Using AI to Increase Your Intelligence and Enrich Humanity

Artificial intelligence can do far more than retrieve information. Used well, it can help us learn difficult material, communicate more clearly, discover new treatments, and turn ideas into things we could not have built on our own.

But the important phrase is used well. A tool that strengthens our ability to think is very different from a tool that removes our motivation to think. A robot that helps a nurse care for patients is very different from a system deployed without enough training or human oversight.

My discussion with Dr. Fei-Fei Li kept returning to one central idea: AI should enhance human agency, not take it away. Agency means having the knowledge, motivation, and freedom to make meaningful choices. Understanding what AI can do, how it learns, and where it differs from the human brain is essential to retaining that agency.

Table of Contents

• Why Vision Is Central to Intelligence
• The Three Ingredients Behind the Modern AI Revolution
• How the Same Recipe Expanded Into Speech, Language, and Video
• How AI Recognizes Context, and Why Children Learn Differently
• What the Internet Captures, and What It Leaves Out
• Intuition, Motivation, and Empathy Are Not Just Fluent Answers
• Using AI to Strengthen Your Own Agency
• AI in Scientific Discovery and Healthcare
• Children Need Both AI Access and the Effort of Learning
• Embodied AI: Intelligence Beyond Words
• Spatial Intelligence and the Work of World Labs
• AI, Filmmaking, and the Future of Creative Work
• Why AI Governance Requires More Than Technology Companies
• Teachers and Parents Cannot Be an Afterthought
• A Human-Centered Future Is Something We Build

Why Vision Is Central to Intelligence

To understand modern AI, it helps to start with something much older than computers: the ability to see.

Vision transformed the relationship between animals and their environment. Detecting light, locating food, avoiding predators, and finding mates created powerful evolutionary advantages. The emergence and expansion of visual systems around the Cambrian period was part of a profound acceleration in animal complexity.

Vision remains central to human intelligence. A substantial portion of cortical processing participates in visual function, and children begin learning about the visual world before they develop language. They encounter objects, movements, and spatial relationships long before they can describe them.

That relationship between seeing and learning also helped shape artificial intelligence.

From visual neuroscience to neural networks

Research by David Hubel and Torsten Wiesel revealed how neurons in the mammalian visual system respond to particular features of the visual world. Across successive stages of processing, the nervous system transforms incoming signals into increasingly complex representations.

This hierarchical organization helped inspire artificial neural networks. In a simplified description, an artificial network consists of computational units that take inputs, apply mathematical functions, and pass outputs to other units. Stacking those units allows a system to learn increasingly complex patterns.

However, inspiration is not equivalence. Modern neural networks can contain hundreds of billions of parameters or more, and their architecture and learning processes differ substantially from those of biological brains. Calling something a neural network does not mean it operates like a complete nervous system.

The Three Ingredients Behind the Modern AI Revolution

For years, researchers explored different algorithms for getting machines to recognize objects, understand speech, and interpret language. Progress was real, but limited.

One important question was whether researchers were focusing too heavily on algorithms and too little on the amount of information those algorithms needed to learn.

When Fei-Fei was a young professor at Princeton, she and her students looked to the human visual system for clues. Children encounter enormous amounts of visual information. By early childhood, they can recognize a remarkable variety of objects. Yet many computer vision systems were being trained on comparatively small datasets.

That observation helped motivate ImageNet, a project that assembled roughly 15 million images to support machine learning research. The goal was to help computers recognize everyday objects, from cups and chairs to animals and vehicles.

The major breakthrough came from the convergence of three ingredients:

• Better algorithms: Decades of research had improved the capabilities of artificial neural networks.
• Large datasets: Internet-scale collections supplied far more examples from which systems could learn.
• More powerful computing: Graphics processing units, or GPUs, enabled the parallel computation needed to train larger networks.

No single ingredient explains the transformation. Their combination changed what was possible.

