Fei-Fei Li

1. Key Themes

Spatial Intelligence as the Next Frontier of AI

WorldLabs is positioning itself as a frontier model lab focused on a new category of AI — spatial intelligence — distinct from language models. The claim is that understanding, generating, and acting within 3D spaces (physical or virtual) is the next major unlock, analogous to what transformers did for language.

"WorldLab is a two year old startup. I think we should just recognize it's a frontier model lab. We are building the next frontier of AI, which is what we call spatial intelligence. And spatial intelligence is about creating AI that has the ability to generate, understand, reason with, and interact with spaces, whether it's physical or virtual." — Fei-Fei Li 00:01:49

The Real-to-Sim-to-Real Pipeline as a Core Infrastructure Bet

Scenics, now part of WorldLabs, is building a pipeline where real environments are digitized into physics-consistent digital worlds, and those worlds are used to train and evaluate robots at scale — then transfer back to the real world. This is not just academic; they are already working with customers near deployment.

"We are developing what we call a real-to-sim-to-real pipeline. We're going to map the real environments into the digital world that has the best alignments with the real environments. By alignments, we mean that whatever happens in the digital world is also going to happen in the real environment. Such that we can replace all the data, all the evaluation we need in the real environments by using the data that can generate at a scalable way in our digital world." — Yunzhu Li 00:05:19

Simulation's Unique Role: Counterfactual Reasoning

The argument for simulation is not merely about data scale — it is about generating data for scenarios that cannot or do not happen in the real world. This is a philosophically distinct and stronger argument than "simulation is just cheaper data collection."

"There's a very important role simulation plays that real-world data doesn't play, which is counterfactual reasoning. It's that you play out events that hasn't happened or cannot happen or you don't have enough data to make it happen in real world. And while you play it out, you learn how to act in it." — Fei-Fei Li 00:20:06

"Waymo has officially said they use billions of hours of simulation. And actually, Waymo is more simulation heavy than just real-world data heavy. So these are real examples. And as you know, Martin and Yunzhu too, cars are the simplest kind of robots." — Fei-Fei Li 00:21:04

Simulation Enables Both Reliability and Efficiency — Two Distinct Value Propositions

Yunzhu Li articulates two separable and concrete business values that simulation provides to robotics customers: systematic coverage of state space for reliability, and the ability to train robots to operate faster than human speed for efficiency.

"For reliability, if you're thinking about a robotic system working reliably in the real environment, you need data to provide systematic coverage of all the state space and the variations that robots might encounter... With simulation, you can do systematic randomizations and control and variations of lighting, frictions, geometries, object types... Second is about efficiency. So right now, many people are doing teleoperation... the data at a speed that is actually slower than humans actually doing the task. But for many of our clients, human speed to them is not good enough. They want faster than human speed." — Yunzhu Li 00:21:57

Robotics Will Follow a Structured-to-Unstructured Progression

The path to general-purpose robots goes through fully structured environments (car factories, already automated for decades), then semi-structured (Amazon warehouses, restaurants, hotels), and only eventually to unstructured environments (homes). The pragmatic near-term opportunity is in the middle tier.

"If you look at all the progressions of robotic applications in the real environments, it has always followed the trend from going from fully structured environments into semi-structured environments and then into unstructured environments... it's so much easier and more approachable, at least like right now, to focus more on the semi-structured environments before we move on to fully unstructured environments." — Yunzhu Li 00:30:09

Evaluation Is the Overlooked Bottleneck in Robotics

Evaluation in robotics — knowing how well a model checkpoint performs — is orders of magnitude slower than in language models because atoms have to physically move through space. Simulation-based evaluation is a direct unlock for iteration speed.

"If you really think about also the robotic evaluations right now people are doing in the real environments, the iteration speeds is multiple orders of magnitude slower than iterations of those language models... Not only is it slow, it's dangerous, it's costly, but at the same time the speed is also like multiple orders of magnitude slower." — Yunzhu Li 00:24:35

Embodiment-Agnostic and Model-Agnostic Infrastructure Is the Moat

WorldLabs/Scenics is deliberately not building a specific robot or a specific model. The bet is on being the environment and infrastructure layer that any robot company can plug into — regardless of hardware form factor or software stack.

