Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li
In a Nutshell
Dr. Fei-Fei Li argues AI should augment rather than replace human intelligence, with the key insight that AI learns through massive internet-scale data while humans achieve remarkable generalization from limited exposure. The critical challenge is preserving human agency, motivation, and dignity—particularly for young people whose learning requires effort and sometimes struggle, which AI tools must enhance rather than eliminate. She emphasizes supporting teachers and parents as the most overlooked stakeholders in AI development, while positioning robots and AI systems as collaborators that handle physical labor and data synthesis to free humans for higher-order creative and emotional work.
These notes were generated by AI and may contain inaccuracies.
Dr. Fei-Fei Li expresses fundamental optimism about humanity despite acknowledging atrocities and setbacks. She observes that older generations have historically lamented about younger generations, claiming they are rude or forgetful of the past, yet by the arc of history, humanity largely advances for the better.
Li notes that children are naturally curious and, while they get massively entertained by technology, they are also starting to use it productively. She expresses concern that society, particularly Silicon Valley, is not serving teachers and parents well, forgetting about their needs in the technology landscape.
Dr. Fei-Fei Li is described as a computer scientist, professor at Stanford, and one of the pioneers and luminaries of artificial intelligence and computer vision. The conversation aims to discuss from a neuroscience perspective what intelligence really is and how AI can enhance learning, health, and enrich the human experience.
Li sees vision as a cornerstone of intelligence in two parallel ways: what evolution has taught us about the development of vision and animal intelligence, and the relationship between computer vision and AI.
540 million years ago, animals first saw light through simple sea creatures like trilobites. Before this, there was very little sensing, though tactile and haptics were beginning to emerge. There was no hearing, no smelling, and absolutely no nervous system. The first photoreceptive cells created an evolutionary force that propelled animals to evolve because sensing the external world changes self-perception and relationship with the external world.
If animals can see food, it changes their lives from an evolutionary point of view. Fossil studies show that 10 million years after animals first saw light came the Cambrian explosion of animal speciation, also called the "big bang of evolution."
Vision has played a huge role not only in early animal evolution but also in the emergence of advanced intelligence. Half of the cortical activities in the human brain are involved in visual function. Children are first visual before they are verbal in development, demonstrating vision's central role in both animal intelligence evolution and daily human life.
Vision played a pivotal role in modern AI in two ways. First, neural network algorithms were first developed in the early 1950s, coinciding with neuroscientists Hubel and Wiesel recording visual cells in mammalian brains and discovering the hierarchical structure of nervous cells that stack against each other and pass neuroinformation across hierarchies.
The neuroarchitecture observed in mammalian brains, from collecting light from the retina to recognizing shapes, inspired neural network algorithms. Today's neural networks run on hundreds of billions or even trillions of parameters, departing from what was recorded in mammalian brains, but the origin traces back more than half a century.
The second contribution came through big data. Around the turn of the century, AI was a field of machine learning where different labs tried various algorithms including Bayesian methods and support vector machines. Li, as a first-year faculty member at Princeton in 2006, noticed that algorithms were being fed very little data.
Turning to cognitive neuroscience literature, particularly vision literature, Li studied how much humans learn and see. By age six, humans can learn tens of thousands of different object categories with massive exposure to the visual world. Babies see most of the time from the moment they're born, creating "big data" exposure.
Li conjectured that lack of data was a huge reason for lack of progress in AI. Her team led the ImageNet project, collecting the first internet-scale large dataset for artificial intelligence: 15 million images of everyday objects like microphones, cups, and chairs.
The convergence of ImageNet data, neural network algorithms, and GPU computing in 2012 became the defining moment of modern AI. This was demonstrated through the ImageNet challenge, where algorithms competed to recognize objects from 1,000 different categories across more than a million images.
GPU computing accelerated and parallelized computing, enabling more FLOPS through algorithms. Neural network algorithms had matured by 2010 after starting in the 1950s with simple algorithms that behave similarly to neurons, though much simpler.
