Fireside Chat, Interview
How Fei-Fei Li Is Rebuilding AI for the Real World
- Fei-Fei Li is credited by Martin Casado as the "godmother of AI" for singularity introducing data as a central component of the field, contrasting with peers who focused primarily on neural network architectures.
- Casado notes that during a lunch, Li articulated the missing critical component in current AI as the lack of a "world model" to enable navigation of physical reality.
- World Labs was established with the conviction that 3D spatial intelligence is a fundamental requirement for intelligence, surpassing the limitations of language-based models (LLMs).
- Li argues that language is a "lossy" encoding of the physical world, whereas the evolutionary history of animals (spanning 500 million years for spatial navigation) relies on embodied intelligence rather than generative text.
- The company's "North Star" problem is to create systems capable of 3D reconstruction and generative capabilities simultaneously, moving beyond the 2D constraints of current vision systems.
- World Labs aims to enable a multiverse of digital environments for robotics, creativity (design, architecture, film), socialization, travel, and storytelling.
- Technically, the company focuses on converting 2D views into full 3D representations (including occluded areas like the "back of a table") to allow for measurement, manipulation, and interaction.
- Co-founder Ben Mildenhall pioneered Neural Radiant Fields (NeRF) for 3D reconstruction; co-founder Christoph Lassner contributed to Gaussian splat representation; and co-founder Justin Johnson was a pioneer in image generation using GANs and style transfer.
- Li shares a personal anecdote of losing stereo vision for a few months, which highlighted the necessity of 3D depth perception for safe driving and spatial navigation, as monocular vision lacks reliable distance metrics.
- The team integrates expertise across computer vision, diffusion models, computer graphics, optimization, and data science to productize the technology, aiming to solve the "unit economic positive" limitations of current AI.
- Li posits that while LLMs solved language problems rapidly due to the recent evolutionary emergence of language processing, solving spatial reasoning requires concentrated industry-grade effort in compute, data, and talent.
- Applications are described as "horizontal" and foundational, enabling robots to navigate unknown 3D spaces and allowing humans to manipulate digital objects with physical properties (e.g., stacking, measuring) in virtual environments.
- Casado emphasizes that World Labs represents a departure from previous trends where research was siloed; the company aims to concentrate the world's top talent on this specific singular problem.