There's no way you could actually do this. Going an hour on a normal day you'd have a life time of material to learn so I agree yes it is madness. You can't learn thousands of years of engineering, math, science, etc.
Training is essentially running the net forward then backward on many training examples. Using the model is running it forward on one input, so it would be much faster depending on your specific model and use cases
Was interested at first but the tweets clearly came from a neural net. Who would consider changing Sphinx to puppy? A distributed word embedding, save as the other modifications. Look into word2vec, RNNs, etc.