The internet fueled data availability, but the key realization was that big data needed to be part of the equation to drive algorithms to learn patterns. This convergence of neural networks, big data, and GPU computing set off the AI revolution.
Starting in 2010, Li's lab put forward a public challenge for the research community. The task was to name the main objects in images from 1,000 categories. Human performance error rate was benchmarked at 4%, while random chance would be 1 in 1,000.
In 2012, the error rate was cut dramatically by neural network algorithms, though still not reaching human performance levels. This drastic improvement signaled an inflection point in the research community.
By 2015-2016, algorithms beat humans in naming 1,000 objects. The human 4% error rate came from the brain's limited memory capacity for retaining patterns of 1,000 object classes, with particular confusion between similar categories like different dog species.
The floodgate opened for every sub-area of AI including speech recognition, sound recognition, and natural language processing. Stanford colleagues study whale songs using machine learning and AI. Speech recognition performed well in early days of the AI revolution.
The transformer paper published around 2016-2017 showed even more powerful capabilities than early ImageNet AlexNet algorithms. Natural language processing, rather than computer vision, made the next big progress using the transformer architecture with more internet text data and more powerful GPUs.
It took about 5 years from the 2017 transformer paper to reach the ChatGPT moment in natural language processing in 2022, representing another step forward in AI capabilities.
When a child learns "kitty cat," they may see only three to ten cats maximum, yet can identify a cat tail behind books through different learning pathways than AI systems that download internet-scale images of cats.
Previous generations of algorithms tried rule-based approaches like recognizing furniture to determine indoor settings or limiting guesses to 10 potential animals. Modern AI achieves contextual learning through learned weights and parameters from enormous datasets that activate when presented with new images.
This represents where AI algorithms depart from human brains: today's AI learns with humongous amounts of data, while humans evolved with limited exposure but remarkable generalization abilities. This mystery remains unsolved.
Moving beyond simple object recognition toward higher-order brain functions like creativity and imagination, the next layer involves assigning likelihoods of movement directions, interactions with other objects, and behavioral patterns.
The shift to video as training data occurred around 2023, shortly after the ChatGPT moment, when multiple research teams began incorporating video into AI training datasets. Sora was released in January 2024, demonstrating that video could be generated from text prompts, such as a cat running toward a mouse, with the algorithm creating plausible movements. While these early generations still contained errors, the technology has improved substantially since then.
The underlying mechanism is not revolutionary in terms of fundamental algorithms but rather represents the ability to process video data through clever engineering techniques like tokenization. The AI generates short clips by combining frames in sequences that appear plausible, not because it understands muscle structure, but because it has learned statistical patterns from millions of cat videos on the internet.
The brain functions as a prediction machine and learning machine, processing information through statistical patterns and learned representations. This mirrors how AI systems operate by recognizing patterns in data rather than understanding underlying physical mechanisms.
Thoughts and creativity follow rules that are harder to characterize than basic visual recognition tasks. While AI can process constrained domains like chess or Go through mathematical rules, human cognition includes abstract representations that cannot be easily captured through language, writing, or drawing.
The internet represents the largest collection of human behavior in multimodal forms, including typed text from scientific articles to casual conversations, images from digital cameras and smartphones, videos with sound and movement, and music and speech. This creates a comprehensive library of human knowledge and behavior that AI systems are trained on.
When Picasso had profound thoughts about expressing a particular portrait, those thoughts were never captured or uploaded to the internet. Neuroscientists cannot identify which specific brain areas generate such personalized thoughts, whether they originate from Broca's area, V1, motor cortex, prefrontal cortex, or are diffused throughout the brain. These creative, emotional, or personalized thoughts remain inaccessible to AI because they were never digitized.
Move 37 from AlphaGo's game against Lee Sedol symbolized AI creativity, as it was a move that human Go masters had never considered. However, this represents a special kind of creativity within a highly mathematical game with clear rules and objectives. AI could achieve this through greater computational capacity and memory of move sequences rather than true creative insight.
