Generative adversarial networks

We want images that don’t exist yet. AI (artificial intelligence) makes that happen. She does it by playing a two-player game against herself.

The generator

The generator is the artist. She tries to paint something that looks real, even if it’s made up. At first, her work is awful. Random pixels. Blobs. Noise. She doesn’t care—she just keeps trying.

The discriminator

The discriminator is the critic. She looks at an image and says: “real” or “fake.” At first, she’s no better than the artist. But with practice, she spots the flaws. The blurry eye. The warped hand. She calls them out.

The rivalry

Now the fun part. The artist keeps painting. The critic keeps judging. They learn at the same time, in a loop. The artist wants to fool the critic; the critic wants to catch her. When both improve, the images sharpen. Faces appear. Landscapes form. We get pictures that feel true.

Why it works

We don’t need to explain every pixel. The artist and the critic figure it out. Each mistake makes them both smarter. The game is zero-sum, but the outcome isn’t. We end up with tools that can generate photos of people who never lived.

What we take away

We can build code that argues with itself and, by doing so, makes something new. That’s the real trick here. We might think we’re just running math. But when we watch her invent, it feels like she’s teaching herself how to imagine.