Generative AI refers to deep-learning models specifically designed to create new and original data rather than just classifying, identifying, or predicting existing data. Unlike traditional AI that might recognize a cat in a photo, Gen AI can draw or describe a cat that has never been seen before.
These models are trained on massive, diverse datasets (e.g., billions of text documents, millions of images, or hours of audio). During training, the model learns the underlying probability distributions and latent spaces of the data. This allows it to understand the fundamental rules and patterns (e.g., how words form sentences, how light and shadow behave, or how code functions) and use these rules to generate entirely novel, high-quality outputs that mimic the training data.
Novelty: Creates content that did not exist in the training set.
Multimodality: Can generate text (LLMs), images (Diffusion Models), audio, video, and code.
Foundation Models: Many large Gen AI models are considered 'Foundation Models' because they can be adapted or fine-tuned for a vast range of downstream tasks (e.g., summarizing, translating, writing code).