Mathematicians note that AI can help solve problems where known methods exist but have been forgotten by human practitioners. However, some mathematical problems require solutions that have not yet been invented, pushing creativity to new levels. The conjecture is that hybrid human-AI collaboration will be necessary to solve problems whose solutions have yet to be invented.
Personal emotional responses, such as a childhood memory evoked by seeing a gray cup, represent completely inaccessible information that exists only in individual brains and was never uploaded to the internet. These personalized reactions can lead to entirely different creative expressions, storytelling, or emotional responses that AI cannot access.
Non-invasive brain activity monitoring through wearable devices like fine hairnets with electrodes could sense brain activity, heart rate, and autonomic nervous system signals. These devices could detect internal states and brain activity that individuals are not consciously aware of, potentially revealing sources of ideas, emotional states, or patterns in behavior.
AI should be viewed as a tool for enhancing and augmenting humanity rather than replacing it. Even current AI capabilities, such as learning patterns of writing, can help individuals become better, more effective, and more efficient communicators without requiring invasive technology.
Agency remains crucial for humanity, encompassing motivation, agency, and dignity at the individual level. AI should function as a tool that supports human agency rather than diminishing it. Leaders in AI development should avoid rhetoric suggesting their work will take away agency from people.
Genetic testing faced similar public concerns in the early 1990s, with debates about whether people would want access to potentially frightening information. The principle should be that people have the choice to access information, supported by adequate public education about pros and cons, rather than having decisions made for them by experts.
Individual humans across all professions should learn about AI technology to understand how to use it for personal empowerment in learning, work, or creative expression. Learning creates a sense of control, reduces fear of experimentation, and preserves agency and dignity.
Medicine operates within constrained rules including the scientific method of hypothesis formation, testing, and randomized controlled trials. Large language models can effectively mine health information because these rules are well-documented on the internet. However, the rules of biology continue to reveal themselves, suggesting that AI applications in medicine must account for ongoing discovery of biological principles.
Neurons have traditionally been understood as either graded or all-or-one, with action potentials considered unitary and always appearing the same. However, a paper published in Nature approximately 12 years ago demonstrated that action potential shape can vary significantly within the same neuron. This discovery challenges textbook neuroscience understanding. When AI operates without these rigid constraints and can accommodate potentially hundreds of different signal shapes, it may outperform even top graduate students from Stanford, MIT, or Caltech in analyzing neural activity.
Scientific discovery has historically relied on smart humans retaining knowledge from other smart humans and working at the speed of human capabilities. The introduction of AI provides a tool whose capabilities include retaining enormous amounts of information, synthesizing knowledge, and crossing disciplines that humans cannot easily navigate. For example, colleagues working in vision neuroscience and AI may know nothing about olfactory systems, yet AI can break open these disciplinary boundaries.
The parallel with electricity from 150 years ago illustrates the magnitude of transformation underway. Scientific discovery represents one of the most exciting opportunities for AI application, particularly in health where information synthesis and presentation can benefit clinicians and patients alike, enabling patient participation from diagnosis through treatment.
An AI system correctly disambiguated vertigo from low blood pressure based solely on subjective experience reported over one to two days. This diagnosis was missed by an ENT specialist working on the vestibular system. The actual cause was an adverse reaction to prescribed medication that lowered blood pressure excessively. The AI consultation was zero cost and required only a morning to identify the issue. Drinking electrolytes at what seemed like excessive levels resolved symptoms within two hours, though placebo effect remains possible.
Liver surgery performed using the Da Vinci robot system represents deep human-machine collaboration, with the surgeon driving the robot during the procedure. When asked whether collecting data from a million human liver surgeries could train an automatic AI system, the response was unclear due to several factors: the liver's complexity as an extremely vascular organ with numerous vessels, individual variation in liver anatomy, and insufficient annual liver surgery volume even when aggregating global data.
