Welcome to McConomy Originals. Original stories of founders, brands, and global business power.
I hear you love walking talks.
Yes, because we sit too much.
Yeah. And also our company is just next to this park, so it's great.
Is it all AI talk all the time here? I don't know other people's talk, but I talk AI all the time, but that's the only thing I know.
Fei-fei Li is a giant in the field of computer vision, a branch of AI dedicated to teaching machines to see. She made a breakthrough with ImageNet, a visual library that sparked the modern AI revolution, earning her the title godmother of AI. And she's been a pioneer at AI center ever since as a Stanford professor, a Google executive, an adviser to US presidents, and a relentless advocate for AI that serves people.
I feel the leaders of AI are not talking enough about empowering humanity. They're talking too much about replacing humanity, replacing jobs.
Now, she's adding startup co-founder to her star resume. What's it like running your own shop?
While the rest of the AI world doubles down on large language models like ChatGPT and Claude, Li is betting on a new frontier, world models. She's building AI that aims to predict what happens next in the real world, not just the next word in a sentence.
So, what can world models do ultimately that LLMs will never be able to?
Can words put down fires? Can words cook an omelette? Li and her startup, World Labs, is competing in a crowded field of rivals, all making the same bet that world models are the next big leap in AI.
Did you ever imagine AI would be this big?
Have I imagined the entire civilization is going to be redefined by AI? The cultural shifts, the geopolitical landscape. I would not have imagined that.
Before she became a leading AI scientist, Li was just a curious kid looking at the world. What was young Fei like as a kid?
It was 80s and 90s in China. My city was Chengdu. My dad is quite a curious soul about nature. So, I got quite a dosage of chasing after bugs and running after puppies and going to the mountains to draw and sketch with my dad and, you know, reading.
Cutting your hair short, loving aerospace and physics. Did you see yourself as a rebel?
As a kid, right, you're pre-teen. You almost have to be a rebel just because you want to make a statement for yourself. But the love of nature and science and physics, the love of aerospace and outer space, those are part of my identity.
You immigrated from China to the US in the '90s.
What was it like coming here as a teenager back then?
That was tough. Shifting the whole scene, culture, language, life as a teenager is perhaps one of the hardest times because you're finding who you are and then suddenly you don't even know the world, right?
Li learned English, finished high school, studied physics at Princeton while running the family dry cleaning business on weekends, and followed her dreams to Caltech where she earned her PhD and became a computer scientist.
You've been studying artificial intelligence since it was super niche. What was the field like back then?
That was a beautiful period. It was really pure curiosity because there was no money, there was no power, there was no fanfare, there's no hype. We were there just the same as scientists look into the sky, go down in the ocean. I was deeply curious.
How does AI see and how is it different from how you or I see?
We use our eyes to drive the sensory system to turn that into signals in the brain and the brain does computing. It's not that different for computers. We need to learn the pattern of the world and understand objects, colors, environments. A lot of seeing is preparing humans to act. You wake up this morning, you hug your kids, you go downstairs, prepare some breakfast, grab your key, drive to your work. Everything I've described so far, you've done that with your spatial intelligence.
So, it's got to be hard to get computers to do that.
Li was working on this back in 2006. She built ImageNet, a massive catalog of 14 million pictures in over 21,000 categories, the largest assembled at that time. Because Li saw that algorithms needed data to get smarter.
So what was the spark for ImageNet?
Nobody was paying attention to data. My students and I had that epiphany that learning needs to be driven by data.
In 2010, Li turned ImageNet into a competition. Who could build the best algorithm to correctly identify the most images? In 2012, Geoffrey Hinton's University of Toronto team entered with AlexNet, an algorithm powered by Nvidia graphics cards. The result marked a turning point. That combination, massive amounts of data, neural networks, and GPU computing power, became the golden recipe for modern AI and cemented Li's legacy.
What do you feel when you look at this?
It's not the eight of us. It's the entire humanity that's really the main story here because this comes from the photo of what about 100 years ago when industrialization was in a booming phase and big urban skylines are being built. That was a civilizational moment.
It's amazing that you're being recognized here, but like could they have centered the photo?
Come on. It's true cuz there's more space here.
I asked you about being the godmother of AI a few years ago.
How do you feel about that title?
You know, Emily, I would never call myself godmother of anything. Remember, Emily, when you asked me, I was like, 'Oh, I was taken aback because I don't naturally think about myself as godmother of anything. That's not my personality. I focus on work.' If I rejected that on that spot, we once again would be in a situation where women just don't get recognized the same way as men. I want more women to be called godmother of whatever creation, innovation, discovery that they have done.
So, do you embrace the term now?
Embrace is a big word. I don't go around and say, 'Hi, I'm Fei. I'm the godmother of AI.' But people have adopted that term thanks to you.
Artificial intelligence had long been discussed in policy and tech circles as something perpetually right around the corner until ChatGPT. AI went from a niche research field to the most talked about technology on the planet. Large language models, image generators, video generators. Suddenly, it seemed like everyone was or wanted to be in the AI business. And let's be honest, things have come pretty far.
Li saw an opportunity and launched her own startup in 2024.
What's it like running your own shop?
So, here's the thing. I'm by far the most senior in the company. From my co-founders to my engineers and researchers and scientists, they're so talented, but they're also young. I do feel like a mom and often times I feel like a tiger mom. I do run this shop as I would say high standard.
Li is working on advancing world models. If ImageNet taught AI how to see, this technology aims to guide AI through the physical world because describing how to catch a ball and actually catching one are very different things.
And super intelligence, she's betting that lies beyond chat bots. The rest of the AI industry is focusing on LLMs.
What can't LLM do? If we want to create a world where we continue to push scientific discovery, where we want machines like robots to be a partner to us, all this language cannot do it alone. Building world model and spatial intelligence is not about LLM. It's about the next frontier and the next chapter.
So what exactly is a world model?
World model at this point is overloaded term. We look at spatial intelligence as three different kind of functions. One is rendering, one is simulation and the third one is planning. Rendering is the world model outputs beautiful pixels for humans to consume. Most likely you're thinking about Sora. The second class of world model is they do simulation. But it doesn't serve just humans. It serves machines. They are focused on capturing the actual structure of the world, the geometry structure based on physics. The third kind of world model planning helps you to know what is the next thing you need to do. For example, if you have a cup, a world model that serves as a planner will let the robot know pick it up and put it somewhere else. And that is a world model that is very closely coupled with robotics.
Marble is World Lab's first step towards creating a world model. It's a platform that lets anyone generate an explorable, editable 3D world from a single visual or text prompt.
Now this is a showcase page of some user-generated worlds. You can see they created a room, a beautiful room. Now you look at this, you're like, 'Fei, this is no different from a picture.' But here's the thing. It actually is a world because I can navigate in it. So like a game, I can move forward. I can turn around. I can go anywhere. Here it's a fully consistent 3D world.
So who or what companies are using Marble right now?
So why is Marble useful? For example, virtual production in movies, people are using Marble because they can shoot actors in any environment. Game developers, Marble has completely lowered the amount of resource and time one needs. We also have robotics example. This is a collaboration with Nvidia where they're using Marble environments to augment the training of robots.
What are you training these models on that others don't have?
We are training these models on two things, right? One is specially prepared data. Pixel data are more nuanced than language data in a sense that there's more information in a picture. There's camera information. Another one is algorithmic and architectural innovation because no one has created generative world models as we have in terms of creating 3D and eventually 4D worlds.