What is Generative AI?

Generative AI is a type of artificial intelligence that enables users to generate novel outputs resembling human-created content, including text, images, videos, sounds and 3D models. Recent breakthroughs in the field have drastically improved the capabilities of these models, making generative AI the fastest adopted technology in history.

The crux of generative AI is in its name – generative. While traditional AI algorithms have focused on identifying patterns within data for predictive purposes, say to predict whether the next image in a sequence of images is a cat or a dog, generative AI leverages the learned patterns to create entirely new outputs. Given the above example, a generative AI model could create an entirely new species of cat, or dog, or cat-dog!

How does Generative AI work? (SH)

Generative AI is grounded in machine learning techniques that draw inspiration from the neural systems of the human brain, known as neural networks. These neural networks are ‘trained’ on extremely large amounts of data, during which they learn to capture and internalise the essential features, patterns and relationships within the data. This enables the model to generate new data instances that are similar yet distinct from the original training data.

For instance, a generative AI model trained on images of faces learns to understand facial features like nose shape, eye placement, and smile curvature, allowing it to generate new faces that possess the same structure and features but do not match any specific face. Similarly, models like ChatGPT, trained on vast amounts of language data, learn to understand grammar, sentence structure, common phrases, context, tone and style, enabling them to generate coherent and contextually appropriate text.

Generative AI can be run on a variety of models, each employing unique methods for training the AI and generating results. These include generative adversarial networks (GANs), Variational AutoEncoders (VAEs) and denoising diffusion probabilistic models (DDPMs).