[gwern]

OP's mind would probably be absolutely blown by the demo scene if he thinks 10MB of data is impressive. Also, the NN example is interesting because no one in deep learning thinks that you actually need all those parameters: most NNs can be shrunk by 90%+ with some simple tweaks like quantizing neural weights and pruning them (speculation is that you need to train large parameterized NNs to make the path through high-dimensional space to the optimal program smooth enough for gradient descent to travel).