LLenny's Podcast
← All episodes
Dr. Fei-Fei Li16 November 2025

The Godmother of AI on jobs, robots, and why world models are next

5Frameworks
16Insights

Frameworks in this episode

Mindset4 steps

Mission-Over-Minutiae Opportunity Choice

Choose your next big move by passion, mission, and team — not by optimizing every minor variable.

Read →
Strategy4 steps

Northstar Problem Commitment

Pick one foundational building-block problem and commit years to it, not a scatter of trendy ones.

Read →
Innovation4 steps

Objective-Data Alignment Test

Before betting on an AI capability, check that your training data is the same shape as your desired output.

Read →
Innovation4 steps

Physical-System Reality Check

For anything embodied, budget for three things — a brain, a body, and real application scenarios.

Read →
Strategy4 steps

The Golden Recipe of Modern AI

Every AI breakthrough sits on three legs: big data, the right neural architecture, and enough compute.

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 3

Hot Take06:30

Why She's an AI Optimist but Not a Utopian

Fei-Fei Li frames her optimism as humanist, not utopian: AI's impact on jobs and people is real, but the outcome is up to us. She sees technology as a net positive across the long arc of civilization, while insisting every technology is a double-edged sword that a careless society can misuse.

  • She rejects both doom and utopia; the outcome depends on human choices
  • Humans are a fundamentally innovative species that keeps improving its tools
  • Technology is a net positive but always a double-edged sword
  • Individuals, communities and society can 'screw this up' if they act irresponsibly

“So, Lenny, let me be very clear. I'm not a utopian.”

Fei-Fei Li · 06:30

“So I do believe technology is a net positive for humanity.”

Fei-Fei Li · 06:30
#ai-optimism#ethics#humanism#future-of-ai
Hot Take24:00

AGI Is a Marketing Term, Not a Scientific One

Li pushes back on the AGI hype, noting no one has really defined the term and that its definitions range from machine superpowers to economically viable agents. As a scientist she sees no meaningful distinction between AI and AGI, and suspects even Alan Turing would just shrug at the question.

  • No agreed definition of AGI exists; definitions vary wildly
  • Her north star has always been simply 'can machines think and do what humans do'
  • She won't go down the rabbit hole of AI vs AGI
  • Turing asked the same question in the 1940s

“I don't know if anyone has ever defined AGI.”

Fei-Fei Li · 24:00

“I feel AGI is more a marketing term than a scientific term.”

Fei-Fei Li · 25:30
#agi#hype#definitions#hot-take
Hot Take47:30

The Brain Runs on 20 Watts — Dimmer Than a Lightbulb

Asked whether the work gives her awe for the brain, Li points out the human brain operates on about 20 watts — dimmer than any lightbulb in the room — yet does extraordinary things. The deeper she goes into AI, the more she respects humans.

  • The human brain runs on roughly 20 watts
  • That's dimmer than a lightbulb, yet capable of enormous cognition
  • It took a half billion years of evolution to produce that power
  • Working in AI increases her respect for human intelligence

“We we operate on about 20 watts. That's dimmer than any light bulb in in the room.”

Fei-Fei Li · 47:30

“the more I work in AI, the more I respect humans.”

Fei-Fei Li · 48:00
#neuroscience#brain#awe#human-intelligence

Explainer· 4

Explainer27:00

What Today's AI Still Can't Do (That a Toddler Can)

Li argues we're far from done innovating by naming concrete gaps: current models can't reliably count the chairs in a video of office rooms — something a toddler manages. Nor can AI derive Newton-style physical laws from data, or match the emotional intelligence of a mentor talking a student through their motivation.

  • AI struggles to count chairs in a video — a task a toddler can do
  • It cannot derive Newton's laws of motion even given modern celestial data
  • It lacks the emotional and cognitive intelligence of a real mentoring conversation
  • No serious scientific discipline has ever declared itself done innovating

“ask the the model to count the number of chairs. And this is something a toddler could do”

Fei-Fei Li · 27:30

“that level of creativity extrapolation abstraction we have no way of enabling AI to do that today.”

Fei-Fei Li · 28:00
#ai-limitations#spatial-intelligence#reasoning#emotional-intelligence
Explainer34:30

What a World Model Actually Is

Li explains that unlike a chatbot, a world model lets anyone generate a world in their mind's eye from an image or sentence — and then interact with and reason inside it. It's a foundation for browsing, picking up objects, changing things, or, for a robot, planning a path to tidy a kitchen.

  • You prompt with an image or sentence and get an explorable world
  • You can navigate, pick up objects, and change things inside it
  • A robot consuming the output could plan a path and act in the world
  • It's a foundation to reason, interact, and create worlds

“this model can allow anyone to create any worlds in their mind's eye by prompting whether it's an image or a sentence and also be…”

Fei-Fei Li · 35:00

“world model is a a foundation that that you can use to reason, to interact, and to create worlds.”