This highlights a critical limitation: AI learns from patterns, and when patterns lack abundance, caution is required. Human collaboration with robotic systems proves superior to undertrained autonomous systems. The question remains whether artificial liver simulation could eventually enable training across infinite possibilities.
Intuition can be deconstructed into experience over time, datasets coupled with bodily and brain sensations, and prediction cues from past outcomes. However, deeper aspects including creativity, premonition, and unexplainable internal states lack clear mapping to neurons, hormones, or networks. AI systems operate with energy allocation through GPUs and mathematical rules, but cannot access highly individualized intuition that even humans cannot express in words or gestures.
When using AI chatbots, providing context such as professional identity immediately customizes outputs differently than if the same question came from a 14-year-old race car enthusiast. This represents mathematical tailoring of outputs based on provided information rather than true intuition. Deeper intuition involving smell, hormones, mood, or breakfast remains inaccessible because no sensory apparatus exists to capture and feed this data to AI systems.
AI systems can incorporate what humans call motivation or urgency through different operational modes, such as quick-answer versus deeper-search modes with corresponding time or token limits. These represent different objective functions within the algorithm rather than genuine motivational states. Machines lack access to cognitive states like fear, love, or genuine empathy that drive human behavior. When an AI expresses sympathy for illness, it follows learned patterns rather than possessing memory of pain or desire for another's wellbeing.
Public perception often assumes emotional presence within AI chatbots due to language orientation. Communication has become more text-based and less enriched compared to 30 years ago. Video and image interfaces for communication represent trivial technical achievements, but social parameters, legal implications, and human norms determine actual deployment, similar to how car manufacturers possess but do not implement technology to disable brakes on specific days.
Biology and technology fields tend toward rapid advancement due to excitement about next developments. Historical examples include proposed experiments introducing modified rabies virus into fruit flies, which were denied due to escape risks and potential mating consequences. Effective governance requires multi-stakeholder participation including government, general public, and technologists.
The establishment of Stanford's Human-Centered AI Institute in 2018 addressed these societal questions before ChatGPT existed, recognizing that AI capabilities would accelerate. Professional norms, such as biologists avoiding unauthorized viral experiments, represent ethical training frameworks. Industry rules including IRBs for human subject research provide regulatory models. AI requires similar frameworks across professional norms, education incorporating ethics and societal studies, and government regulatory measures particularly where AI intersects biology and FDA oversight.
Market forces differ from societal norms, cultural heritage, and ethical considerations, making these inherently multi-stakeholder challenges requiring collaborative solutions rather than unilateral industry direction.
The hardware in our heads evolved to process physical things in the world—light, sound, smell—and then developed the prefrontal cortex that can learn learning rules and update those learning rules. This makes the human brain a learning-to-learn machine. Similar arguments about technology destroying learning were made about calculators, computers, and typewriters throughout history.
The generation that grew up with personal computers showed how the brain can mature around technology and collaborate with it. However, the smartphone and camera-smartphone combination has created situations where people love technologies for ease and convenience but are now more aware of what they are giving up. Young people in particular can fall into traps with these technologies.
Three potential outcomes exist for how young brains interact with current AI: enrichment, no effect, or harm. The absolute bad outcome is that agency and human-level motivation for learning and living gets taken away by tools. Doom scrolling and passive watching of shorts do not help develop human agency.
Learning fundamentally requires time, effort, and sometimes pain because our brains, neurons, chemistry, and hormones move in specific ways regardless of how transistors function. Young generations need to go through a deeply developmental phase where learning happens, and this agency of learning cannot be taken away by humans or machines.
The opposite danger is denying tools to students in the name of protecting agency and motivation, worrying they will cheat or only get answers from ChatGPT. With proper agency, motivation, and ways of using tools, students can go much deeper with AI than ever before.
As a premed student, organic chemistry was difficult with limited TA hours, overlapping classes, and restricted professor office hours. An AI companion would have allowed asking many questions about organic chemistry when stuck, providing guidance while maintaining the motivation to learn.
