Episode Summary
Executive Summary: Fei-Fei Li discusses launching World Labs to build spatially intelligent AI that can understand, generate, and interact with plausible 3D worlds. She argues 3D world models are as fundamental as language, but harder and underdeveloped, with major implications for robotics, creativity, simulation, and healthcare. She also reflects on ImageNet, data-centric research, and human-centered AI.
Main Topics: Why start World Labs now (Priority: 5/5): Li explains that she wants to build at this moment because spatial intelligence and 3D world models are a critical frontier with broad real-world impact. She emphasizes building with a strong team of young technologists. Defining spatial intelligence (Priority: 5/5): She defines spatial intelligence as the ability to understand, reason about, interact with, and generate 3D worlds, framing it as a core component of intelligence for humans and animals and a prerequisite for complete AI. 3D world models as a foundation model problem (Priority: 5/5): Li says World Labs is tackling 3D generation foundation models and aims to create realistically accurate or at least plausible worlds that encode geometry and physics, unlocking applications in design, navigation, simulation, and AR/VR. Robotics, embodiment, and haptics (Priority: 4/5): The conversation explores robotics data hierarchies, simulation, synthetic data, teleoperation, and embodied data. Li argues robotics is multimodal and that haptics is an underappreciated but critical modality for manipulation. Creativity and commercial applications (Priority: 4/5): Li sees near-term value in using spatial AI to superpower designers, 3D artists, VFX artists, game developers, and marketers, especially for content creation in XR/AR/VR and collaborative creative workflows. ImageNet and the role of data in AI progress (Priority: 5/5): Li revisits the creation of ImageNet, explaining how limited data motivated a large-scale dataset that helped enable deep learning breakthroughs like AlexNet and validated the importance of scaling data. Human-centered AI and future priorities (Priority: 4/5): Li argues AI should augment people rather than replace human values, and highlights healthcare as a major area where AI can help with discovery, diagnosis, treatment, aging, and access.
Key Arguments: Spatial intelligence is as fundamental and difficult as language, because humans and animals live in and reason about 3D physical worlds. Current AI is incomplete without robust 3D world understanding and generation. World models should be geometrically and physically plausible, even when they are fantastical. Robotics will not converge on a single humanoid form; task demands will drive diverse, energy-efficient morphologies. Simulation and synthetic data are important for robotics, but haptics is especially underappreciated for manipulation tasks. Creative workflows are a near-term product opportunity because AI can collaborate with humans the way coding assistants do for software engineering. ImageNet demonstrated that large, well-curated datasets can unlock major advances even before scaling laws were fully understood. Fearlessness is essential for scientific and entrepreneurial progress because important breakthroughs require boldness in uncertain, underexplored areas. Human-centered AI should superpower people while preserving values like love, justice, prosperity, and relationships. Healthcare is one of the clearest areas where AI can create net good by expanding discovery and access to care.
Data Points: ImageNet labeled images: 15 million - Li describes the scale of ImageNet as a breakthrough dataset for object recognition and deep learning. ImageNet categories: thousands of categories - She explains ImageNet spanned many object classes, not just a small taxonomy. ImageNet early target size: 101 categories - Li recalls choosing 101 classes for her PhD dataset, slightly exceeding her advisor’s suggestion of 100. PhD era: 2003-ish - She situates the ImageNet work in the early internet era when large datasets were scarce. Timeline for storytelling problem: 100 years - Li says she once thought image storytelling might be a problem she could work on for the rest of her life. Family help with data cleaning: mother assisted - She recalls recruiting her mother to help clean images manually using a simple interface.
Pivotal Quotes: "Because, in my heart, I want to build." — Fei-Fei Li: Explaining why she started World Labs now. "Without spatial intelligence, AI would be incomplete." — Fei-Fei Li: Describing why 3D understanding and generation are foundational to intelligence. "Be fearless." — Fei-Fei Li: Her core advice for scientists, technologists, and entrepreneurs pursuing hard problems.
Implications: Listeners should expect 3D world models, robotics, and embodied AI to become major next-wave opportunities. The episode suggests the biggest wins will come from combining data, simulation, haptics, and human-centered design to create AI that augments people across creativity, industry, and healthcare.