Deep Learning has led to breakthroughs in many previously unsolved problem domains, from image classification to machine translation to medical imaging analysis. Venture capital firm Andreessen Horowitz recently cooked up an AI playbook, which posits that AI will impact software as broadly as relational databases have since the late 20th century. It’s hard to think of a technological problem that AI doesn’t touch.

In this talk, we will explore the limits of today’s most popular approaches to AI. In particular, what kinds of problems can’t we solve today and how might the solutions shape the way we approach software development? Training a model for your particular domain is easier than ever, but why is it so difficult to make sense of what is going on inside the model? How can we move toward a more intuitive and accessible model for understanding what our AI has learned?