Today's AI Cannot Know Why That Gray Cup Matters to You
Fei-Fei Li says AI learns only what humans upload, so our most private memories stay out of reach.
Andrew Huberman · Using AI to Increase Your Intelligence & Enrich Humanity | Dr. Fei-Fei LiToday's AI is missing something essential, computer scientist Fei-Fei Li argues, because it can learn only what humans have managed to upload.
That gap includes thoughts, creativity, and abstract art that makes people feel something they cannot put into words. AI learns from public material, Li says, so private experience stays outside its reach. "That thought is not captured therefore it's not on the internet therefore AI has not seen it," she said.
By internet, Li meant something specific. She called the internet "the biggest collection of human behavior in multimodal forms," built from decades of typing, from teenage chatter to digitized scientific articles, plus digital photos, videos with sound and movement, and music and speech. Because so much language and imagery is already there, she said, AI is powerful at recognizing patterns and combining them.
That is why, Li argued, even a profound private thought like the one that led Picasso to paint a young woman in a particular way remains outside AI's reach. No one captured that thought when it happened, and neuroscientists cannot even say which brain area produced it.
Li gave AI credit for a different kind of creativity. She pointed to Move 37, when the Go-playing program AlphaGo, in its third game against human masters in the ancient board game Go, made a move that human masters had never considered. But she added that Go has very clear mathematical objectives and rules, which let a computer use its memory and computing power in a way human brains typically do not.
Li said her current guess is a hybrid, with humans working alongside AI to solve "these problems whose solutions have yet to be invented."
Then she made the limit personal. She imagined looking at an ordinary gray cup that evokes a childhood moment shared only with a best friend. Her reaction to the cup, and what she would do with it, would come from a memory AI has never seen.
Questions people ask
What exactly does Li count as being on the internet?
She means the vast public record of what people type, photograph, film, and record. That includes years of messages and articles, smartphone photos, videos with sound and motion, and music and speech. AI trains on that library, so it is strong at language and patterns found there.
Couldn't AI just predict my memory from similar stories other people shared?
It can guess based on what many people have shared, such as typical reactions to cups or childhood nostalgia. But Li argues it cannot know your specific association if you never wrote it down, photographed it, or recorded it. Without that data, the model has no pattern to learn from.
Does she think AI can be creative at all?
Li says yes, in a certain way. AI can combine what it has learned in new and surprising ways, and she points to AlphaGo's Move 37 as an example. Her distinction is that this is different from inventing out of private experience, especially for problems whose solutions have not yet been invented.