Andrew HubermanUsing AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li
Andrew Huberman
Optimism about the next generation 0:00
Fei-Fei Li opens by pushing back on the common complaint that younger generations are careless or ignorant just because they use new technology differently. She points out that across the arc of history, humanity tends to advance, even though setbacks and atrocities are real. She says kids today are curious and are starting to genuinely use AI, not just be entertained by it. Her real worry is aimed elsewhere: she thinks teachers and parents are being underserved by a society, especially in Silicon Valley, that has not done enough to prepare them for this technology.
Vision as the root of intelligence 4:32
Andrew Huberman introduces Fei-Fei Li as a computer scientist, AI pioneer, and director of the Stanford Institute for Human-Centered Artificial Intelligence, known for insisting that ethics stay central to AI development. Li explains why vision matters so much, tracing it back 540 million years to when simple sea animals called trilobites first evolved light-sensing cells. Before that, there was almost no sensing at all, not even a nervous system. Once animals could see, they could find food and avoid becoming food, and this pressure sped up evolution dramatically, leading to what fossil records call the Cambrian explosion of animal diversity about ten million years later. She notes that half of the human cortex is devoted to visual function, and that children develop visual understanding before verbal language.
How neural networks borrowed from brains 7:32
Li connects the biology of vision to the history of AI algorithms. In the early 1950s, around the same time computer scientists began experimenting with early neural network algorithms, neuroscientists Hubel and Wiesel were recording visual cells in mammalian brains and discovering a layered structure that passes information from raw light in the retina up to recognizing shapes. That layered brain architecture partly inspired the design of artificial neural networks, even though today's networks, running on hundreds of billions or even trillions of parameters, have grown far more complex than anything found in biology.
Building ImageNet to feed the algorithms 10:00
As a first-year faculty member at Princeton in 2006, Li noticed that AI algorithms were being fed very little data compared to how much visual information humans absorb. Research showed that by age six, children can already recognize tens of thousands of object categories, having been exposed to visual input since birth. This led her to conclude that the missing ingredient in AI was not just better algorithms but far more data. She led the creation of ImageNet, a dataset of 15 million internet images meant to teach machines to recognize everyday objects like microphones, cups, and chairs. By 2012, the combination of this large dataset, maturing neural network algorithms, and GPU computing power converged into what is now seen as the defining moment of modern AI.
Machines catching up to human vision 15:30
Li describes the ImageNet Challenge, a public competition her lab launched around 2010 asking algorithms to correctly name the main object in each of over a million test images across a thousand categories. A Stanford graduate student later benchmarked human error on this task at about 4 percent, confusion often arising from closely related categories like different dog breeds rather than from time pressure. In 2012, error rates dropped sharply due to neural networks, signaling a major inflection point, though it still took until around 2016 for algorithms to actually surpass human accuracy on the task.
The same recipe spreads beyond vision 22:02
Li explains that once the combination of big data, neural networks, and GPU power proved successful in vision, the same approach quickly spread to other domains, including speech recognition, sound recognition such as researchers studying whale songs, and natural language processing. When the transformer algorithm was published around 2016 to 2017, it proved even more powerful than the earlier ImageNet-era AlexNet approach, and companies like OpenAI and Google built on it using ever more internet text data and stronger GPUs. It still took about five years, from 2017 to 2022, to reach the ChatGPT moment.
Why AI still learns differently than children 27:31
Huberman and Li discuss how a child can learn to recognize a cat from just a handful of real examples, while modern AI systems rely on learning from vast amounts of internet data. Li explains that today's systems recognize a cat's tail peeking from behind a bookshelf because they have absorbed so many patterns from huge datasets, not because they reason the way a child does. She admits this remains a real mystery in the field, since a child who has only seen a few cats can still generalize successfully, through some other kind of learning pathway that has not been fully explained.
From still images to moving video 29:30
Li describes the next major leap: teaching AI to generate motion, such as animating a cat running toward a mouse without explicit instructions about limb movement. This became possible once video was added to training data. Shortly after the ChatGPT moment, in January 2024, OpenAI released Sora, which could generate short video clips from text descriptions, showing plausible but still imperfect animal movement. Li stresses that this advance was still fundamentally about data rather than any deep understanding of anatomy. The algorithm does not know the actual muscle structure of a cat's legs, but it has seen so many cat videos that it has learned what plausible movement looks like, much as humans recognize natural cat motion without knowing the underlying muscle mechanics themselves.