What the ImageNet challenge demonstrated

Beginning in 2010, the ImageNet challenge gave researchers a common object recognition benchmark. Systems had to identify objects from a set of 1,000 categories using a large collection of images.

In 2012, a neural network system produced a dramatic reduction in classification error. It did not immediately surpass human performance, but the size of the improvement made clear that the field had reached an inflection point. Over the following years, systems improved further and surpassed the human benchmark on that particular task.

The human error estimate discussed was roughly 4 percent. Some mistakes involved distinguishing closely related categories, such as similar dog breeds. Remembering and consistently applying 1,000 category labels is not trivial, even for a person.

The qualification matters: exceeding a human benchmark on a defined task is not the same as exceeding human intelligence in general.

How the Same Recipe Expanded Into Speech, Language, and Video

The advances did not remain confined to computer vision. Speech recognition, sound analysis, and natural language processing also benefited from better algorithms, more data, and faster computation. Researchers have even used machine learning to study whale vocalizations.

Another major step came with the transformer architecture, introduced in 2017. Combined with enormous amounts of text and increasingly capable computing hardware, transformers helped enable the language systems that eventually reached broad public use through tools such as ChatGPT.

Video added another layer. A still image can show a cat, but video provides examples of how cats move, which direction they tend to travel, and how their limbs change position over time.

As video became part of training, systems became increasingly capable of generating plausible movement. A request for a cat running toward a mouse could produce a short sequence that resembles that event.

That does not necessarily mean the system understands a cat's muscles or biomechanics. It has learned patterns in what movement looks like. In some respects, humans do something similar: most of us can recognize a plausible walking cat without knowing its anatomy. But plausible appearance should not be confused with a complete understanding of the physical processes involved.

How AI Recognizes Context, and Why Children Learn Differently

Imagine a cat hidden behind a bookshelf with only its tail visible. A child may identify the animal correctly despite seeing very little of it.

That seemingly simple act requires more than matching a complete picture. The child uses partial information, previous experience, and context. A furry tail inside a home is more likely to belong to a cat than to a fox.

Earlier computer vision approaches tried to build such reasoning through explicit rules. A system might first identify the room as an indoor environment, or narrow its choices to a small set of animals.

Today's large models can often make the same identification because their training has exposed them to so many related patterns. The new image activates learned parameters associated with cats, tails, bookshelves, and indoor scenes.

Yet there is a major difference between that process and childhood learning. A child does not need to download the internet to learn what a cat is. A relatively small number of encounters may be enough to support recognition across different angles, movements, and partial views.

We do not fully understand all the mechanisms that make human learning so efficient. The fact that machines and children can arrive at similar answers does not establish that they arrived there by the same route.

What the Internet Captures, and What It Leaves Out

To understand AI's strengths and limitations, consider what its training data represents.

The internet is an enormous collection of recorded human activity. It includes everyday conversations, scientific papers, photographs, videos, speeches, music, and many other forms of expression. These records provide a vast resource for learning patterns in language and behavior.

But a record of human expression is not the entirety of human experience.

You may look at an ordinary gray cup and feel nothing in particular. Someone else may associate that cup with a childhood friend, a family ritual, or a deeply personal moment. If that association has never been expressed or recorded, a model has no access to it.

The same issue applies to artistic creation. A finished painting can be photographed and analyzed. The particular internal experience that led an artist to make it may never have been captured at all.

This is not an argument that human thought is mystical or beyond scientific study. It is a practical distinction: a system cannot learn a particular piece of private information that it has never received.

AI creativity is real, but it requires careful definition

AI can combine information in surprising and useful ways. One famous example is AlphaGo, which produced an unexpected move during its match against Lee Sedol. Move 37 became a symbol of machine creativity because expert players had not anticipated that approach.

That achievement deserves recognition. It also occurred within a game with explicit rules and a clearly defined objective. A machine's ability to explore possibilities within that structure is different from the personal, emotional process behind a painting or a story.

Mathematics presents a related opportunity. A human mathematician cannot remember every method developed over centuries. AI may connect a difficult problem with an existing technique that a person has forgotten or never encountered.

