CEOInterviews.AI
Start App
Get an email when Fei-Fei Li gives a new interview
One email per new interview. Unsubscribe in one click.
Fei-Fei Li
Co-Founder and Chief Executive Officer; incoming Executive Vice President and Chief Scientist, AMD, World Labs

Fei Fei Li | 2026 Frontier & Pioneer Symposium

📅 Aug 21, 2026 Asian American Scholar Forum 53 MIN 114 SEGMENTS · 7 SPEAKERS
In this keynote fireside chat, Fei-Fei Li, Founder and CEO of World Labs and Sequoia Professor at Stanford University, discusses ...

What Fei-Fei Li said

Written from the verified transcript and checked against it. Every figure links to the moment it was said.

Fei-Fei Li discussed her perspective on AI as a 'messy middle' between extreme rhetoric, emphasizing the need for nuanced communication and responsibility among those with a platform. She reflected on her journey from ImageNet to founding World Labs, noting 2026's acceleration in spatial and physical intelligence, and identified data as the harder challenge compared to model development. She proposed a three-tier functional taxonomy for world models: rendering, simulation, and planning, with World Labs focusing on the simulation layer. Li highlighted healthcare and caregiving as critical applications for robotics, and argued for human agency in using AI as a tool. She discussed the difficulties of robotics data collection compared to language, compared driving to consumer robotics, and stressed the importance of working with young technologists, fearlessness, and accepting uncertainty in entrepreneurship.

Key takeaways

  1. Data is harder than model development for physical AI, and remains underappreciated more than a decade after ImageNet.
  2. World Labs focuses on the simulation layer of world models, which includes rendering, simulation, and planning.
  3. Consumer robotics lacks mature hardware and business use cases, making data flywheels difficult, unlike driving data from Tesla and Waymo.

Numbers and commitments

FigureWhat it refers toTypeAt
2012 Year of the ImageNet moment timeline 9:06
2024 Year World Labs was founded timeline 12:21
$15,000 to $30,000 Benchmark cost for caregiving robots price 14:51

Chapters

  1. 0:00AI's messy middle and responsibility
  2. 7:22Journey from ImageNet to World Labs
  3. 17:22World models: rendering, simulation, planning
  4. 29:00Robotics data challenges compared to language
  5. 33:02Confronting dystopian AI rhetoric and cyber risks
  6. 38:28Immigrant journey and fearlessness
  7. 44:11Human agency in the AI era

