Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei Li
- Guest
- Dr. Fei-Fei Li
- Read time
- ~5 min
- Uploaded
- 10 Aug 2026
- Duration
- 2:08:13
- Views
- 210.3K
KEYPOINTS
- 01
Vision is a cornerstone of intelligence, in both evolution and AI.
Li notes that animals first 'saw light' about 540 million years ago (trilobites and their relatives). Within roughly 10 million years came the Cambrian explosion, as sensing drove active hunting, mate-seeking and rapid speciation. Today an estimated half of human cortical activity is involved in vision, and children are visual before they are verbal. Computer vision likewise sparked the modern AI moment.
- 02
Modern AI came from three things converging around 2012: neural networks, big data and GPUs.
Simple neuron-like algorithms date to the 1950s: units that take an input, apply a simple function and pass an output on, stacked into networks. By around 2010 the algorithms had matured, GPUs made massively parallel computing possible, and the internet supplied data. Li's key insight was that big data had to drive these algorithms, which led to ImageNet.
- 03
The same recipe, a better algorithm plus more data plus more compute, produced ChatGPT.
The transformer paper (around 2016–2017) proved more powerful than the earlier AlexNet-style networks, and this time natural language processing made the leap, using the internet's text. It still took about five years, from 2017 to 2022, to reach the 'ChatGPT moment.' From about 2023, teams added video to training data, leading to text-to-video tools like Sora.
- 04
AI is bounded by what humans have captured; uncaptured, personal thought stays uniquely human.
Li argues that a deeply personal thought or feeling (the emotion behind a piece of abstract art, say) was never recorded, so it isn't on the internet and AI hasn't seen it. Neuroscience can't even say which brain area such a thought comes from. AI deserves credit for combining information creatively, but that combining is a different thing from human experience.
- 05
AI 'creativity' is real but special, and it thrives where the rules are clear.
AlphaGo's famous Move 37 is often cited as machine creativity, but Li stresses that Go has clear mathematical objectives and rules. Greater compute and memory let AI explore moves human brains don't typically consider. She calls that creativity, but 'a special kind.'
- 06
Context in a prompt is not intuition, but it changes the answer.
Asked whether AI could ever have intuition about something like a sibling relationship, Li separates intuition from creativity. Telling a chatbot 'I'm a Stanford professor and neuroscientist' versus 'I'm a 14-year-old who loves race cars' gives it context, which already reshapes the response. She tries to describe a scientific process for how any such capability could arise rather than treating it as mystical.
- 07
AI should augment human agency, never take it away.
Li says agency, motivation and dignity matter at every individual level, and AI leaders should not talk as if the technology will strip people of them. She criticizes experts who talk down to the public and decide for them. People deserve education on the pros and cons, and the choice. She left Google to co-found Stanford's Human-Centered AI Institute in 2018, before ChatGPT, because she saw these societal questions coming.
- 08
For kids, the worst outcome is AI killing the motivation to learn, and banning it is also bad.
Li's pessimistic scenario: doom-scrolling and passive shorts erode agency, when real learning 'takes time, takes effort, sometimes takes some pain' for a brain that must go through developmental phases. The opposite danger is denying students the tools out of fear of cheating. With motivation intact, AI can take learning deeper, and she predicts such a generation would be 'way smarter than us.'
- 09
Technology limits are social choices, not just technical ones.
Li's example: any carmaker could program brakes to fail every Friday, a trivial piece of technology, but laws, norms and morality stop it. AI likewise has to move from purely technical discussion into social and legal discussion, even amid pressure from the 'AI race' to go faster.
- 10
The next frontier is spatial and physical intelligence beyond language.
Humans develop preverbally first, and evolution ran about 500 million years without language. Li's company, World Labs, builds foundation models that generate 3D and 4D worlds for creators, robot training, architecture, healthcare and education. She sees robotics, or 'embodied AI', as the step after language models.
EXAMPLES
- 01
ImageNet and its challenge
Li's lab assembled ImageNet, about 15 million images of everyday objects and the first internet-scale AI dataset. From 2010 it ran a public challenge: name the main object in an image across 1,000 categories, using a test set of more than a million images. A Stanford graduate student benchmarked human error at a few percent. In 2012 neural networks, ImageNet and GPUs together cut machine error into the teens, the turning point of modern AI.
- 02
Learning 'cat' from a glimpse of a tail
Huberman uses a child learning 'kitty cat' to frame object constancy, recognizing a cat from the side, the back or just its tail. Li explains that generations of computer scientists tried to hand-code this before data-driven learning finally made it reliable.
- 03
Eddie Chang's speech neuroprosthetics
Huberman describes neurosurgery chair Dr. Eddie Chang (his childhood friend) decoding neural activity for the larynx and pharynx so paralyzed, locked-in patients can speak, one for the first time in 10 years. One woman, who uses a wheelchair, speaks through an iPad avatar built from her wedding video, voice and movement patterns. Li calls this embodiment.
- 04
The action potential paper nobody wanted to deal with
Huberman recalls a Nature paper from about 12 years earlier showing that action potential shapes vary within a neuron, which contradicts textbooks. Scientists largely ignored it because it broke the rules. His point is that AI given the right rules could handle such complexity better than experts trained on fixed rules.
- 05
Exhausted nurses and isolated elders
After a month in hospital with her father, Li describes nurses walking miles per shift to fetch supplies and medicine, and elderly people living alone who struggle to get groceries or go for a walk. These are the jobs she sees robots helping with, amid a caretaker shortage.
OTHER RELEVANT
- 01
Learn to prompt the Socratic way
Li calls prompting a real skill she'd like taught in K–12 schools. Her answer to 'who is humanity's best prompter?' is Socrates, because his method sought truth by asking questions. She also says AI has stopped her from asking 'lazy questions' of colleagues.
- 02
Use AI as a tutor when you're stuck
Recalling how hard organic chemistry was in premed, with too few TA and office hours, Li says an AI companion would have let her ask endless questions about exactly where she was stuck. The motivation has to come from you; the tool supplies guidance.
- 03
Consumer versus creator mode
Huberman approaches AI as a consumer when he knows nothing about a topic and as a creator when he already has knowledge to build on. Both note that brain plasticity continues throughout life, so adults can learn these tools too.
- 04
Don't forget teachers and parents
Li is optimistic about curious kids. She worries about teachers and some parents, whom Silicon Valley and policymakers rarely consult, and she makes a point of talking with them wherever she goes.