Children acquire language using significantly less data than large language models (LLMs), according to a report by MIT Technology Review AI. Why this difference exists remains unknown.
Understanding this distinction could help create more efficient language models while simultaneously providing new insights into the development of human cognition.
Available data is limited to the publication's synopsis and lacks details on the comparison methodology, the magnitude of the gap, or specific models. Therefore, it is premature to draw conclusions about the causes.
editorial commentary
Why it matters
A likely consequence is increased interest in training methods focused on data efficiency, as well as in child development research. The next verifiable signals will be published details of the comparison methodology and results from work linking human learning to model improvements. Significant uncertainty remains: the source reports neither the exact scale of the difference nor its cause.