The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li

Summary of The Godmother of AI on jobs, robots & why world models are next | Dr. Fei-Fei Li

by Lenny Rachitsky

1h 19m•November 16, 2025
Listen to Episode

Overview of The Godmother of AI on jobs, robots & why world models are next (Dr. Fei‑Fei Li — Lenny Rachitsky podcast)

This episode features Dr. Fei‑Fei Li — creator of ImageNet, former Google Cloud Chief AI Scientist, former director of Stanford SAIL, co‑founder of the Stanford Human‑Centered AI Institute (HAI) and co‑founder of World Labs. The conversation traces the history of modern AI (how ImageNet reignited the field), explains the importance of world models and spatial intelligence, introduces World Labs’ product Marble (a prompt‑to‑3D‑world tool), and covers implications for jobs, robotics, policy and responsible AI.

Key takeaways

  • ImageNet + neural nets + GPUs were the pivotal combo that reignited modern AI; AlexNet (2012) used that recipe (initially with just two GPUs).
  • The three core ingredients powering large AI advances remain: lots of data, powerful neural architectures, and compute (GPUs).
  • World models (spatial/3D models) are a complementary and necessary step beyond language models for embodied AI, robotics, creative tools and scientific reasoning.
  • Marble (from World Labs) is an early consumer product that generates navigable 3D worlds from prompts and images — useful for VFX, games, robotics simulation, therapy research and more.
  • Fei‑Fei emphasizes human‑centered AI: everyone has a role in AI, it should augment human dignity and agency, and technologists must act responsibly.
  • Robotics faces unique data and physical challenges unlike language models — scaling alone may not be sufficient (the "bitter lesson" is necessary but not a complete solution).

Brief history & why ImageNet mattered

  • AI has a long lineage (Turing, Dartmouth 1956, decades of work). Modern momentum came from the machine‑learning shift (statistical learning vs. pure rule systems).
  • Fei‑Fei Li began focusing on visual intelligence and object recognition as a “North Star” problem: objects are fundamental to human interaction with the world, but very hard to learn without massive data.
  • ImageNet (started ~2006–2007) curated ~15 million labeled images across a large taxonomy (WordNet concepts) and was open‑sourced. It provided the "big data" missing piece.
  • The 2012 ImageNet challenge win (AlexNet, Hinton et al.) paired large labeled data + convolutional neural networks + GPUs, catalyzing deep learning progress.
  • The same trio of ingredients later scaled to language (foundation models / GPT) and now spatial models.

Notable quote: “There’s nothing artificial about AI — it’s inspired by people, created by people, and it impacts people.”

What are world models (spatial intelligence) and why they matter

  • Simple definition: models that internally represent and generate richly structured, navigable, interactive worlds (3D / 4D + dynamics), not just sequences of text or pixels.
  • Capabilities expected from world models:
    • Create immersive, explorable 3D environments from prompts (images/text).
    • Enable agents/robots to plan, reason and act in those worlds.
    • Support design, virtual production, game creation, simulation for training robots, and scientific/psychological experiments.
  • Rationale:
    • Language models are powerful but incomplete — many real‑world tasks require spatial, embodied reasoning (e.g., disaster response, manipulating objects, understanding 3D structure from 2D cues).
    • World models can augment human embodied intelligence and enable new classes of applications.

Analogy: Fei‑Fei referenced Rosalind Franklin’s 2D X‑ray photo enabling Watson/Crick to deduce a 3D double helix — humans use spatial reasoning to derive higher‑order structure; we want AI to do the same.