Questions asked in this interview

12
  1. 3:22I do want to talk with all of these great thinkers, students, and leaders." What do you think is the most important message on today, on this day, on this Saturday?
  2. 5:01What are some of the potentially the thoughts that we need to entertain about the messy middle?
  3. 10:10How is the answer to the question that you're trying to solve at World Labs faring?
  4. 16:56Did I cut you off, though, on what you wanted to reflect on?
  5. 21:43Don't you want her for hours?
  6. 23:41How do you tell them to look at all of this and negotiate through that as they're making the decisions for the future to work with you or any other organization?
  7. 28:28And how much higher is that difficulty that you're trying to solve for right now?
  8. 35:41Who would one or two of those be, and what did they teach you?
  9. 40:56But in the beginning, was it like that, and what did you say to yourself?
  10. 44:59And that is to you on what you—you've been asked this before, but I love to ask this question, too: What do you wish young Fei knew that you know today?
  11. 46:54And so what's your hack to the loneliness?
  12. 48:26What else can bridge the gap, and do you have any timeline to release the product?
Kai Li 0:00 ↗
Next session is another fireside chat. And for this one, I asked both the keynote speaker and the host, how do I introduce them? They told me you can introduce any way you want. Okay, so here it goes. For Fei-Fei, it's my great pleasure to introduce her because she is my long-term friend, colleague, and collaborator. And she's currently a professor at Stanford, and she co-founded the Stanford Institute for Human-Centered AI when many people didn't know human-centered is really important. And she's a member of National Academy of Engineering and National Academy of Medicine. She's a founder and CEO of World Labs. And just very briefly, I think I will just share with a secret story that you don't know about Fei-Fei because I'm long-term friends with her.
I think what she distinguished herself from many people, she's a visionary. You know, when she and I start talking about ImageNet, her vision, and she wrote a proposal trying to get funding, didn't get funding. She asked me to help do the next version. We also didn't get funding. Basically, the reviewer said they couldn't believe those well-known professors working on this, they should be working on algorithms. And what they didn't know is what come out of ImageNet have changed the world, played a big role, and how much trouble he has caused. That's what I will do with Fei-Fei. It is also my great pleasure to introduce Richard. He's an award-winning journalist, author, filmmaker, longtime anchor, and correspondent of CNN, NBC News, MSNBC, and MS Now. He's been a wonderful friend and support of AASF. And fortunately for us, he keeps agreeing to come back. So please join me to welcome both Fei-Fei and Richard.
Richard Lui 2:39 ↗
Well, thank you, Kai Li, as always. It's great to be here at Stanford, as well as AASF. And look who's here.
Fei-fei Li 2:46 ↗
Same.
Richard Lui 2:47 ↗
Wow. I myself am a little starstruck by the number of cameras, Fei. I think there are others who feel the same. And that was such a great introduction by your father.
Fei-fei Li 3:01 ↗
No, just—
Richard Lui 3:02 ↗
I'm joking. They're not related.
Fei-fei Li 3:04 ↗
It's a joke.
Richard Lui 3:05 ↗
It's a joke. They're very good friends. They do share the same last name. But it's an important relationship, the friendships you build in your career as well as outside of your career.
Fei-fei Li 3:18 ↗
Yeah, I'm very grateful. Yeah, Kai and I have known each other for decades.
Richard Lui 3:22 ↗
And this is why AASF is so important that we sit here and be able to have a conversation with you. And, Fei, when you are asked to speak all around the world, I have to ask you what makes you consider, well, yeah, I'll go, and number two, this is what I want to say. So today you said, "Yes, I do want to sit down. I do want to talk with all of these great thinkers, students, and leaders." What do you think is the most important message on today, on this day, on this Saturday?
Fei-fei Li 3:53 ↗
Well, great question. I always say that I'm the most boring AI speaker because—
Richard Lui 4:02 ↗
Not true.
Fei-fei Li 4:04 ↗
You haven't heard what I'm about to say. So, AI is full of extreme rhetoric right now. The louder the polarized voices, the more airtime they get. But I'm a scholar. I'm a scientist, and also I'm now an entrepreneur. AI is the messy middle. As a technology, it's a double-edged sword. As a civilizational technology, it not only is incredibly powerful, it's also butting against the most messy of our world, which is our own humanity. And when you have a powerful technology meeting the messy humanity, you have everything in the middle. So if you ask me what's my voice, I want to speak as a scientist and entrepreneur from the middle, not from the extreme sides.
Richard Lui 5:01 ↗
And you've been very consistent with that. And if you have done your homework before arriving today, you'll know that Fei has consistently—and up to today, I was looking at some of the other interviews you've done—you're a pragmatist. And yet we are pushed, and I know that in my business we are often talking these extremes. How do you help us understand that messy middle? What are some of the potentially the thoughts that we need to entertain about the messy middle?
Fei-fei Li 5:32 ↗
