My guess is human learning requires some major differences in training.

For instance, we learn to process a firehose of visual, sound, tactile and motor information, before we start thinking in well-defined concepts.

That is a major foundation, grounding, and highly organized state, on which we learn language. And is entirely missing for LLMs.

We learn a lot metaphorically, so it may be that the physical effects we learn to recognize act as highly efficient sub-vocabulary, for quickly learning new concepts both verbal and nonverbal.

So a series of different learning stages, each making the next more efficient, is likely to be a big part of any solution.

No amount of data/exposure will help a baby trade stocks. General fast learning requires other things to be learned first.