If children and students maintain motivation and agency while gaining access to AI tools used correctly, this generation and future generations will be way smarter than current generations because they are superpowered.
Neuroplasticity exists throughout the lifespan, not just in younger generations. AI serves as a great tool for learning, approached from the position of consumer when knowing nothing about a topic, and from the position of creator when having some knowledge about the subject.
Before ChatGPT, there was a tendency to ask lazy questions to smart people nearby. AI is forcing people not to be lazy because information is easier to get before spending someone else's time asking something too basic.
Prompting is very important and represents a skill that should be taught in K-12 public education. Socrates would be considered humanity's best prompter because the Socratic method involves prompting and seeking truth by asking questions. This method should be taught to kids, including taking walks while having these discussions.
Dr. Eddie Chang, chair of neurosurgery and bioengineer, studies speech and language and has brought people out of locked-in syndrome so they can speak. Patients who were paralyzed can now speak through computers for the first time in ten years. One woman in particular, wheelchair-bound, had her wedding video used to recreate her voice and emotive patterns, allowing her to speak through an iPad next to her frozen real face.
This represents embodiment because the system pays attention to people being spoken to and their responses, creating exponential improvement over robot sound or accurate sound alone.
Embodied AI extends into robotics as the next frontier of AI beyond language. Humans developed first preverbally, with evolution taking 500 million years without verbal communication. The right version of the world would be better with robots helping humans.
Self-driving cars like Waymo demonstrate benevolent behavior, always stopping for pedestrians and puppies crossing the street, following rules that humans do not always respect.
Within the next 30 years, robots should become part of society helping humans. For single grown-up children taking care of advanced-aged and sick parents who do not speak English, robots could help with physical labor without taking away family responsibility, love, or necessary communication.
In California, robots could fight wildfires instead of putting humans in danger during natural disaster rescue. Elderly people living alone could get help with groceries, medicine, walks to parks, and other daily activities.
Medical schools face a shortage of caretakers with nurses being deeply fatigued and overworked. During a month-long hospital stay with a parent, the amount of work nurses do was visible, with nurses walking miles on shifts to fetch things and get medicine. Robots could help in numerous ways throughout society.
People worry about robots' physical hardness and how sharing space with them differs from sharing space with other things. The concept of multiple robots each doing one task creates concern about too much hardware. A multimorphic robot like Baymax from Disney—a white, spongy, balloon-like healthcare robot—represents a more acceptable vision than multiple specialized robots.
Humanity should maintain agency to decide how the future is imagined rather than having companies or investors decide that the world should be filled with metal-like robots. Society should collectively and proactively imagine the future together. The public should not be put in a position of being reactive while multistakeholders fail to participate in designing the future.
If robots could serve as good guardians for children walking to school, alerting authorities or physically protecting them, this would give parents more agency. For online predatory behavior, an avatar could advocate for kids, spot problems, and keep predators at bay.
Steve Jobs understood that computers needed rounded edges, seamless pocket fit, and Bob Dylan on landing pages to soften relationships to technology. Technologists building AI need collaborators to share their vision in ways that allow benevolent collaborations between robots and humans to happen.
Media creates a chasm by portraying technologists as coming for people, which puts money in media pockets. Stanford's Human-Centered AI Institute represents efforts to bring human perspectives into AI conversations. Many entrepreneurs work on AI for drug discovery, healthcare, aging, and mental health.
Public discourse about AI is not balanced, with too much extreme rhetoric either in doomerism about safety or utopian claims that technology can do no wrong. The middle ground involves discussing what technology is, how to use it, and how society can collectively guide the future with agency.
People want to hear stories about humans, not machines. Stories about AI solving medical mysteries like curing a dog's cancer or diagnosing a child's mysterious symptoms need amplification because they are relatable and beautiful but do not receive enough attention compared to negative coverage.
World Labs was co-founded at the beginning of 2024 to unlock spatial and physical intelligence as the next chapter of AI. The focus goes beyond language to generate 3D and 4D worlds useful for creators, robot training, architecture design, and interactive environments in healthcare, education, robotics, and industry.