Other problems may require genuinely new methods. Whether those advances come from humans, AI, or collaboration remains an open question. Fei-Fei's working expectation is that hybrid creativity, with humans and AI working together, will be especially important.

Intuition, Motivation, and Empathy Are Not Just Fluent Answers

Because chatbots use language, it is easy to attribute human internal states to them. We hear a responsive sentence and instinctively imagine someone behind it.

That instinct can obscure important differences.

Context is not the same as private intuition

If I tell a model that I am a neuroscientist, it can tailor its explanation differently than if I say I am a 14-year-old interested in race cars. That is context affecting the output.

Some things we call intuition may also reflect learned patterns: a scientist recognizes a cell type from subtle features, or an experienced clinician notices a familiar combination of signs. When those patterns can be captured in data, machines may be able to learn them.

But other intuitions involve information that is difficult even for the person experiencing it to identify. A mixture of bodily state, memory, mood, and sensory cues may shape a judgment without becoming expressible in words.

Until relevant information can be measured or communicated, a model cannot use it. Future sensors might make additional signals accessible, but that possibility should not be confused with an established ability to read a person's private experience.

Computational priorities are not felt motivation

A model may offer a quick-answer mode and a deeper-reasoning mode. It can operate under different time limits, token budgets, or computational objectives.

From the outside, that might resemble urgency or deliberation. Mathematically, however, it is a system operating under different constraints. It does not establish that the machine feels motivated, anxious, or eager.

Similarly, a chatbot can produce a sympathetic response when someone reports being ill. It has learned that such language is appropriate in that context. A friend's concern involves a different set of processes: attachment, memories of pain, and a genuine desire for that person's well-being.

Supportive language can be useful without being evidence of felt empathy. Public communication about AI needs to preserve that distinction.

Using AI to Strengthen Your Own Agency

The most useful framing is not that AI replaces you. It is that AI can help you become more effective at things you care about.

Even without speculative brain interfaces, a system can help improve your writing, organize information, or explain a concept in several different ways. It can support your expression without becoming the source of your values or goals.

One of the most important steps is simply learning enough about the technology to use it deliberately. That does not mean everyone needs to become a programmer. An artist, physician, teacher, or student needs an understanding appropriate to their work.

Learning reduces the distance between a person and an unfamiliar tool. It creates room for informed choices, including the choice not to use it.

That principle also applies to more personal technologies. Some people want as much information about their health as possible. Others do not. Education should explain the benefits and limitations rather than assume that technical experts should decide on everyone's behalf.

The same standard should guide AI: explain how it works, make its limitations understandable, and leave meaningful decisions with people.

AI in Scientific Discovery and Healthcare

Scientific discovery is one of the most exciting applications of AI because it addresses a basic limitation of human expertise: no scientist can retain and integrate everything.

Even researchers in neighboring fields may know very little about one another's specialized methods. AI can help synthesize literature, connect disciplines, and identify relationships that would be difficult for an individual to find.

This matters especially in biology, where our understanding continues to change. Scientific rules are useful models, but discoveries can reveal that they are incomplete. Systems capable of processing many variables and large bodies of evidence may help researchers explore patterns that exceed unaided human memory.

The opportunity extends beyond research laboratories. AI can help organize information for clinicians and make explanations more accessible to patients, potentially allowing patients to participate more fully in diagnosis and treatment discussions.

A personal example, with an important boundary

I experienced this when trying to distinguish vertigo from symptoms related to low blood pressure after starting a prescribed medication. Describing the experience to an AI tool helped surface the blood pressure explanation, which had not initially been identified in my conversations with clinicians.

That was useful, but it was a personal experience, not evidence that AI is better than physicians. A clinical examination could have provided information that was unavailable over the phone. The broader point is that AI can help generate and organize possibilities for further evaluation.

Helping someone ask a better medical question is different from safely replacing medical care.