Fei-Fei Li · 35:30
#world-models#spatial-intelligence#robotics#world-labs
Explainer42:30

Why the 'Bitter Lesson' Alone Won't Crack Robots

Li explains that language models enjoy a perfect setup — words in, words out — while robots must output actions in 3D worlds that training data barely captures. Worse, robots are physical systems closer to self-driving cars than to LLMs, and even self-driving took 20 years despite being 'simpler' robots that never have to touch anything.

  • Language models have perfect alignment: word training data, word outputs
  • Robots need actions in 3D worlds, but training data lacks those actions
  • Teleoperation and synthetic data are needed to fill the gap
  • Robots are physical systems, closer to self-driving cars than to LLMs
  • Self-driving cars took ~20 years and are 'simpler' robots that avoid touching anything

“You hope to get actions out of robots. But your training data lacks actions in 3D worlds.”

Fei-Fei Li · 43:30

“self-driving cars are much simpler robots. They're just metal boxes running on 2D surfaces. And the goal is not to touch anything.”

Fei-Fei Li · 46:00
#bitter-lesson#robotics#self-driving#training-data
Explainer57:30

How World Models Differ From Video AI (Plato's Cave)

Li distinguishes World Labs' work from video generators like Veo using Plato's allegory of the cave: vision is the task of making sense of a 3D or 4D world from flat 2D projections. Spatial intelligence is deeper than creating flat video — it's the ability to create, reason within, and interact with genuinely spatial worlds.

  • The world is not something you passively watch pass by like video
  • Plato's cave: vision means reconstructing a 3D/4D world from 2D projections
  • Spatial intelligence is deeper than generating flat 2D video
  • World Labs also shipped real-time video generation on a single H100 GPU
  • Marble gives creators worlds with real 3D structure to work with

“the world is not passively watching videos passing by, right?”

Fei-Fei Li · 57:30

“spatial intelligence to me is deeper than owning creating that flat 2D world.”

Fei-Fei Li · 59:00
#world-models#video-generation#spatial-intelligence#plato

Story· 4

Story17:30

How ImageNet and the 'Golden Recipe' Sparked Modern AI

Li recounts building ImageNet from 2006-07 by curating 15 million labeled images across a 22,000-concept taxonomy, then open-sourcing it via an annual challenge. The 2012 breakthrough came when Jeff Hinton's Toronto team combined ImageNet's big data, a neural network, and two Nvidia GPUs into the 'golden recipe' still underlying ChatGPT today.

  • ImageNet curated 15 million internet images across a 22,000-concept taxonomy borrowed from WordNet
  • She conjectured big data was the critically overlooked ingredient for AI
  • 2012: Hinton's Toronto team used ImageNet data plus two Nvidia GPUs to crack object recognition
  • The trio of big data, neural networks and GPUs is still the core of ChatGPT-era AI

“We curated very carefully 15 million images on the internet. Created a taxonomy of 22,000 concepts”

Fei-Fei Li · 18:00

“that combination of the trio technology uh big data neuronet network and GPU was kind of the golden recipe for modern AI”

Fei-Fei Li · 19:00
#imagenet#deep-learning-history#big-data#gpus
Story21:30

When Calling Yourself an 'AI Company' Was a Death Knell

Li recalls that as recently as 2015-16, some tech companies avoided the word 'AI' because they weren't sure it was a dirty word. She encouraged everyone to embrace it, and by roughly 2017 companies began branding themselves as AI companies — a stark contrast to today when you can't not call yourself one.

  • In 2015-16 some tech companies avoided the term 'AI' entirely
  • Li pushed people to use it, proud of it as one of humanity's most audacious questions
  • By ~2017 Silicon Valley marketing began calling companies 'AI companies'
  • Less than a decade later, not calling yourself an AI company is unthinkable

“some tech companies avoids using the word AI I because they were not sure if AI was a dirty word.”

Fei-Fei Li · 21:30

“2017ish was the beginning of companies calling themselves AI companies”

Fei-Fei Li · 22:00
#ai-history#branding#silicon-valley
Story51:00

The 'Dots' Loading Effect Users Love Was Intentional

Lenny praises the dotted point-cloud visualization you see before a Marble world renders. Li reveals it's an intentional visualization the engineers designed to guide people into the world — not part of the model itself — and she was delighted to learn how many users find the experience magical.

  • The dot/point-cloud effect before rendering is an intentional design feature
  • It is not part of the model — the model just generates the world
  • Engineers built it to guide users into the world and converged on the dots
  • Many users, not just Lenny, found the experience delightful

“the dots that lead you into the world was a an intentional feature uh visualization. It is not part of the model.”