Behind closed doors, I talk to my journalist friends about the need of self-education about what this technology is about, and also on camera, that sense of responsibility. Today, 2026, the global public—not just US, global public—needs responsible conversation and nuanced communication about how to use this technology. And I consider that a responsibility for anyone who has a following, anyone who has a megaphone. I think journalists still have a megaphone, and I'm lucky that sometimes I have this speaker, and I take it very seriously. Especially we're in the heart of Silicon Valley. Man, I love Silicon Valley, but Silicon Valley is about saying the most outrageous things, and I think it's very important we're more nuanced than that.
Richard Lui 6:38 ↗
Thank you for being nuanced. And it's a voice, it's a beacon, it's a tentpole, very much needed. And one of the ways that you're doing that, as we have moved amazingly fast thanks to your work, the godmother of AI with ImageNet. And when I've heard your name over the years, and me not being an AI specialist, but having to self-learn as you were just describing, it's really amazing. And I'm sure when you started that journey, did you know that we would be here today? If you go back to the ImageNet mindset that you had at that time, working hard, and now you see where we are in 2026, is this way above, way farther than you thought we'd be?
Fei-fei Li 7:22 ↗
Yeah, I get asked that question a lot. In fact, just two days ago, I was offstage with my good friend Geoff Hinton. And you know, those of us who started decades ago in AI, we really were a group of scientists chasing a curiosity that was not fashionable or hot or guarantees jobs. It's very different from this generation doing AI, and I'm not passing judgment. I'm just telling my own story that I started in AI as a scientist just being curious about how to make thinking machines, what does intelligence mean. So of course, for example, you talk about ImageNet. ImageNet, I understand, is one of the most prominent work in my career, but it's not the only work.
My entire career is after this problem of visual intelligence, and now spatial and physical intelligence. All of this is rooted in the deep belief that there is a frontier of knowledge and innovation that as scientists it's our responsibility and frankly our enjoyment to push that frontier forward. And when you do that, when you do that out of love, out of curiosity, you don't think about, is it going to be a big deal out there in the VC world or in whatever. So from that point of view, early years, I did not expect 2026 or whatever the moment we're in will be this huge. I did not expect.
But I did after the ImageNet moment. Many people think 2012 was the ImageNet moment. And then I became Google Cloud Chief Scientist, 2017, 2018. I think that was a personal inflection for me at that moment, seeing machine learning and AI potentially impacting every vertical business because I was at Google Cloud, so I get to see vertical businesses. I think it did dawn on me that this is no longer a personal curiosity; this is a civilizational technology. It changed me from a scientist that just was chasing curiosity to a scientist that recognized my own societal-level responsibility.
Richard Lui 10:02 ↗
Do you want that responsibility?
Fei-fei Li 10:07 ↗
I think that's what responsibility is. You don't have a choice.
Richard Lui 10:10 ↗
Yeah, exactly. When I think of what you've just said, Fei, and we move forward to your next big question that you're trying to solve at World Labs, how is that question faring? How is the answer to the question that you're trying to solve at World Labs faring?
Fei-fei Li 10:32 ↗
Yeah. So the question we're trying to ask here is, how do we push the frontier of AI beyond language intelligence into spatial and physical world? Why is this important is because evolution has told us a story that so much of the development of intelligence systems from animals to humans rely on spatial and physical intelligence. In fact, this intertwined relationship between perceiving and interacting is the story of evolution. Most of evolution's development of intelligence is preverbal. You know, before humans as an animal, there isn't much language in the animal world. Yet we see incredible ability in navigating and manipulating and changing the physical world based on animal intelligence. So that capability would be incredible to unlock.
Think about all the potential scientific discovery work. Think about all the potential help humans can have, whether it's natural disaster relief, or exploration of dangerous places, or manufacturing, or agriculture. And personally speaking, every day I am a caretaker of elderly parents and sick patients, and I understand that incredible physical toll that humans and caretakers go through. So unlocking spatial and physical intelligence would be incredible for many, many vertical use cases. So how is it going? I think the progress is accelerating.
When we started World Labs back in 2024, we were a little over a year after the ChatGPT moment. So the whole world's attention was in language models. But today, fast forward to 2026, everybody and their aunt is talking about world models, to the point that two months ago, I was like, "Wow, we really need to define world models because everybody was talking about it in their own definition." So I wrote a technical blog about how I see the functional taxonomy of world models. So it's a long way of saying that the progress is accelerating. The problem is extremely daunting.
Compared to language modeling, world modeling, spatial intelligence, physical intelligence, it comes with its own challenges that language doesn't have. To start with, data is a challenge. Second is that the world is extremely vast, whether it's virtual or physical. And on top of that, if you want to apply an interactive change to the physical world, then we're talking about the actuation and the interface between hardware. So these are extremely challenging technical problems that go beyond language models.
Richard Lui 13:49 ↗
