Fei-Fei Li: Spatial Intelligence is the Next Frontier in AI

Fei-Fei Li argues spatial intelligence — understanding and generating 3D worlds — is the hardest unsolved problem in AI and a prerequisite for AGI.

• Li claims AGI cannot be complete without spatial intelligence, calling it more fundamental than language.
• Human language evolution took under 500,000 years; vision and 3D spatial reasoning took 540 million — arguing vision is combinatorially harder.
• World Labs is founded with Justin Johnson (neural style transfer), Ben Mildenhall (NeRF author), and Christoph Lassner (precursor to Gaussian Splatting).
• Language is 1D and purely generative; the 3D world is 4D with time, physically constrained, and requires balancing generation with reconstruction — making it mathematically ill-posed.
• The core spatial data problem: language data is abundant on the internet; 3D spatial data is not, requiring hybrid real-world and synthetic approaches.
• Li told Andrej Karpathy in ~2015 to reverse image captioning and generate images from text — he said ‘I’m out of here’; that is now standard generative AI.
• For PhD students, Li recommends problems not on a collision course with industry compute advantages: interdisciplinary AI, theory, causality, and small-data regimes.
• Li ran a dry-cleaning shop at 19 to fund her Princeton physics degree — frames it as her first founder-CEO exit after seven years.

2025-07-01 · Watch on YouTube