Kids outlearn AI–and we still don't know why (www.technologyreview.com)

A new study highlights a significant data efficiency gap between human language acquisition and the capabilities of large language models (LLMs) like ChatGPT. While LLMs can process vast amounts of text—often training on trillions of tokens—children only need exposure to around 100 million words to achieve fluency in their native languages. This stark contrast poses profound questions for cognitive scientists and AI researchers about how children outperform these sophisticated models, despite their seemingly limited experience. Experts suggest that unraveling the mechanisms behind children's language learning could lead to the development of more efficient AI systems that require less data. The implications extend beyond mere curiosity. Exploring how kids master language can inform the creation of AI tools that are not only more efficient but also tailored to serve underrepresented language communities. The emergent BabyLM competition is a notable example, pushing researchers to build AI models that learn from smaller, developmentally appropriate datasets. As the race to improve AI capabilities continues, understanding the nuances of human language learning may bridge the gap, leading to more human-like communication in machines and offering insights into fundamental cognitive processes.