If you can, for us, level of difficulty and level of importance to the desired outcome that you're trying to solve for here, looking at the model opportunity and the data opportunity—you alluded to both sides of that—when we're thinking about physical AI, which has the higher difficulty: the model development or the data side of physical AI?
Fei-fei Li 14:17 ↗
It's very hard to answer this question, Richard, because they're so intertwined. But I understand that as a journalist, you are pushing for a crisper answer. So if you force me, I would still say it's data. It's much harder to get that kind of data. Data is, you know, it really makes me chuckle that more than 10 years after ImageNet, data is still very much underappreciated in AI.
Richard Lui 14:51 ↗
Yeah. As an outsider and as a journalist, a small practitioner, that's the way I see it, too. I'm not an expert like you, and the whole world doesn't want to listen to me; they want to listen to Fei-Fei. You brought up caregiving for a moment. You and I are both caregivers, and I saw that in your background. And one of the things I am fascinated with—and this comes out of conversations at Davos in healthcare—and that we were saying, you know, if we can solve this physical AI data set problem in caregiving, potentially we would have robots, $15,000 to $30,000 was sort of like the benchmark that they were talking about. We can have effectively four caregivers working full shifts if we can just figure that out. And you and I know that that would be amazing.
Fei-fei Li 15:39 ↗
Yeah, I think so. I think one of the most important applications in robotics, in my opinion, is healthcare. First of all, I don't think there is a single country in this world that has excessive nurses.
Richard Lui 15:55 ↗
Yeah, that's right.
Fei-fei Li 15:56 ↗
Right. If you look at nursing and caregiving, healthcare caregiving profession, this is a highly critical, yet highly desired population. And American nurses, at least I know, are very fatigued.
Richard Lui 16:16 ↗
Yes.
Fei-fei Li 16:16 ↗
There is way too much work, and every shift, the demand is too much. And then home care is an even messier problem as we have longer life expectancy. And this is where—you know, I'm a working mom and an entrepreneur—it impacts women a lot.
Richard Lui 16:41 ↗
Mhm.
Fei-fei Li 16:41 ↗
And this is disproportionately—
Richard Lui 16:43 ↗
Unaccounted GDPs.
Fei-fei Li 16:45 ↗
You know, the amount of work that—
Richard Lui 16:48 ↗
About $3 trillion here in the United States.
Fei-fei Li 16:50 ↗
Yeah, okay. I didn't know that. Thank you for that number.
Richard Lui 16:53 ↗
$11 trillion globally.
Fei-fei Li 16:55 ↗
Yeah.
Richard Lui 16:56 ↗
And we agree on this point for sure. And I can see we have some folks that are lining up for questions. We will have that at the end. Did I cut you off, though, on what you wanted to reflect on?
Fei-fei Li 17:09 ↗
No. No.
Richard Lui 17:11 ↗
Okay. You know, this idea of world models, is it one or the other when we think of LLM versus world models, or are they working together? Does one take over the other?
Fei-fei Li 17:22 ↗
Yeah. Okay, so let's dive deeper. What is a world model? Clearly, in fact, in my tech blog, I alluded to even back in ancient Greece that defining the word "world" becomes contentious. Which elements, what does it mean? So I would like to look at it in a functional taxonomy way, which, in a three-tier definition of world model, is what we call rendering, simulation, and planning. What does that mean? Rendering is really about a model that can show you what the world appears to be. And that is very much what we see today: Sora, a lot of the video generation models that give you beautiful pixels of moving or static pixels that shows an astronaut is riding a horse, there's an avocado chair. These are early renderings of the world. So that is one layer of world modeling. It focuses on giving you what meets the eye. The consumer of the rendering aspect of world model tends to be humans. Humans get to enjoy these pixels, and some humans get to manipulate these pixels. So that's one level.
The second level is simulation. Simulation goes deeper than rendering. Here it's not just about beautiful pixels and entertainment; it's about are we capturing the physical and dynamic reality of the world. For example, when I simulate the world that's in front of Richard and me, I would like to know, how does this bottle move? If I pour the water, how does it move? So this goes beyond just putting pixels on the screen. It's about how do I faithfully simulate the physics and the dynamics and the movements and the structure of the world. Obviously, this is related to rendering. You can say if I have perfect rendering, you might get to very good simulation, and vice versa. But the consumer of simulation is likely both humans and machines. For humans, for example, gamers, we love playing simulated games, right? But machines, for example, using CAD models to design machines or moving robots in simulation, the user might be machines and programs themselves.
The last but not least is planning. This is where not only we want to understand the structure of the world, it's that when we put in action, the world model predicts what's the next step, what do we do next. And this has direct implication to self-driving cars, to robotics. And these three layers are interconnected. As I said, if you have perfect one thing, it might actually lend itself to a very high quality another thing. What we are seeing now in this current trend is that these three functions, or modeling of these three functions, are starting to merge. And so even at World Labs, we are really looking at the linchpin of these three layers, which is the simulation layer. But our model can render very well, and also the model architecture potentially can lend itself to planning as well.