"What we are building, you can imagine, is an infrastructure with the softwares around these infrastructures for people to, for them, build worlds such that robots can learn and evaluate. And this infrastructure is naturally model agnostic and embodiment agnostic." — Yunzhu Li 00:27:46

The Multiverse Business Model: Physical and Virtual Spaces Are Co-Equal Markets

WorldLabs' thesis is not exclusively about physical robotics. Creative fields — VFX, gaming, design — are equally valid addressable markets for spatial intelligence. This positions WorldLabs as a platform play across both physical and virtual world interaction.

"WorldLab's thesis has always been that the world we live in can be multiverse, that we create technology to allow people, builders, developers to act within different spaces... Look at the creative field, whether it's VFX or gaming or design, many use cases, you can create and act within virtual spaces." — Fei-Fei Li 00:03:07

2. Contrarian Perspectives

The Humanoid Robot Hype Is Premature — Specialized Bodies in Semi-Structured Environments Win First

Against the mainstream excitement about humanoid robots operating in homes, the argument here is that humanoid form factors are optimized for evolutionary generalism, not economic specialization. Purpose-built hardware for constrained environments is the right near-term bet.

"From a business point of view, from a pragmatic technology point of view, this unstructured environment and a generalized body is actually the hardest problem to solve. It's not necessarily even the right way to solve the problem. It's we specialize. So we take more specialized body to solve a narrower problem." — Fei-Fei Li 00:31:55

Video-Only World Models Are Insufficient for Robotics

The dominant industry approach of using video prediction models as the foundation for robot learning is critiqued for lacking physical consistency. Objects disappearing mid-interaction is not a stylistic flaw — it is a fundamental failure mode for robot training.

"Imagine if a robot is pushing an object forward, the object just magically disappears, which has been a problem on many of the existing video prediction models. It's one provides good enough signal for the robot to know, like, what is the right thing to do." — Yunzhu Li 00:13:33

Robots Must Work Out of the Box — The LLM Error-Tolerance Model Does Not Transfer

The widespread practice of shipping AI products that require human review of outputs is not viable for robotics. This is a non-obvious constraint that makes robotics fundamentally harder and creates a higher reliability bar from day one.

"That is very different from how people will be using, for example, robotic models. Because for robotic models, out of the box, the robots have to work reliably in the real environments... we don't even have all the necessary infrastructures around those for the robots to just out of the box work reliably in the real environments." — Yunzhu Li 00:35:56

Simulation vs. Real-World Data Is a False Dichotomy

Against the view (attributed to Sergey Levine) that simulation always deviates from reality and real-world data is essential, the argument here is that physics-based simulation and learning-based models are on a continuum, and the right system evolves from physics-heavy to data-heavy as real-world data accumulates.

"They don't contradict with each other... Maybe at the very beginning, we have stronger emphasis on we have more physics to making sure we have the right consistency and right structure for us to learn the world... But as we accumulate more and more data... we'll have the data that will be moving towards more learning-based, like modeling of the environment." — Yunzhu Li 00:19:05

Human Brain Power Efficiency Is the Real Standard — and AI Is Far From It

Against the narrative that AI has achieved or is approaching human-level capability broadly, the point is made that the human brain runs on 30 watts and AI systems are nowhere near that efficiency even for language, let alone physical tasks.

"Even LLMs does not have human brain efficiency. Human brain operates on 30 watts. So we are far from that." — Fei-Fei Li 00:34:50

3. Companies Identified

WorldLabs

Frontier AI model lab focused on spatial intelligence and large world models. Building the Marble base model for 3D-consistent generative world creation. Recently acquired Scenics. Actively seeking robotics company customers.

