FROM THE FRONTIER
The ‘godmother of AI’ is building AI models that can understand the physical world
Image Source: Andria Lo, IEEE Spectrum
AI models have conquered language, but they’re still blind to the physical world. Fei Fei Li, the prominent researcher behind the legendary ImageNet breakthrough and founder of World Labs, argues that’s about to change.
The missing piece. In a new blog posted earlier this week, Li argues that AI models like GPT-5 excel at abstract knowledge like language, but lack ‘spatial intelligence’ — they can’t understand how objects relate in physical space or predict basic physics. Li believes spatial intelligence is the scaffolding of human cognition itself. Without it, AI remains a brilliant wordsmith trapped in the dark, unable to truly interact with our physical world.
Li’s solution? “World models” that understand geometry, physics, and spatial relationships. Her startup World Labs already launched Marble, a world model that lets creators build fully explorable 3D worlds from text prompts. Think filmmakers conjuring entire sets without budgets, or scientists simulating molecular interactions in real-time.
The endgame is transformative: robots that navigate collapsing buildings, AI tutors using immersive 3D lessons, and drug discovery accelerated through spatial modeling. As Li puts it, spatial intelligence is “AI’s next frontier.”
via Superhuman