AI Basics Glossary: Models, Training, Inference, and Tokens Explained
Open any AI article and you will hit words like "model," "training," "token," and "context window." If you cannot understand them, the article is basically wasted. This guide has one goal: to explain the most common AI terms in one sentence plus one real-life example each. These terms are connected: training determines model ability, tokens drive cost and window size, fine-tuning and RAG are two ways to improve answers, and agents turn conversation into action. After reading, you can follow most AI beginner content and move on to our prompt engineering for beginners guide. To keep up with AI application news, bookmark the AI-related category.
1. Model: An Intern Who Knows How to Write
In one sentence: a model is a program that has learned a certain ability — think of it as an intern who has read a huge amount of material and knows how to write.
GPT, Claude, and Gemini are all language models. They are not databases that "look up" answers; they "write" the most plausible answer based on patterns they learned. Interns differ in ability, and so do models — the difference comes from training.
2. Training and Inference: Learning vs. Answering
Training: the process of teaching a model to find patterns in huge datasets — like an intern reading a whole cabinet of materials until they know how to write a polite reply.
Inference: the act of the model actually answering a question — like the intern sitting down to write your letter. Training happens once and is expensive; inference happens on every request and is cheaper but still costs money. To compare model quality, see our AI model evaluation guide.
3. Token: AI's Counting Unit for Text
In one sentence: a token is the smallest unit AI uses to process text — roughly half a word or one character.
In English, a word usually splits into 1-2 tokens ("ChatGPT" might be 3), while Chinese roughly counts 1-2 characters per token. So "100 words is about 150 tokens" is a common estimate. Tokens drive two things: the size of the context window and your bill — most APIs charge per token.
4. Context Window: How Much Conversation It Can Remember
In one sentence: the context window is the total number of tokens a model can "see" in one go — how long a conversation it can remember.
The longer the window, the longer the documents and chats it can handle. Entry-level models may have only a few thousand tokens, while mainstream models reach 128K (roughly 100,000+ Chinese characters). But a bigger window does not mean better recall — middle parts of long conversations can still be overlooked, so put key points at the start and end.
5. Fine-Tuning and RAG: Two Ways to "Catch Up"
When a model answers poorly, you usually have two paths:
- Fine-tuning: keep "tutoring" the model with your own examples so it learns a style or a domain — good when you always want the same writing style. See our fine-tuning tutorial.
- RAG (Retrieval-Augmented Generation): "look up" material in your own knowledge base before answering, then answer with it — good when answers must follow the latest data, such as company document Q&A. See the RAG implementation guide.
In one sentence: fine-tuning changes the brain; RAG is an open-book exam.
6. Agent: An Assistant That Gets Things Done
In one sentence: an agent is an AI program that can break down a task, call tools, execute step by step, and deliver a result.
Normal chat is question-and-answer; an agent is like a proactive assistant: "book a meeting room for next week" — it checks the calendar, sends notices, and confirms with you. To go deeper, read AI agent development basics.
7. Embedding: How AI Sees Text
In one sentence: embedding is the process of turning text into a string of numbers (a vector), letting AI compare whether two texts are "similar in meaning."
Whether two sentences mean the same thing is intuition for humans, but for AI it is about vector distance. "Nice weather today" and "The sun is lovely today" sit close together, while "We have a meeting today" sits far away. Chat, retrieval, and vector databases all rely on it. To go deeper, see the vector database guide.
8. One Table to Remember All Terms
| Term | One-Sentence Explanation | Real-Life Analogy |
|---|---|---|
| Model | A program that learned an ability | Intern who can write |
| Training | Teaching a model from data | Reading to study |
| Inference | The model actually answering | Sitting down to answer |
| Token | Smallest unit of text billing | Half a word / one character |
| Context window | Total tokens it can remember | Working memory size |
| Fine-tuning | Reshaping the model with your examples | Changing the brain |
| RAG | Checking material before answering | Open-book exam |
| Agent | An AI that executes tasks on its own | Proactive assistant |
9. FAQ
Q: Is a bigger token count always better?
Not necessarily. Tokens decide how much text you can process, but models may still miss middle information in long contexts. Put key points at the start and end.
Q: Can fine-tuning and RAG be used together?
Yes. Fine-tune for tone and format, then use RAG for fresh facts — a common production combination.
Q: How do I run a small model locally?
Start with an open-source model of a few GB. For GPU and server sizing, see our local AI deployment hardware guide.
Q: With so many concepts, where should I start?
First grasp model, token, and context window — they decide "how big you can use, how much it costs, and how much it remembers." Fine-tuning, RAG, and agents are advanced abilities; learn them when you need them.