Optimized for Engagement: Artificial Intelligence and the Future of Entertainment

You may have seen this striking chair designed by Autodesk's generative CAD software: It began with a primitive description of a chair: a platform of certain dimensions held above the ground by a certain height, able to carry a certain amount of weight. From there, a computer tested billions of possible configurations, evaluating each one for strength, stability, and material usage. As it worked, the software developed what's known in machine learning as a "gradient"--a map that relates how outcomes change with respect to objectives as you vary inputs. In this case, the inputs are bits of matter added to or subtracted from any part of the chair, and the outcome is a combination of strength, stability, and material usage. As the computer varied the design of the chair, arbitrarily at first, it began to identify the most promising avenues of design and spent its efforts in those directions. "Gradient descent"--this process of informed search through an almost infinite field of possibilities--is one of the foundations of artificial intelligence, and the ability to perform gradient descent operations has that has advanced rapidly in the last five years, thanks to both algorithmic advances and cheaper, more powerful computing.