Children still learn language better than any model, and nobody can explain the gap
MIT Technology Review·medium signal
MIT Technology Review's August 24 piece frames the comparison starkly: for 100,000 years exactly one thing could learn a human language to full fluency, and now there are two, but the child does it on orders of magnitude less input. The unresolved question is what the child has that scaled pretraining does not, and the article treats the answer as genuinely open rather than settled by any current curriculum-learning or grounding hypothesis. Relevant to anyone reasoning about sample efficiency ceilings rather than parameter counts.