Most artificial intelligence you might have encountered analyzes existing information. It can categorize photos, identify spam in your email, or translate languages. This is incredibly useful, but it's fundamentally about understanding what's already there.

Generative AI is different. It doesn't just analyze; it creates. It learns the underlying patterns and structures from vast amounts of data—like text, images, or music—and then uses that knowledge to produce something entirely new.

Think of it like the difference between a music critic and a composer. A critic analyzes a symphony, breaking it down and explaining its parts. A composer uses their understanding of musical theory and patterns to write a completely new piece. Generative AI is the composer.

The idea of a machine that creates isn't new, but for a long time, the results were clunky. Early attempts relied on rigid rules, which struggled to capture the nuance of human creativity. The real breakthrough came when models started learning directly from data. This shift allowed them to pick up on the subtle patterns that define a style, whether in writing, art, or code.

Two early and influential approaches paved the way for the powerful models we see today: Generative Adversarial Networks and Variational Autoencoders.

Generative Adversarial Networks, or GANs, work through a clever kind of competition. Imagine an art forger (the generator) trying to create a convincing fake Picasso. At the same time, an art critic (the discriminator) is trying to spot the fakes.

At first, the forger is terrible, and the critic easily identifies the fakes. But with each attempt, the forger gets feedback and improves. The critic, in turn, must get better at spotting more convincing forgeries. This back-and-forth continues until the generator becomes so skilled that its creations are nearly indistinguishable from the real thing.

Variational Autoencoders, or VAEs, take a different route. Instead of competition, they focus on compression and reconstruction.

A VAE first learns to compress an input, like an image of a face, into a much simpler, condensed representation. This is the encoder part. It’s like summarizing the key features of a face—eye shape, nose length, etc.—into a short description. Then, the decoder part tries to recreate the original image using only that compressed description. By learning to do this accurately, the VAE gets very good at understanding the essential features that make up a face. It can then be used to generate new faces by creating new compressed descriptions and decoding them.

While GANs and VAEs are often discussed in the context of creating images, the applications of generative models extend far beyond digital art. They are being used across a wide range of fields to create, simulate, and innovate.

These foundational models helped establish the principles of teaching an AI to be creative. They set the stage for the more complex and powerful models, like Large Language Models, that are becoming increasingly common.

Let's check your understanding of these core ideas.

Understanding these early models gives us a solid foundation for exploring the more advanced architectures that power today's generative AI applications.