The Limits of AI's Training Data 33:32
Andrew Huberman raises a question he has held onto for a year and a half: can AI capture aspects of human cognition that were never uploaded to the internet, things like a fleeting thought, an abstract painting, or a piece of music that evokes an unnamed feeling. Fei-Fei Li agrees this is an important limit. She explains that the internet is essentially the largest collection of human behavior in multimodal form, built from decades of typing, photos, videos, and music. AI is trained on that vast library, which is why it is so powerful at recognizing and synthesizing patterns in language and imagery. But something like the exact thought behind a Picasso portrait was never captured in words, images, or sound, so it was never uploaded, and AI has no way to access it. Even neuroscientists cannot say which brain region produced that thought, which underscores how personal and uncaptured such moments remain.
Move 37 and the Question of Machine Creativity 41:30
Fei-Fei brings up Move 37, the famous move AlphaGo made against Lee Sedol that no human master had ever considered. She calls it a real form of creativity, but a special kind tied to the fact that Go is a highly mathematical game with clear rules, where greater computing power let the machine explore possibilities humans never had reason to try. She recounts a conversation with a leading mathematician who was optimistic that AI could help solve many unsolved problems in math simply because it can retain and apply known methods better than any human, even a Fields medalist. But he also doubted AI could solve every problem, since some solutions have not been invented yet and will require a new level of creativity. Fei-Fei's own guess is that the answer will be hybrid, humans and AI working together to reach solutions neither could invent alone.
Personal Memory as Uncapturable Information 45:01
Fei-Fei uses the example of a gray cup that might evoke a private childhood memory shared only with a best friend. That association lives entirely in her brain, was never typed, photographed, or spoken aloud, and so was never uploaded anywhere AI could learn from it. She argues this is why AI, however powerful, cannot access certain personal reactions, whatever you choose to call them, creativity, expression, or storytelling.
Brain Sensing and Collaborating With Yourself 45:31
Huberman imagines a near future, maybe five to ten years out, where a lightweight, non-invasive device could sense brain activity, heart rate, and alertness, comparing that internal state to what a person says and does. He frames this as something that could stay private and benevolent, letting someone collaborate with unconscious parts of their own mind to produce something meaningful, like an image drawn from an experience they could not consciously access. Both agree people are already curious about their own patterns, yet currently rely on crude self-experiments, like tracking how many cups of coffee produce a good day, to understand themselves.
AI as Augmentation, Not Replacement 48:01
Fei-Fei stresses that AI should be understood as enhancing humanity rather than replacing it, even without exotic tools like brain-sensing headsets. Something as ordinary as AI learning a person's writing patterns can make them a better communicator. She names agency, motivation, and dignity as central to what makes people human, and insists AI should support agency rather than take it away. She criticizes voices in the AI field who talk down to the public, deciding for people what is good for them instead of offering honest education and letting people choose. Huberman connects this to the early days of genetic testing, when people feared what a blood test might reveal, noting that some people still prefer not to know, and that this choice should be respected. Fei-Fei adds that learning about AI, not necessarily coding, helps anyone, whether artist, teacher, or doctor, feel more in control and less afraid to try.
Constrained Rules Versus the Messiness of Biology 58:01
Huberman contrasts fields with fixed rules, like Go or the visual patterns that identify a cat, with medicine and biology, where the rules keep shifting. He gives the example of action potentials, long taught as strictly uniform, until a Nature paper about twelve years ago showed their shape can vary substantially, a finding that unsettled basic neuroscience but never fully overturned the textbook framework. He suggests AI could handle this kind of rule instability better than even top graduate students, since it can absorb far more variation without being confused by it. Fei-Fei responds that scientific discovery is one of the most exciting uses of AI precisely because it can synthesize knowledge across disciplines that no single human brain could master, comparing the moment to the transformative arrival of electricity. She sees huge potential for AI to help synthesize medical information for both clinicians and patients, and to involve patients more directly in diagnosis and treatment. Huberman shares a personal case where AI correctly identified that his vertigo-like symptoms were caused by a medication lowering his blood pressure too far, a diagnosis that had eluded a specialist in the vestibular system.