"WorldLab is a two year old startup... We are building the next frontier of AI, which is what we call spatial intelligence... a means to an end towards spatial intelligence is building large world models. And that's what WorldLab is mostly focused on." — Fei-Fei Li 00:01:49

Scenics

Robotics simulation infrastructure startup co-founded by Yunzhu Li, Changxi Zheng, and Sunny Hu. Built the real-to-sim-to-real pipeline for robot training and evaluation. Acquired by WorldLabs after organically becoming a Marble customer. Embodiment-agnostic and model-agnostic platform.

"At Scenics, we are developing what we call a real-to-sim-to-real pipeline... we can replace all the data, all the evaluation we need in the real environments by using the data that can generate at a scalable way in our digital world." — Yunzhu Li 00:00:06

Waymo

Self-driving car company cited as the premier real-world proof point that simulation-heavy development works at scale.

"Waymo has officially said they use billions of hours of simulation. And actually, Waymo is more simulation heavy than just real-world data heavy. So these are real examples." — Fei-Fei Li 00:21:04

Weta (Weta Digital)

World-class VFX studio where Scenics co-founder Changxi Zheng previously worked, cited as evidence of deep simulation and rendering expertise on the Scenics team.

"The other two are Changxi Zheng, another Columbia professor who has been a world-class technologist in simulation... Changxi has his background in also VFX. He worked at Weta." — Fei-Fei Li 00:07:37

Amazon

Cited as a prior employer of Scenics co-founder Sunny Hu, who worked across computer vision tech stacks following an acquisition.

"There's Sunny Hu, who is a phenomenal engineering leader who was also in a startup that was acquired by Amazon many years ago. So, he worked in many different tech stacks in the computer vision field in Amazon." — Fei-Fei Li 00:08:02

Tencent

Cited as a prior employer of Changxi Zheng, further substantiating his breadth of industry experience in simulation and technology.

"He worked at Weta. He worked at Tencent. He's been an entrepreneur." — Fei-Fei Li 00:07:37

4. People Identified

Founder and CEO of WorldLabs. Former Stanford AI Lab director, co-inventor of ImageNet. Architect of the spatial intelligence thesis and the WorldLabs/Scenics integration strategy.

"WorldLab is a two year old startup... We are building the next frontier of AI, which is what we call Spatial Intelligence." — Fei-Fei Li 00:01:49

Yunzhu Li

Co-founder of Scenics, assistant professor at Columbia University. PhD from MIT, postdoc with Fei-Fei Li at Stanford. Full-stack robotics researcher from hardware to modeling, now joining WorldLabs.

"Throughout my career, my goal has been very simple. Trying to help the robots better perceive and interact with the physical world. I'm a very practical person. I want my robot to work in the real physical environments." — Yunzhu Li 00:04:24

Changxi Zheng

Co-founder of Scenics, Columbia University professor. World-class simulation and rendering technologist with industry experience at Weta Digital and Tencent. Described as one of the most senior researchers in simulation.

"Changxi Zheng, another Columbia professor who has been a world-class technologist in simulation... Changxi has his background in also VFX. He worked at Weta. He worked at Tencent. He's been an entrepreneur." — Fei-Fei Li 00:07:37

Sunny Hu

Co-founder of Scenics. Seasoned engineering leader with startup experience (acquired by Amazon) and deep computer vision expertise across multiple tech stacks at Amazon.

"There's Sunny Hu, who is a phenomenal engineering leader who was also in a startup that was acquired by Amazon many years ago. So, he worked in many different tech stacks in the computer vision field in Amazon." — Fei-Fei Li 00:08:02

Sergey Levine

Prominent robotics researcher (UC Berkeley). Mentioned specifically for the view that simulation always eventually diverges from physical reality and real-world data is therefore irreplaceable — a position the speakers address and complicate.

"I've heard other researchers say like Sergey Levine say simulation will always eventually deviate from the physical world and real world data collection is absolutely critical." — Martin Casado 00:18:38

Martin Casado

General Partner at a16z. Host of this episode and long-time board member / close collaborator with Fei-Fei Li. Frames the investor perspective throughout, including pointed questions on unit economics of humanoids and the viability of simulation-only approaches.