So it's a rapidly evolving field, but it's also a very exciting field. And it also does interface with language. I'm sure a lot of you are seeing that using language models, we can drive traditional simulators to do a lot of cool things. So I don't want to geek out too much, but I can speak for hours.
Richard Lui 21:43 ↗
Don't you want her for hours? All right, well, here we are. Let's just—
Fei-fei Li 21:50 ↗
Kai would not be happy.
Richard Lui 21:52 ↗
I'm just going to unbutton Fei-Fei, and we'll just hang out here. What has surprised you and made you smile as you've been working at World Labs? You know, it's early on, right? But I'm tapping into your love of curiosity, your sensibility of different segments and communities, countries coming together and doing things differently. What has surprised you in the journey? And it's early on, I know, at World Labs, but there's always something, "Wow, that is so cool."
Fei-fei Li 22:23 ↗
I can tell you hands down what makes me smile and wake up is working with these incredible young technologists who just have so much passion and craftsmanship to unlock the technology challenge at their hand. Being a builder is in my heart. And the privilege I've had in my entire career is that the people I work with, they don't age. They're always young. They're always in their 20s and their 30s. I love, love, love that. And I do age, but working with the young people, the builders, the engineers, and scientists is by far my biggest privilege, and it really does energize me. I go to World Labs every day. We have our box lunch, we're a scrappy startup. But no matter which hour I wake up, Slack is always up. There's someone posting on Slack and—
Richard Lui 23:35 ↗
I love it, too.
Fei-fei Li 23:36 ↗
That's great. Yeah, so it just makes me smile.
Richard Lui 23:41 ↗
You know what you're reflecting on here, Fei, is actually who you are, too. The 20-year-old face and energy is what you have, what Kai has, and so many others who are part of AASF. And that's what I think is so redeeming about being in this community is you never lose the energy, the verve, the love of trying to solve the big question. When you are trying to pass that on to those who you are working with, you're hiring your students, how do you teach them this sensibility that you've described? Because clearly it's more than two different ends of answering the question for science versus the economic question. There's a lot of gray space in between that, too. We seem to have potentially leaned too much on the economic outcomes in some subsectors of this sector. And on this side, has this voice been lost, this curiosity of creating the big solution? How do you tell them to look at all of this and negotiate through that as they're making the decisions for the future to work with you or any other organization?
Fei-fei Li 24:53 ↗
That's a very interesting question, Richard. I don't think they need to be a dichotomy. And it depends on the role, right? For example, if we are squarely in a university setting and one is a young PhD student or graduate student starting at the beginning of their scientist career, it's a very highly protected environment. There isn't much of an economic drive in that phase, and then you tend to lean towards curiosity. But in a startup setting, I think that healthy balance between curiosity and economic drive is very important. In fact, it's not about making money for ourselves; it's about providing economic value to your customer and user. And to really unlock that, it takes curiosity, too. It takes the curiosity for what customers need, what customers' pain points are. So I don't pit them against each other. I actually think, like you said, there's so much in the middle. These energies can flow together. And it's very important that today's AI provide economic value because we together collectively have painted such a big story for AI, and I do believe in that big story. But I think when the rubber meets the road, that big story has to translate into true value for individuals as well as businesses.
Richard Lui 26:33 ↗
Yeah. And a way to kind of understand that in industry as well as in your sector, as well as the economic parts of what our discussion has been so far, you and I were talking about Davos, and I would just walk down the Promenade or go through the sessions, and I would see "AI plus apples" or "plus bananas." I mean, they would attach it to just about anything. And for me, that was always an indicator of the complexity of the way we understand what AI is at the boardroom level. I mean, at the boardroom level, we still use the word, as you know very well, "cyber," and "cyber" is a word from the 1950s. So when you look at that dynamic of how do we understand that gray better, what would be the lesson for boardroom voices that still, I believe in many cases—if not all—AI plus bananas, right? That's, I don't want to see those titles at Davos anymore, basically.

50 more exchanges in this transcript

Sign in free to read the rest of this interview. No card required.

Sign in to read the full transcript

Cite this transcript

APA, MLA, BibTeX
APA

Li, F. (2026, August 21). Fei Fei Li | 2026 Frontier & Pioneer Symposium [Interview transcript]. Asian American Scholar Forum. CEOInterviews.AI. https://ceointerviews.ai/interview/1295606/

MLA

Fei-Fei Li. "Fei Fei Li | 2026 Frontier & Pioneer Symposium." Asian American Scholar Forum, 21 Aug. 2026. Transcript, CEOInterviews.AI, https://ceointerviews.ai/interview/1295606/.

BibTeX
@misc{li2026_1295606,
  author       = {Fei-Fei Li},
  title        = {Fei Fei Li | 2026 Frontier \& Pioneer Symposium},
  howpublished = {Interview transcript, Asian American Scholar Forum. CEOInterviews.AI},
  year         = {2026},
  month        = {aug},
  url          = {https://ceointerviews.ai/interview/1295606/},
  note         = {Speaker-attributed transcript with timestamps}
}