Understanding how neural networks learn through interactive visualizations

TL;DR -- Interpretability might be way more tractable than we thought, and the guy who helped convince everyone otherwise is now saying we can understand models by directly studying neurons and their interactions...