AI glossary: what all those terms you hear every day mean

When it comes to artificial intelligence, strange words inevitably appear: hallucinations, fine-tuning, embeddings, transformer. People nod, but few actually understand what is meant. It's time to fix that.

Why a vocabulary is needed now

Over the past two years, AI has moved from niche technical circles into the mainstream. Everyone talks about ChatGPT, Claude, Gemini — but the original terminology has remained the same: technical, confusing, full of Anglicisms. People hear words on podcasts, see them in articles, hear them at conferences and at work, but don't really understand their meaning. The problem isn't human stupidity, but rather that the AI community has become accustomed to speaking its own language, built on decades of research. It's time to translate these ideas into regular English — not marketing gloss, but real, working explanations that will help you not look confused in a conversation.

Main categories of AI terms

All AI words are conventionally divided into several groups, and each group has its own logic. If you understand the logic of one group, the remaining terms will fit in organically.

Models and architectures — these are fundamental concepts about how the neural network itself is structured: transformer, diffusion model, large language model (LLM), convolutional neural network. It's like architectural styles of a building — each determines how everything else will be organized.

Training and adaptation — processes that a model goes through to learn: fine-tuning, prompt engineering, reinforcement learning, backpropagation. It's like a system of exercises for an athlete — it determines what skills the model will acquire.

Problems and limitations — this is what developers and users actually encounter: hallucination, bias, alignment, overfitting, mode collapse. It's like diseases a system can have, and they need to be taken into account.

From hallucinations to prompt engineering

Let's take five of the most important and frequently used words:

• Hallucinations — when AI generates information that doesn't exist in the training data. It makes up links, quotes, figures that weren't there. It sounds like mystical fantasies, but it's actually an error in text prediction.

• Fine-tuning — additional training of an already-built model on specialized data. It's as if you took a ready-made textbook and adapted it for a specific school.

• Embeddings — transforming words or text into mathematical vectors that the model understands. It's translating human language into the language of numbers.

• Transformer — the network architecture underlying all modern LLMs (GPT, Claude, Llama). It was a breakthrough in 2017 that changed the entire world of AI.

• Prompt engineering — the art of correctly formulating a task for AI to get the desired result. Not just a command, but a command that works.

Why this matters

A vocabulary is not just a reference guide. It's a pass into the world of serious conversations about AI. When you understand that hallucination is a specific, explainable error in how a model works, conversations become completely different. You no longer just nod blindly at conferences, but ask the right questions. At work, you're not just listening to colleagues, but participating in the discussion. Understanding terminology is the first step toward understanding what AI can and cannot do, why it works and why it sometimes breaks.