The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li
Dr. Fei-Fei Li explains why world models and spatial intelligence are the next AI frontier, and launches Marble, the first large world model product.
- ImageNet (2006–2012) combined 15M labeled images with neural nets and Nvidia GPUs, producing the trio that still underlies every major AI model including ChatGPT.
- As late as 2015–2016, major tech companies avoided the word ‘AI’ as a brand liability; widespread self-identification as ‘AI companies’ only began around 2017.
- Li considers AGI a marketing term, not a scientific one, and says today’s AI cannot derive Newtonian mechanics even given modern celestial-body data Newton never had.
- World Labs (18 months old, ~30 people) launched Marble, claimed to be the world’s first generative model outputting navigable, genuinely 3D worlds from text or image prompts.
- Virtual production teams using Marble reported 40x faster production time; Sony collaborated on World Labs’ launch video using the product.
- Robotics faces a structural data mismatch language models avoided: training data is passive web video, but robots need action data in 3D worlds — the bitter lesson is unproven for robotics.
- Self-driving cars, far simpler robots running on 2D surfaces, took 20 years from Stanford’s DARPA win (2005) to Waymo street deployment, signaling a long road for general robotics.
- HAI, co-founded at Stanford in 2018, now spans hundreds of faculty across all eight schools and helped pass a national AI research cloud bill during the first Trump administration.
2025-11-16 · Watch on YouTube