Fei-Fei Li · 51:30
#marble#product-design#ux#world-labs
Story53:00

Marble in the Wild: 40x Faster VFX, Robot Sims, and Therapy

Li shares early Marble use cases: a Sony virtual-production team cut their VFX shoot time by 40x, researchers are generating synthetic environments to train robots, and — unexpectedly — a psychology team wants immersive scenes for psychiatric research. Lenny connects the last one to exposure therapy for phobias like heights.

  • A Sony virtual-production collaboration cut shoot time by 40x
  • Robotics researchers use it to generate diverse synthetic training environments
  • A psychology team reached out to create immersive scenes for patient research
  • Potential exposure-therapy applications for phobias (heights, snakes, spiders)

“they were saying this has cut our uh production time by uh 40x.”

Fei-Fei Li · 53:30

“a psychologist team called us to use marble to do psychology research.”

Fei-Fei Li · 55:30
#marble#use-cases#vfx#robotics#exposure-therapy

Tool· 1

Tool48:00

Marble: The First 'Prompt-to-Worlds' Product

Li introduces Marble, World Labs' first product, built on a frontier model that took over a year to create and outputs genuinely 3D worlds. Users prompt with a sentence, image, or multiple images and get navigable worlds — even walkable in VR goggles — which she calls 'prompt to worlds.'

  • Marble is built on World Labs' frontier spatial-intelligence model
  • It took over a year to build a model that outputs genuinely 3D worlds
  • Prompt with a sentence, image, or multiple images to create a world
  • Worlds are navigable and even walkable in VR goggles
  • Available at marble.worldlabs.ai

“We've spent a year and plus building the world's first uh generative model that can output genuinely 3D worlds.”

Fei-Fei Li · 49:00

“it's the world's first uh product that allows people to just uh prompt we call it prompt to worlds.”

Fei-Fei Li · 50:30
#marble#world-labs#product-launch#3d-worlds

Takeaway· 4

Takeaway07:30

There's Nothing Artificial About AI

Li reflects on the line she delivered to Congress and repeats to every graduating student: the field is called artificial intelligence, but there is nothing artificial about it. It is inspired by, created by, and most importantly impacts people.

  • AI is inspired by people, created by people, and impacts people
  • She reminds every graduating student of this framing
  • The framing reframes AI as a human endeavor, not a machine one

“your field is called artificial intelligence but there's nothing artificial about it.”

Fei-Fei Li · 08:00
#ai-philosophy#human-centered-ai#quotable
Takeaway1:05:30

Intellectual Fearlessness: The Throughline of Her Career

Asked how she kept landing at the centers of AI breakthroughs, Li credits intellectual fearlessness — the same quality she now hires for. She left a near-tenure position at Princeton to restart the clock at Stanford, joined Google to work with the best, and founded World Labs, refusing to overthink everything that could go wrong.

  • She hires young people for intellectual fearlessness and courage
  • Making a difference means diving into something new that people haven't done
  • She gave up a near-certain tenure track at Princeton to move to Stanford
  • She doesn't dwell on failure cases; she focuses on people and mission
  • 'I don't overthink all possible things that can go wrong because that's too many'

“I'm I'm an intellectually very fearless person and I have to say when I hire young people I look for that”

Fei-Fei Li · 1:06:00

“I don't overink of all possible things that can go wrong because that's too many.”

Fei-Fei Li · 1:08:00
#career#risk-taking#hiring#mission
Takeaway1:08:30

Advice to Young Talent: Optimize Passion, Not Every Variable

Li worries that young AI talent over-analyzes every dimension of a job offer. Slipping into mentoring mode, she urges candidates to focus on what actually matters — passion, mission alignment, faith in the team, and the impact and people they'll work with — rather than every minute variable.

  • Young candidates often weigh every single aspect of a job decision
  • She finds herself in mentoring mode, not just recruiting mode
  • The most important questions: where's your passion, do you align with the mission, do you trust the team
  • Focus on the impact you can make and the people you'll work with

“where's your passion? Do you align with the mission? Do you believe and have faith in this team?”

Fei-Fei Li · 1:09:30
#career-advice#hiring#mission#young-talent
Takeaway1:14:30

Everybody Has a Role in AI

Closing out, Li answers the question she hears everywhere — does a musician, nurse, farmer, or accountant have a role in AI? — with a resounding yes. She argues no technology should take away human dignity, and that dignity and agency must sit at the heart of how AI is developed, deployed, and governed.

  • Everyone — artists, nurses, farmers, teachers — has a role in AI
  • No technology should take away human dignity or agency
  • Dignity and agency must anchor AI's development, deployment, and governance
  • She wants healthcare workers augmented by AI as society ages
  • Silicon Valley too often talks past ordinary people with abstractions

“Everybody has a role in AI.”

Fei-Fei Li · 1:15:30

“no technology should take away human dignity and the human dignity and agency should be at the heart of the development, the deployment as well…”

Fei-Fei Li · 1:16:00
#future-of-work#human-dignity#ai-ethics#human-centered-ai