My boss asked me to watch the CS231n lectures and learn TensorFlow, but he didn't give me a dataset. So I downloaded the history of the 9gag hot page, over 60k images. I'm also interested in learning languages, and I find that comics with dialogues are a great way to learn how to talk to people. I then wrote my own image viewer to let me tag pictures as a comic/not comic.
So now I'm still wasting time looking at memes, but I can pass it off as "productive" work, training a machine learning model to find dialogues for language learning.