Fei-Fei Li: How Vision Built Human and AI Intelligence

Andrew Huberman sits down with Dr. Fei-Fei Li, the Stanford computer scientist known as the godmother of AI and director of Stanford's Institute for Human-Centered Artificial Intelligence, for a wide-ranging conversation about where intelligence comes from and where it is headed. The discussion moves from the biology of vision to the mechanics of machine learning, and finally to the human values that should guide AI as it enters medicine, creativity, and daily life.

Vision as the root of intelligence

Li frames vision as the true cornerstone of intelligence, both evolutionary and technical. The Cambrian explosion, when animal life diversified rapidly hundreds of millions of years ago, is widely linked to the emergence of eyes and the ability to perceive the world. That same logic shaped Li's own career: she built ImageNet, a massive labeled image dataset, which helped trigger the 2012 deep-learning inflection point when neural networks trained on visual data began dramatically outperforming older approaches. That breakthrough set off the modern AI boom and eventually led to today's video-generation systems, such as Sora, which can synthesize moving, realistic scenes from a text prompt. For Li, vision was never a side project. It is the sensory foundation that both brains and machines use to build a model of the world.

How AI and human brains learn differently

One of the clearest threads in the conversation is the contrast between how AI systems and human brains acquire knowledge. Large AI models learn from enormous quantities of internet data, while a child can learn a new concept from just a handful of examples. Huberman and Li connect this to the prefrontal cortex, the brain region that essentially gives humans a learning-to-learn machine, capable of updating its own learning rules over time. That is why children who grow up alongside chatbots and AI tools are not necessarily worse off than earlier generations who grew up with calculators or the personal computer. Li argues the deeper risk is not the technology itself but how it is used: tools that respect a child's need to develop agency and motivation through effort, versus tools or habits, like passive scrolling, that quietly erode both.

What AI still cannot access

Huberman and Li spend real time on creativity and intuition, distinguishing shallow intuition, which is really just AI using context clues from a conversation, from deeper intuition rooted in personal, embodied experience. A chatbot can tailor its answer based on who you say you are, but it cannot access the private, often wordless internal states, hormonal shifts, memories, felt emotions, that never get written down or uploaded anywhere. Li is careful to demystify this rather than make it sound mystical: if information can be expressed through language, images, or physiological signals, it can eventually be learned by a machine. If it cannot be expressed at all, no human or machine can access it. That is also why, she argues, an AI saying "I'm sorry you're sick" is fundamentally different from a friend saying the same words: the machine is pattern-matching, not remembering what pain feels like.

AI and robotics in medicine

The conversation turns concrete when it comes to healthcare. Li describes an AI system that helped identify vertigo linked to low blood pressure, a pattern common enough in medical data for AI to recognize reliably, and useful for people without immediate access to a doctor. She also shares a personal story: her father underwent liver surgery at Stanford performed with robotic assistance through the Da Vinci system, guided by a human surgeon, and lost ten times less blood than in a typical procedure. But she is candid about the limits: liver anatomy varies enormously between patients, and even aggregating every liver surgery performed worldwide may not generate enough data to train a fully automated surgical AI. In domains where patterns are not abundant, she says, human and AI collaboration remains far safer than either operating alone.

Embodied AI and the road ahead

Looking further out, Li describes AI's next frontier as embodiment, moving beyond language into robots and physical assistance. She points to self-driving vehicles that already yield to pedestrians more reliably than some human drivers, and imagines healthcare-style robots, evoking Disney's Baymax, helping elderly people get groceries or medicine, and easing the load on overworked nurses and family caregivers. Rather than replacing human connection, she frames these tools as freeing people up for the parts of caregiving, storytelling, and teaching that remain deeply human.

Keeping humans at the center

Throughout the episode, Li returns to one word: agency. She warns against both extreme AI doomerism and uncritical utopianism, arguing that professional norms, education, and regulation all need to evolve together, the way society adapted to cars, airplanes, and biotechnology. Her strongest concern is for teachers, parents, and students, the groups she says are talked about constantly in AI debates but rarely talked to directly. Her own response, after ChatGPT launched, was to go teach her child's elementary school teachers what the technology actually does.

The episode closes on a note of grounded optimism: intelligence, biological or artificial, is built on perception, learning, and the choices humans make about how to use these new tools. Li's closing message is that preserving human dignity, motivation, and agency should stay the design goal, not an afterthought, as AI keeps advancing.