Vision Is the Cornerstone of Intelligence, and AI Should Augment Rather Than Replace Human Agency
Vision is the cornerstone of intelligence; AI must augment human agency rather than replace it; and humanity's most precious cognition has never been uploaded to the internet.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The starting point of intelligence is not language, it is vision
Fei-Fei Li places vision at the core of intelligence, for reasons on two levels. In evolutionary history, 540 million years ago marine animals such as trilobites gained the ability to sense light for the first time, and the Cambrian explosion followed; in AI, vision has driven modern artificial intelligence on both the algorithm and the data side. She also notes that roughly half of human cortical activity is related to visual function. Understanding vision, then, is not just understanding the eye — it is understanding where intelligence itself begins.
— Fei-Fei LiThe bottleneck for AI was never algorithms, it was data
When Fei-Fei Li was an assistant professor at Princeton in 2006, she realised the real bottleneck for AI algorithms was the lack of data. Inspired by cognitive neuroscience, she decided to build a large-scale dataset, and that became ImageNet: 15 million images, with the goal of teaching machines to recognise everyday objects. Before 2012, machines performed far worse than humans, whose error rate on the 1,000-category recognition task was about 4%; in 2012 neural network algorithms sharply cut the error rate, marking the inflection point, and by around 2016 algorithms surpassed humans. ImageNet brought algorithms, data and GPUs together, and became the turning point of modern AI.
— Fei-Fei LiSora does not understand a cat's muscles, and still makes the cat run
Around 2023 several teams added video to their training data, and in January 2024 Sora was released: users type text and it generates a few seconds of video, such as a cat running toward a mouse. Fei-Fei Li stresses that the algorithm does not understand the cat's muscle structure; it has simply learned plausible patterns of cat motion from a huge amount of video. This is similar to how humans learn by observation — not biological knowledge first, but massive data first. The significance of video generation is that it moves AI from static images to predicting a dynamic world.
— Fei-Fei LiThe boundary of AI is not compute, it is the thoughts never digitised
Fei-Fei Li agrees that AI cannot capture highly individualised, unrecorded human cognitive behaviour — for example the unique ideas Picasso had while creating. The internet is the largest multimodal collection of human behaviour, but the thoughts and emotions that were never uploaded are not in it, and no matter how powerful AI becomes it cannot access them. On this basis she argues humans remain unique, while AI can also combine existing information in creative ways. The boundary of AI is not compute, but whether humans have digitised a given cognitive process.
— Fei-Fei Li"You don't understand" is the most dangerous sentence in AI
Fei-Fei Li stresses that AI should augment human agency, not strip it away. She criticises some leaders in the AI field who talk about AI in condescending language, implying AI will replace humans — rhetoric that is neither healthy nor useful. She also points out that "you don't understand"-style public rhetoric is dangerous, and argues for public education so people understand the pros and cons, while the final choice must stay in the public's hands. Genetic testing is one example: there was fear early on, but people ultimately need to decide for themselves.
— Fei-Fei LiAI will become scientific infrastructure, not just generated conversation
Fei-Fei Li says AI will fundamentally change how science is done. Traditional science relies on the intelligence and speed of the human brain, but AI can store vast amounts of information and synthesise knowledge across disciplines, which brings enormous opportunities in fields such as biomedicine. Using vision and smell as examples, she explains that AI can cross the knowledge boundaries of any single human expert and connect different fields. That capability makes AI not just a generator of conversation, but part of the infrastructure of scientific research.
— Fei-Fei LiThe best prompt writer in history is Socrates
Fei-Fei Li worries that AI may strip young people of agency and motivation to learn — endless short-video scrolling, for instance; but she opposes throwing out the baby with the bathwater and banning students from using AI outright. With the right guidance, she argues, AI can be a powerful learning tool that makes this generation smarter than us. She even offers a quiz: who is humanity's best prompter? The answer is Socrates — his method was itself about seeking truth through questioning. Prompting is not a technical detail; it is part of public education.
— Fei-Fei LiLanguage is not the endpoint, spatial intelligence is the next frontier
Fei-Fei Li founded World Labs in early 2024, focused on spatial intelligence and physical intelligence, which she sees as the next frontier of AI. The company's foundation models generate 3D and 4D worlds, serving creators, robot training and architectural design, with goals clearly beyond language itself. She also repeatedly stresses that humans should actively imagine and design the future of AI, and must not leave it to companies or investors to decide unilaterally. Spatial intelligence is the next landing point for the judgment that vision is the cornerstone of intelligence.
— Fei-Fei LiIn their own words · checked verbatim
I see vision as a cornerstone of intelligence.
Fei-Fei Li4:37
we need to think about AI as a tool that helps us in our agency. It it should not take away our agency.
Fei-Fei Li49:41
The absolute bad outcome is that our young generation. There. Agency. And human level motivation of learning and living is taken away by tools.
Fei-Fei Li1:29:33
Who is humanity's best prompter? I'm going to flunk this quiz. Socrates, if he were alive. Because that is the method of prompting. Right, think about it. What is Socrates method? Is prompting and seeking truth by asking questions.
Fei-Fei Li1:35:00
They are the most important people in our society. We should be talking to them. We should be uplifting them. We should be supporting them. We should be providing resources to them.
Fei-Fei Li2:02:02
Figures
| When animals first gained light sensitivity | 540 million years ago | 4:37 |
| Share of human cortical activity related to vision | About half | 6:53 |
| Number of images in ImageNet | 15 million | 9:24 |
| Human error rate on the ImageNet challenge | About 4% | 16:03 |
| Father's surgical blood loss compared with a typical operation | 10 times less | 1:07:02 |
| Protein content of a David protein bar | 20 grams | 1:19:36 |
| Calories in a David protein bar | 150 | 1:19:36 |
| When World Labs was founded | Early 2024 | 1:50:48 |
| When ChatGPT was released | November 2022 | 2:03:03 |
Glossary
- ImageNet
- The large-scale visual dataset Fei-Fei Li led the construction of: 15 million images across 1,000 object categories.
- Spatial Intelligence
- The ability to understand and generate three-dimensional space, which World Labs sees as the breakthrough point of AI's next stage.
- Embodied Intelligence
- Intelligence in which AI interacts with the real world through a physical body, for example a robot caring for the elderly.
- Multimodal
- Processing text, images, sound and other forms of information at the same time; the internet is regarded as the largest multimodal collection.
How to listen
Educators, parents, AI founders and engineers, and anyone concerned with the boundaries of technology and public discourse.