In their first years of life, a child hears only millions of words โ and that's enough to start speaking. Training a powerful AI model, by contrast, takes trillions of words. How children manage this, scientists still don't fully understand. Researchers at Princeton recorded the speech children hear in their first three years and trained an AI on it.
The model learned thousands of words โ but only after it was given an analogue of human sleep.
๐ป Cameras and microphones were set up in several families' homes, recording most of each child's day until they were nearly three. Then, for each of the eight children, the researchers built a separate model that learned only from what that child heard, day by day.
It started from scratch, with no idea where one word ended and the next began, and it never saw written text. At first, the system barely retained what it heard from one day to the next, and by the equivalent of age two it hadn't learned a single word.
Then the model was given a way to revisit the day โ an analogue of sleep. "At night," it replayed recent recordings little by little, and learning took off. The model mastered the sounds of English and, by "age two," learned to recognize thousands of words.
โก๏ธ The approach worked across all eight families, even though conditions varied. But a new problem emerged: the children developed at different rates, while the model progressed at the same pace in every home. The researchers adjusted three parameters: how much the model "replays" at night, how quickly it learns, and when it starts remembering words. After that, each model developed at the pace of its own child.
"For more than a century, psychology studied development from the outside. Now we can approach it from the other side: build a learning system, give it a child's real experience, and ask whether it develops the same way," says project lead Uri Hasson.
Who's smarter?
โค๏ธ โ A three-year-old child
๐ค โ GPT-6 Astra
@hiaimediaen