A Symptom Checked Against Big Data 1:04:00
A listener describes calling a clinic about vertigo and low blood pressure, and how an AI-guided check let them confirm at zero cost that drinking electrolytes would resolve the issue within two hours. Fei-Fei Li notes this kind of reassurance is powerful, though it does not replace a doctor. She points out that patterns like vertigo tied to low blood pressure have likely been reported so often that AI has learned them well, making it useful for people without immediate access to medical care.
The Liver Surgery Robot 1:04:59
Andrew recounts his father's liver surgery at Stanford, performed with a surgeon guiding a Da Vinci robot, a case of deep human machine collaboration. He asked the surgeon whether enough data exists across all human surgeons to train a fully automatic AI for this kind of surgery, and the answer was unclear because the liver is highly vascular and different in every patient. Fei-Fei explains that even aggregating global liver surgery data might not be enough to train such an algorithm, which shows that AI depends on abundant patterns, and where patterns are scarce, human and machine collaboration works better than an undertrained robot working alone. She raises the open scientific question of whether an artificial liver simulation could someday generate unlimited training data. The surgery itself succeeded, with ten times less blood loss than typical thanks to the robot's laparoscopic precision.
Two Kinds of Intuition 1:07:33
Andrew asks whether AI could ever develop something like human intuition, the sense that comes from experience, bodily sensation, and prediction. Fei-Fei separates a shallow, accessible form of intuition, which is really just context a user provides in a prompt, such as mentioning you are a neuroscientist versus a teenager, from a much deeper form rooted in things like smell, hormones, or mood that a person cannot even express in words. She argues this deeper intuition is inaccessible to AI not because of mysticism but because there is no sensory apparatus yet to capture that data, whether for another human or a machine, unless tools like brainwave or skin-conductance sensors eventually change that.
Motivation as Objective Functions 1:14:32
Andrew asks whether machines can have something like motivation or urgency. Fei-Fei explains that such states can be built into mathematics through what machine learning calls objective functions, giving the example of a chatbot's "think deeper" versus "quick answer" modes, which differ only in token limits or which part of the model activates, not in any felt urgency. She stresses that today's machines do not truly feel emotions like empathy or love; when a chatbot says "I'm sorry you're sick," it is pattern-matched language, unlike a friend's genuine empathy rooted in memory of having felt pain. She insists on keeping this distinction clear for the sake of honest public communication.
Why Technology Holds Back 1:21:01
Discussion turns to how synthetic video or image communication is already technically trivial but held back by social, legal, and ethical considerations, much like car manufacturers could disable brakes on a schedule but don't. Andrew shares an anecdote about a researcher wanting to put a modified rabies virus into fruit flies, an idea rightly denied despite its scientific appeal, illustrating how professional norms restrain technologists eager to push boundaries. Fei-Fei explains this is why she left Google to help found Stanford's Human-Centered AI Institute in 2018, arguing that AI's societal implications require professional norms, ethics education, industry rules like IRBs, and government regulation working together, not decisions made by a few industry figures alone.
Agency, Learning, and Young Brains 1:27:30
Fei-Fei argues the human brain evolved a prefrontal cortex built for learning how to learn, meaning children can adapt to AI much as earlier generations adapted to calculators and computers. The real danger, she says, is AI stripping away human agency and motivation through things like doom-scrolling and passive content consumption, since learning inherently requires time, effort, and sometimes struggle that no technology can bypass. An equally bad outcome would be denying students AI tools out of fear of cheating, when a tool that offers patient, always-available guidance, like help with organic chemistry, could let a motivated student learn far more deeply. She closes by praising prompting as a real skill worth teaching in schools, comparing good prompting to the Socratic method of seeking truth through questioning.
Giving Voice Back Through Neural Interfaces 1:35:30
Researchers have found ways to translate neural activity into control of the voice box and throat muscles, allowing people with locked-in syndrome to speak again. One woman, paralyzed and wheelchair bound, now speaks through an iPad placed next to her still face, using a system trained on old wedding videos so it could learn her real voice and emotional patterns. This lets her interact with the world at a much deeper level than a flat robotic voice or a simple text-to-speech device would allow, and the system keeps improving through machine learning as it also picks up on the reactions of the people she talks to.
