What if two neural networks competed against each other - could they create things that never existed?

The answer

In June 2014, Ian Goodfellow introduced Generative Adversarial Networks (GANs) while at the University of Montreal, reportedly conceiving the idea during a discussion at a bar. A GAN pits two neural networks against each other: a generator that creates fake data and a discriminator that tries to tell real from fake. As they compete, both improve, and the generator eventually produces astonishingly realistic outputs. Yann LeCun called GANs 'the most interesting idea in the last 10 years in ML.' GANs went on to power photorealistic face generation, image-to-image translation, art creation, and unfortunately, deepfakes.