Glossary
What is generative AI?
In short
Generative AI is a family of neural networks that create new content: text, images, music, video and program code, from a description in plain language.
AI basics · generative AI
MMOOZZGG.com editorsChecked: 1 min read
For a child
You have read many fairy tales, and now you can make up your own, say about a dragon who is afraid of mice. You don’t copy someone else’s book; you invent something new from what you remember. Some computer programs can do the same. Ask one to draw a cat in space, and it draws a picture that never existed before.
In plain words
AI used to mostly recognize and sort: spam or not, cat or dog. Generative models do the opposite: they create. You write a request in words, and the model gives you a letter, a logo, a tune or a piece of a program. It has studied a huge number of examples and builds the result from them: it picks the next word of a text or reveals a picture from noise step by step. So every answer is new, even though it is built from learned patterns. Chatbots that write texts and services that draw from a description belong here.
For business
Companies use these models for drafts of letters and sales offers, support answers, product cards, and summaries of meetings and documents. Savings appear when an employee knows how to set a task and check the result, not when a company simply opens access to a service. Costs depend on the volume of requests: cloud models charge per token, and your own models need servers with graphics cards. Risks: invented facts, leaks when documents go to an external service, and copyright questions about what the model creates. Decide separately which data staff may paste into requests. Our separate analysis covers the skill of setting tasks in detail.
For a specialist
Generative models approximate the distribution of the training data and sample new objects from it. The main families: autoregressive transformers that predict the next token (the architecture from the paper by A. Vaswani and co-authors, June 2017); generative adversarial networks (GANs), where a generator and a discriminator train in a minimax game (I. Goodfellow and co-authors, June 2014); diffusion models that restore an image from noise step by step (J. Ho, A. Jain, P. Abbeel, 2020); and variational autoencoders (VAEs). Generation is controlled by the prompt, sampling parameters (temperature, top-p) and conditions such as a text description for an image. Image quality is measured with the FID metric, and texts are compared by human ratings or by judge models (LLM-as-a-judge).
Example
A marketer uploads the specs of a new product, and the model writes three versions of the description for a marketplace. A person chooses one and checks the numbers.
Related terms
Questions and answers
Sources
Where companies use it
Not sure whether your company needs this?
AI solutions map