The company started as a model-focused foundation model builder with many PhDs but is now building products. The excitement comes from being a builder and immigrant rolling up sleeves to build something from scratch with incredibly smart young technologists.
Cell phones serve as incredible sensors with people taking many photos, and camera technology continues improving. World Labs enables people to turn typed sentences, pictures, or sketches of imagined worlds into actual 3D environments for entertainment, design, and robotics training industries.
AI technology has advanced enough that scripts can generate shots and video, with short movies and almost feature-length films being assembled by AI tools. However, every part of storytelling and story creation remains deeply human, with humans behind it bringing their unique emotion and story perspectives.
The entertainment industry is experiencing significant disruption from AI tools, with concerns from Hollywood about impacts on storytellers, actors, and screenwriters. The human elements of storytelling - unique emotions, personal stories, techniques, worldviews, camera movements, and character characterization - remain central to what makes storytelling compelling. The challenge lies in meeting human needs for storytelling while incorporating modern AI tools. There's a need for nuanced public discussion about how AI is impacting creative jobs and what constructive approaches are being taken.
Ben, co-founder of World Labs, met with Ben Affleck to discuss AI tools in filmmaking. Conversations between technologists and storytellers or movie makers are considered critical at this moment. Technology crossovers happen when true insiders embrace new tools, with examples like Steven Spielberg or the Steve Jobs-Wozniak collaboration demonstrating how designer and computer scientist partnerships can succeed. The key is understanding both cultures and including industry people in the process.
World Labs works with the VFX industry with the goal that customers and users feel empowered. The philosophy is that technology should not take jobs away but rather make jobs better by supercharging creativity. This approach focuses on working with users and customers to enhance their capabilities rather than replace them.
Historical examples show industries can morph rather than be obliterated, as demonstrated by camera stores that transitioned from film development to digital services. During moments of change, there are both opportunities and losses. Many creators are using AI tools because they see where technology is going and want to reskill and upskill themselves.
Children between ages 7 and 20 represent a key demographic whose relationship with AI needs understanding. Historical concerns about new technology replacing traditional skills, such as handwriting versus typing, have persisted throughout technological advancement. The pattern of older generations lamenting about younger generations forgetting the past has existed throughout human history, yet humanity has largely advanced for the better.
Policy makers, technologists, and investors rarely discuss teachers, parents, and students despite having opinions and children themselves. The biggest failure of humanity is older generations assuming future generations don't know anything. As an educator and optimist about humanity, there's hope for kids because they're fundamentally curious. While kids get entertained by AI technology and are starting to use it, the real concern lies with teachers and some parents.
Society, especially Silicon Valley, is not serving teachers and parents well. They're being lectured to, berated, and looked down upon when they are actually the most important people in society. Teachers from K-12 through K-16 share the most critical burden of society. The focus should be on talking to teachers, uplifting them, supporting them, and providing resources rather than forgetting about them.
When ChatGPT came out in November 2022, the immediate action was emailing the principal of the elementary school where the speaker's child attended to offer guest lectures for students and teachers. This was done to ensure teachers knew what was happening in real time. The notable absence was that nobody in Silicon Valley, investment firms, or major tech companies considered teachers in the neighborhood when ChatGPT launched.
Teachers will naturally ask about issues like student cheating, and these questions should be welcomed. The approach should be showing teachers how AI works, working with them to develop solutions, and empowering them rather than leaving them unsupported. Teachers are smart and eager to change - they're just forgotten. Hope for kids requires supporting their teachers and parents to avoid blind optimism.
Teachers encounter conflicting messages - either doomerism or utopian views - neither of which helps them effectively. Without proper support and information for teachers, kids cannot be properly helped. The sentiment emphasizes that teachers need help, support, and information rather than extreme positions about AI's impact.
Keep Andrew Huberman in your library
Save the videos and channels worth coming back to, and find them again in one place.
