Humans learn in a very data efficient way.

European child hears 500 to 1000 hours of speach per year (and becomes a fluent speaker in a few years).

Tsimane child hears about 60 hours of speach per year (and learns to speak fluent Tsimane)

LLM's require thousands of times more data to build a moderately competent statistical model of language.

Humans practice one-shot learning very robustly.

Me: Your cat is an expert purrgetter. Again you melted and gave her extra snacks.

You: My cat purrgot me yeasterday again. (Irregular past tense)

You inserted the word in a vast system of symbols and rules, and you did it in one shot.

Humans learn by inserting new knowledge in already existing networks.