Babies Learn Language Far More Efficiently Than AI Chatbots, Experts Say

MIT Technology Review reports that babies are far better at learning language than AI chatbots, and much more efficient. Stanford cognitive scientist Michael C. Frank says that despite recent progress, training an AI model requires data equivalent to 'burning down a forest and scraping nearly all human knowledge' to match what a toddler naturally achieves in about a year at home. Toddlers start speaking clear sentences after hearing 10 to 30 million words, while training GPT-2 with similar data would just produce gibberish. This efficiency gap is a big challenge for AI companies hoping to reach human-level intelligence by scaling up models, and many experts doubt that path will work.