And even if "me learning something" means different neural rewirings than "you learning something" and isn't transferable, you could instead have 100 instances of yourself go and learn 100 different things, then combine all the rewirings at the end. There might be "merge conflicts", but I suspect it would still mostly work—and you could parallelize the work of finding ways to fix the conflicts! Even if none of that worked, at the very least you could have your branch-instances discover by experiment the fastest way for you personally to learn each thing, then have your main instance follow an optimal educational program, which I will arbitrarily proclaim to be 4x faster than normal learning.

books. You just said books. Only faster.

Books contain thoughts encoded in words. And words can be ambiguous at times. Writing is a skill, and not everyone is good at it. It is a lossy process.

Good point. But I imagine that thought transference must be the same - human brains are messes of neurons connected in patterns formed by birth and experience, and a thought process that works fine in one brain may not map correctly to another. In any machine capable of learning, the same kind of semi-random process must occur, and mapping a pattern from one will not fit correctly with another.

You can't take half the nodes from a neural network that recognises dogs and add them to a neural network that knows about laundry to get a network that knows how to wash dogs ;)