Robotic surgery illustrates the value of collaboration

Fei-Fei described her father's liver surgery, performed by a surgeon using the da Vinci robotic system. The surgeon controlled the robot. This was human expertise combined with a technical tool, not an autonomous machine independently performing the operation.

Her subsequent conversation with the surgeon raised an important question: could enough surgical data eventually train an autonomous system?

The answer was not obvious. The liver is highly vascular, and anatomy varies between individuals. Even combining many operations might not provide sufficient examples of every important situation.

This reveals a general limitation. When relevant patterns are rare or insufficiently represented, confidence in a learned system must be limited. A human working with a robot may be far preferable to a robot operating beyond what its training supports.

Simulated organs and training environments could become useful research tools, but their ability to support safe autonomous surgery remains a question to investigate, not an outcome to assume.

Children Need Both AI Access and the Effort of Learning

Concerns about AI and education often become a choice between two extremes: let students use it for everything, or keep it away from them entirely.

Neither approach adequately respects how brains develop.

Learning takes time and effort. It can involve frustration. Tools that repeatedly replace that process with passive consumption or finished answers can undermine the motivation and agency that learning requires.

But denying useful tools creates another problem. A motivated student may be able to learn much more with access to explanations and guidance than without them.

Fei-Fei used organic chemistry as an example. As a student, she struggled with molecular structures and orientations while facing limited access to teaching assistants and office hours. An AI learning companion could have offered repeated explanations at the moment she needed them.

The goal is not to remove the intellectual work. It is to help a learner do that work more effectively.

• The harmful outcome: The tool takes over, and the student loses the motivation to learn.
• The other harmful outcome: Fear of misuse prevents a motivated student from accessing useful support.
• The constructive outcome: The student retains responsibility for learning while using AI to explore questions more deeply.

This is not only about children. Adults retain neuroplasticity and can also develop new ways of learning with these systems.

Prompting is a skill worth teaching

The quality and specificity of a question can strongly influence the usefulness of an AI response. Providing context, identifying where understanding breaks down, and asking follow-up questions are skills.

Fei-Fei's comparison was Socrates: learning through questions, clarification, and continued examination. That is a better educational model than treating a chatbot as an answer dispenser.

There is also a social benefit. Straightforward background questions can sometimes be explored with AI before taking another person's time, allowing human conversations to begin at a more thoughtful level.

Embodied AI: Intelligence Beyond Words

Language is powerful, but intelligence is not limited to language. Humans perceive space, move through environments, manipulate objects, and communicate through faces, voices, and bodies.

Work by neurosurgeon and speech researcher Dr. Eddie Chang illustrates how meaningful those additional dimensions can be. Systems that translate neural activity into communication have helped people with severe paralysis express themselves.

Adding a personalized visual avatar and a familiar voice can create a much richer interaction than synthetic speech alone. Facial expression and vocal characteristics help restore aspects of communication that text or a microphone cannot convey by themselves.

Robotics extends this question into the physical world. What could AI do if it could help with the tasks that occupy so much human time and effort?

Where robots could make everyday life better

Fei-Fei's examples were grounded in real needs rather than spectacle:

• Caregiving: Helping with physical tasks while leaving love, responsibility, and family communication intact.
• Hospital support: Fetching supplies and reducing some of the walking and logistical work that burden nurses.
• Disaster response: Reducing the need to put people in dangerous environments such as wildfires.
• Independent living: Helping older adults obtain groceries, access medication, or move through their communities.

These are possibilities, not promises of imminent availability. Hardware adds complexity, and robotics generally takes longer to develop and deploy than software.

Design also matters. A world filled with separate, rigid machines may not feel welcoming or practical. People should help determine how robots look, how they share space with us, and which jobs we actually want them to perform.

The public should not be left merely reacting to a future designed elsewhere.

Spatial Intelligence and the Work of World Labs

The move beyond language is central to World Labs, which Fei-Fei co-founded in early 2024.

The company's focus is spatial and physical intelligence: building models that can support the generation of three-dimensional environments and worlds that evolve over time.