"Language models transformed how AI understands words. The next frontier is teaching AI to understand and act within the physical world." — Martin Casado 00:01:07

5. Operating Insights

Acquire Your Own Organic Customers First

The Scenics acquisition was discovered not through proactive M&A sourcing but because Scenics signed up as a paying customer for Marble. Fei-Fei Li did not even initially know it was Yunzhu Li's company. The lesson for founders: your customer list is also your acquisition pipeline. Monitor it for strategic signal.

"They came into WorldLab as a customer. When we released the first version of our generative model called Marble last winter around November, December, Scenics just signed up... I didn't even know what it was. And then I realized this is Yunzhu's company... we realized there's so much synergy." — Fei-Fei Li 00:05:34

Don't Rush Post-Acquisition Integration — Protect the Acquired Team's Contained Stack

Rather than immediately blending teams and codebases, WorldLabs is deliberately keeping Scenics' stack and customer relationships relatively contained in the near term. This preserves optionality and avoids destroying value through premature integration.

"We're not rushing to integrate everything from codebase to Teams because I think Scenics does have a very well-thought and I wouldn't call it stand-alone completely, but fairly contained tech stack as well as their customers, as well as the kind of products they're building. We're going to take time." — Fei-Fei Li 00:37:19

Define the Two-Year Lighthouse Customer Vision Before Scaling

Fei-Fei Li's explicit two-year success definition is deliberately narrow: validated customers in a small number of important vertical use cases that become lighthouse examples. This is a disciplined constraint against premature scaling — prove it works deeply before going broad.

"I'm very happy that Scenics team and WorldLabs team will have validated customers in a small number of important vertical use cases where our system, our infrastructure has proven to be truly beneficial to their automation needs. And these customers became our lighthouse examples to scale our business." — Fei-Fei Li 00:39:20

Academic Founders Who Prioritize Industry Design Partners Early Are a Strong Signal

Fei-Fei Li specifically called out as refreshing that Yunzhu and Changxi, despite coming from academia, immediately sought real industry design partners — labs, warehouses, electronics assembly — rather than building in a vacuum.

"Their first instinct is work with design partners and customers in real industry, whether it's labs at industry labs or warehouses or electronics assembly. That is such a refreshing, actually, a refreshing way of approaching robotics." — Fei-Fei Li 00:15:55

6. Overlooked Insights

The Robotics Evaluation Market Is a Standalone, Underserved Business Today

Yunzhu Li casually describes a discrete and currently unmet market: robotics companies need simulation environments specifically to run evals — to distinguish between a checkpoint that performs at 90% versus 92% — because doing this in the physical world takes prohibitively long. This is not just a feature of a broader robotics platform; it is a standalone pain point with paying customers right now, independent of training. Nobody in the conversation stops to name this as a distinct product category, but it is.

"The key criteria people use in industry is how long does it take? How long in work clock time does it take for you to distinguish between a checkpoint that is 90% from a checkpoint that is 92 points? And if you only do that in the real environment, it's just take so long for you to do the distinguishments." — Yunzhu Li 00:24:11

"Some customers, they need only the real to sim part. They want to digitalize the task they care about and be able to do the evaluations of their robotic systems." — Yunzhu Li 00:40:13

The Single Most Demanded Robotic Task Is Cleaning — An Enormous Unserved Market Hiding in Survey Data

In a single throwaway sentence, Yunzhu Li reveals that when they surveyed the general public about what they want robots to do for them, one-third of all responses were about cleaning. This is direct demand-signal research pointing to a massive, specific, near-term market for robotic solutions — one that is far less glamorous than humanoids or autonomous vehicles but may be far more commercially accessible and urgent.

"We actually send out surveys asking the general public what they want the robots to do for them. Among the southern tasks we collected, one-third of the tasks are about cleaning. People just don't like to do those, like, dull and dirty tasks. And those are the scenarios that we really want to make sure we have robotic solutions to deal with." — Yunzhu Li 00:14:58