Marble — World Labs’ first product

  • Marble is a prompt‑to‑world product that generates genuinely 3D, navigable worlds from text and images (launched by World Labs).
  • Core features:
    • Create scenes with 3D structure and textures; navigate and walk through them (can export meshes or video).
    • Intentional UI affordances (e.g., “dot” visualization while worlds load) to help users explore model internals and make the experience delightful.
  • Early use cases:
    • Virtual production / VFX: reported production speedups (Fei‑Fei cited a ~40× acceleration for a demo project).
    • Games and VR content creation.
    • Robotic simulation: rapid creation of diverse synthetic environments for training.
    • Psychological and therapeutic research (e.g., exposure therapy setups).
  • Team & timeline:
    • World Labs formed by Fei‑Fei and co‑founders (Justin Johnson, Christoph Lassner, Ben Mildenhall).
    • Team ~30 (researchers, research engineers, designers, product), heavy GPU usage.
    • Marble was built over ~1+ year and is an early product in a nascent category.

Try it: marble.worldlabs.ai (World Labs site: worldlabs.ai)

Robots, the “bitter lesson,” and why robotics is harder

  • The “bitter lesson” (Richard Sutton): simple learning methods + scale (lots of data & compute) tend to win over hand‑engineered solutions.
  • Fei‑Fei’s view:
    • The bitter lesson is validated in vision and language, but robotics poses extra challenges:
      • Data scarcity and mismatch: robots need action‑annotated 3D data; web videos/videos alone are insufficient.
      • Physical constraints: robots operate in the physical world (3D, contact, safety), meaning hardware, environment maturity, and productization are big parts of the problem.
      • Simulation vs real world: synthetic data and teleoperation data can help, but transferring to real robots remains hard.
    • Robotics may benefit from world models (for simulation, planning and training), but progress will be iterative and multi‑disciplinary (hardware + software + use cases).
  • Realistic timeline: Fei‑Fei emphasizes that while progress is fast, current models are far from general human‑level creativity and broad AGI capabilities.

Societal impact, ethics & human‑centered AI

  • Fei‑Fei is a strong humanist: technology can be net positive, but is double‑edged and requires responsible deployment.
  • Everyone should care about and participate in AI decisions — not just technologists.
  • Key principles she highlights:
    • Preserve human dignity and agency.
    • Design systems that augment human roles (doctors, nurses, teachers, artists, farmers, etc.).
    • Invest in policy, interdisciplinary research, and community engagement (Stanford HAI’s mission).
  • Practical civic engagement: researchers and institutions should inform policy (congressional briefings, national research cloud efforts, regulatory input).

Notable quote: “Everybody has a role in AI.”

Advice for careers, founders and students

  • High agency & intellectual fearlessness: choose missions you deeply care about; don’t over‑optimize every small decision.
  • For young talent: prioritize passion, mission alignment and team over transient metrics (compensation, prestige). Practical experience and impact matter.
  • For founders: expect intense competition for talent and rapid change — integrate deep research with product focus early.
  • Fei‑Fei’s personal pattern: follow curiosity, accept risk (e.g., shifting roles between academia and industry to pursue mission and collaborations).

Practical applications & action items

  • If you create content, games or VFX: try Marble for rapid world prototyping and virtual production.
  • If you work on robotics: explore marble/world models for generating diverse synthetic scenes for training/simulation.
  • If you work in healthcare/psychology: consider immersive scene generation for experiments and therapeutic use cases (exposure therapy, controlled stimuli).
  • Policy and civic actors: engage with interdisciplinary AI institutes and academics to help shape governance and safety frameworks.

Notable quotes from the episode

  • “There’s nothing artificial about AI — it’s inspired by people, it’s created by people, and most importantly, it impacts people.”
  • “AI is up to us.”
  • “Everybody has a role in AI.”
  • “The more I work in AI, the more I respect humans.” (on the efficiency and complexity of the human brain: ~20 watts)

Resources & links

  • Marble (try it): https://marble.worldlabs.ai
  • World Labs (research, jobs, product): https://www.worldlabs.ai
  • Stanford Human‑Centered AI (HAI): search for Stanford HAI (co‑founded by Fei‑Fei Li)
  • ImageNet: the original open image dataset Fei‑Fei Li created (search “ImageNet dataset”)

Summary prepared to give you the historical context, technical framing of world models, concrete product insight (Marble), societal implications, and practical next steps so you can quickly understand what Fei‑Fei Li sees as the next chapter of AI.