Deep learning has achieved human-level performance on reading radiology scans, describing images with idiomatic sentences, playing complex games, and more. Deep learning is however compute-intensive, and in the development of deep learning frameworks something of a trade-off has emerged: increasing usability tends to negatively affect speed, and vice versa. Popular frameworks Caffe, Tensorflow and Theano provide quick computing performance but at the cost of ease of use and flexibility. Then there's PyTorch, developed primarily by Facebook AI and introduced in 2016. A new paper from original PyTorch developers Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan and 17 other researchers explores the inspiration behind the library, and makes the case for its unique marriage of speed and usability.