Like we have 50+ years of science on how human children, and how their parents and caregivers teach them.
We know humans are way more “sample efficient” than current LLMs. E.g. children seem to acquire language and basic facts with orders of magnitude less data than what goes into large model pretraining.
I’m not saying that everything will carry over. But I do think AI researchers should pay more attention to how toddlers actually learn
もっと見る