Beyond Language Into Embodied Robots 1:38:00
Fei-Fei Li argues the next frontier of AI goes beyond language, since humans and other species developed intelligence long before verbal communication existed. She points to self-driving Waymo cars as an early example, noting they reliably stop for her and her puppy in ways some human drivers do not. Looking further out, perhaps thirty years, she hopes robots can take real physical burdens off people, especially those caring for aging or sick parents, helping with tasks like getting groceries or medicine for elderly people living alone, and even assisting overworked hospital nurses who currently walk miles per shift fetching supplies. She also raises wildfire response as an area where robots could take on dangerous work instead of putting human rescuers at risk.
Designing Robots People Can Trust 1:42:30
Discussion turns to why robots and computers can feel unsettling, largely because of their physical hardness and the strange way people have to share space with them. Rather than a house full of single-purpose machines, a multimorphic, soft-bodied design like Disney's Baymax is offered as a more comfortable model, one robot handling many tasks with a spongy, approachable form. Fei-Fei stresses that humanity, not just companies or investors, should have agency in deciding what this future looks like, and gives examples like a guardian robot that could keep a child safe walking to school or an AI companion that could watch for online predators.
Needing a Steve Jobs for AI 1:46:00
Drawing on the memory of Steve Jobs walking around Palo Alto in his early days, the conversation suggests AI needs someone who understands human nature well enough to soften technology's hard edges, the way Jobs gave computers rounded edges and made them fit into daily life. Fei-Fei responds that many such people already exist, entrepreneurs and researchers building AI for drug discovery, healthcare, aging, and mental health, and points to Stanford's Human-Centered AI Institute as one such effort. She feels public discourse is unbalanced, swinging between extreme doom and extreme utopianism, and argues the more powerful stories are quiet ones, like AI helping cure a person's illness, which deserve far more attention than headline-grabbing tech rhetoric.
World Labs and Spatial Intelligence 1:50:02
Fei-Fei describes her startup, World Labs, founded in early 2024, as her life's work, built on the idea that intelligence extends beyond language into spatial and physical understanding. The company builds foundation models aimed at generating 3D and 4D worlds useful for creators, robot training, and architectural design, and is now moving from research toward products. She explains that beyond capturing real-world images the way cell phones and cameras already do, World Labs lets people turn a sentence, picture, or sketch into an imagined, interactive world, a capability she sees as valuable across entertainment, design, and robotics.
AI, Filmmaking, and Creative Jobs 1:54:00
On whether AI can already turn a script into a finished film, Fei-Fei says the technology for generating video shots has advanced enough that short and even near feature-length films have been assembled with AI tools from companies in the US and Asia. But she insists storytelling remains deeply human, carrying each creator's emotion, technique, and way of seeing the world, and acknowledges real fear in Hollywood over jobs for screenwriters and actors. She mentions meeting with Ben Affleck through her co-founder Ben, noting how important it is for technologists and filmmakers to talk directly, and compares the shift to how camera and film-developing stores eventually gave way to digital tools without erasing the industry entirely, as people reskill and adapt.
Listening to Kids, Teachers, and Parents 1:59:31
Asked how children feel about AI, Fei-Fei says kids between seven and twenty are naturally curious and already experimenting with the tools, and she remains an optimist about humanity's long arc of progress despite its setbacks. Her deeper worry is for teachers and parents, who she feels are overlooked by policymakers, technologists, and investors even though they carry a critical burden. She recalls that when ChatGPT launched in November 2022, her first act was emailing her child's elementary school principal to offer a guest lecture for students and teachers, because no one in Silicon Valley was reaching out to neighborhood schools. She closes by urging support and honest information for teachers, since neither doomsday warnings nor blind utopian promises actually help them guide kids through the change.
Podcast Closing Remarks 2:06:01
This closing segment contains standard podcast outro material rather than new content from the discussion. Andrew Huberman thanks listeners, promotes his new book "Protocols: An Operating Manual for the Human Body," and points to his social media presence under the handle Huberman Lab.
Newsletter And Sign Off 2:07:30
He describes the free monthly Neural Network newsletter, available at hubermanlab.com, which includes podcast summaries and protocol PDFs on topics like sleep, dopamine, and fitness, before thanking listeners and closing the episode with Dr. Fei-Fei Li.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
