an image or use a text prompt and then after a few minutes the model will return you a 3D world. And so we built this viewer that shows you the model creation and we needed a name and we think Marble is a great name because it's a nickname for our world perhaps. It's also a material that sculptors and artists use to create something historically, right? We have incredible marble sculptures throughout human civilization, masterpieces. So this interplay between the material itself that allows for creation and interactivity as well as a nickname for the world. We just like this name.
I like it too and I love the double entendre. It is really cool and wonderfully evocative. Maybe just briefly to talk a little bit about the decision to share the work externally. So much in AI right now is not widely shared and there is to some degree less building in public perhaps than has been done in the past. How did you think about that? Like what's your ethos at World Labs in sharing your work and sharing the progress that you're making? How do you guys think about it?
