AI glossary: 20 terms in plain words
Tokens, prompts, hallucinations, agents and RAG — every key AI concept with a human explanation.
- Artificial intelligence (AI)
- Software that learns from examples instead of following hard-coded rules. Modern AI finds patterns in data and applies them to new situations — from face recognition to text generation.
- Neural network
- A mathematical model of millions of 'neurons' with adjustable weighted connections. It learns by the loop 'predict → compare with the answer → adjust weights', repeated billions of times.
- LLM (large language model)
- A neural network trained on huge volumes of text: ChatGPT, Claude, Gemini. It performs one operation — predicting the next token — and answers, translations and code emerge from it.
- Token
- The chunk of text a language model operates on: part of a word, a word or a symbol, 3–4 characters on average. Model context and API pricing are measured in tokens.
- Prompt
- The request text you send to an AI. Answer quality is ~80% determined by prompt quality: role, task, context and format make it strong.
- Prompt engineering
- The discipline of crafting effective AI requests: formulas, few-shot examples, chain-of-thought, structured output. One of the fastest-growing job skills.
- Hallucination
- A confident, plausible, but invented model answer. A consequence of LLMs' generative nature: with no facts available, a 'likely continuation' still looks like an answer. Verify numbers, quotes and links.
- Context window
- The amount of text (in tokens) a model 'holds in mind' within one conversation: your messages, its replies, uploaded documents. What doesn't fit gets 'forgotten'.
- Model training
- The process of tuning a network's billions of weights on large data. It takes thousands of GPUs and weeks of compute — which is why few companies build large models.
- Weights
- The numeric parameters of a network adjusted during training. 'A 70-billion-parameter model' refers to its weight count.
- Knowledge cutoff
- The date after which events are absent from the model's training data. Anything newer reaches it only via web search or materials you provide.
- Generative AI
- AI that creates new content — text, images, music, video, code — rather than only classifying existing content. ChatGPT and Midjourney are generative models.
- Few-shot prompting
- A prompting technique: show the model 1–3 'input → output' samples right in the request so it copies style and structure. Replaces paragraphs of explanation.
- Chain-of-thought
- Asking the model to solve step by step, showing its reasoning. Dramatically improves accuracy on logic and math and makes errors visible.
- RAG
- Retrieval-Augmented Generation: before answering, the system finds relevant chunks in your knowledge base and passes them to the model. The AI answers from your documents, not from memory.
- AI agent
- An AI system that doesn't just answer but acts: plans steps, calls tools (search, code, APIs), checks results and continues toward the goal. The #1 fastest-growing skill.
- MCP
- Model Context Protocol — an open standard for connecting AI to external tools and data: databases, services, files. The 'USB port' of AI agents.
- Fine-tuning
- Additional training of an existing model on your data for a specific task. Costlier and harder than prompting — needed when in-prompt examples no longer suffice.
- Multimodality
- A model's ability to work across data types: text, images, audio, video. Modern flagships understand pictures and voice, not just text.
- Vibe coding
- Building sites and apps through dialogue with AI: you describe the product in words, the model writes the code, you iterate. Word of the year and the fastest-growing route into development.
Terms make more sense in practice
Start for free →