The idea is not simply to photograph more of the existing world. It is also to help people express what they imagine. A sentence, image, or sketch can provide a starting point for creating an environment.

Potential applications include:

• Environments for training robots.
• Tools for architecture and design.
• Creative production and visual effects.
• Interactive settings for education and healthcare.

This is another example of AI as a means of expression. A person may know what they want to create but lack the time or technical resources to construct every detail manually.

AI, Filmmaking, and the Future of Creative Work

AI tools are increasingly capable of turning written descriptions into video shots. Short films and longer productions have been assembled using these tools.

That does not make the human contribution disappear. Stories involve choices about characters, emotional meaning, perspective, pacing, and what deserves to be told. Camera movement and visual style are not merely technical outputs. They can be expressions of how a particular person sees the world.

The technology conversation and the employment conversation need to be distinguished, even though they are connected.

There are genuine concerns about how actors, screenwriters, visual effects artists, and other creators will be affected. Saying that a tool empowers creativity does not resolve those concerns by itself.

Constructive progress requires collaboration between technologists and people who understand the craft and economics of the industry. Fei-Fei described conversations with filmmakers, including Ben Affleck, as an example of the dialogue needed.

Industries can change rather than simply vanish. Photography changed profoundly with digital technology, and creative professionals continue to adapt and develop new skills. But transitions involve both opportunity and loss. That deserves thoughtful attention, not dismissive reassurance.

Why AI Governance Requires More Than Technology Companies

Technical capability does not automatically justify deployment.

A system may be able to generate a convincing imitation of someone's face or voice. That raises questions about permission, identity, deception, and social trust that cannot be answered by asking whether the software works.

Other fields already recognize this distinction. Human research involves ethical review. Medicine has professional responsibilities and regulatory oversight. Engineering involves safety standards.

AI needs a similarly layered approach:

• Education and professional norms: Technical training should include ethics and societal consequences.
• Institutional safeguards: Organizations need processes for evaluating how systems affect people.
• Appropriate regulation: Oversight should reflect the application, including areas where AI intersects with medicine and biology.
• Public participation: People affected by these systems need a meaningful role in shaping them.

These questions helped motivate the creation of the Stanford Institute for Human-Centered Artificial Intelligence.

Market incentives matter, but they are not identical to ethical responsibilities or societal goals. No individual company or small group of industry leaders should be the sole authority deciding how everyone else will live with AI.

Teachers and Parents Cannot Be an Afterthought

One of Fei-Fei's strongest concerns is that teachers and parents are often left out of AI development and public discussion.

They are expected to manage consequences they did not design. Teachers must respond to questions about cheating, assessment, learning, and student motivation, often without enough guidance or resources.

After ChatGPT's public release in November 2022, Fei-Fei contacted her child's elementary school principal and offered to speak with students and teachers. Her point was not that every educator needs a visiting AI scientist. It was that the people responsible for children's learning deserve timely explanations and practical support.

Questions about misuse are reasonable. Educators should be able to explore those questions with technologists, not be lectured or dismissed.

Children are curious and adaptable. That is reason for optimism. But optimism becomes more credible when we support the adults helping them grow.

A Human-Centered Future Is Something We Build

Public discussion of AI often swings between catastrophe and effortless utopia. Neither extreme helps someone decide how to use a tool, teach a class, care for a parent, or evaluate a medical application.

A better discussion begins with specific questions. What problem are we trying to solve? What information does the system actually have? Where is human judgment essential? Who benefits, and who bears the risks?

There are already scientists, designers, entrepreneurs, clinicians, and educators working on those questions. Their work deserves attention alongside the loudest claims about the future.

The opportunity is substantial: deeper learning, more capable scientific discovery, better support for caregivers, and new forms of creative expression. Realizing it requires neither blind trust nor blanket rejection.

AI should help us become more capable participants in our own lives. Keeping that standard central is how technological progress can enrich humanity rather than diminish it.