Why do AIs need so much more data than children to learn?
Serious leads
Why it matters
Children learn language from far less data than AI models. Understanding why could shed light on how human minds learn, and help build AI that makes better use of data at a time when developers risk running out of high-quality training text.
What we know
A child hears tens of millions of words; large language models train on trillions.
What blocks us
We don't know which built-in abilities let children learn so fast.
Current leads
Training on child-scale data, learning through interaction with the world.
Sub-questions
What built-in knowledge, if any, helps babies learn language from so few words?
How much does seeing the world and talking with people, rather than just reading text, speed up language learning?
Can a language model trained on only as many words as a child hears reach human-like language skills?