How-AI-Works Words: LLM, Token, Training and Cutoff

This lesson covers the words people use when they talk about how AI works and how AI models are compared. You do not need to understand the maths. You just need to know what each word points to, so lessons and news stop sounding like a foreign language.

For several of these words we give a one-line meaning and link to a lesson that explains them properly.

What You'll Learn

• The difference between AI, machine learning and an LLM
• Why ChatGPT is a product and GPT is a model
• What tokens, context windows and knowledge cutoffs are
• What training, inference and parameters mean
• What temperature, benchmarks and open models are

AI, machine learning and deep learning

• AI (artificial intelligence): computers doing tasks that normally need human thinking, like understanding language or recognizing faces.
• Machine learning (ML): a way to build AI where the computer learns patterns from lots of examples instead of following hand-written rules.
• Deep learning: a kind of machine learning that uses very large networks of simple maths units. Most modern AI uses it.

Think of them as circles inside each other: deep learning sits inside machine learning, which sits inside AI. AI vs ML vs Deep Learning explains the difference in detail.

LLM (large language model)

Meaning: an AI model trained on a huge amount of text so it can read and write language. It works by predicting the next piece of text, again and again.

Example: GPT, Gemini, Claude and Llama are families of LLMs.

Learn more: What Is an LLM? or the How LLMs Work micro course.

GPT

Meaning: "Generative Pre-trained Transformer", the name of OpenAI's family of models. "Generative" means it creates text. "Pre-trained" means it learned from a lot of text first. "Transformer" is the type of design it uses.

You'll see it: in OpenAI's model names, like GPT followed by a version number. People also say "GPTs" for custom versions of ChatGPT that anyone can build.

Model vs product

This one confuses many learners, so it is worth a minute.

• A model is the AI "brain" itself, like a GPT model or a Gemini model.
• A product is the app you use, like ChatGPT or the Gemini app. A product wraps a model with buttons, memory, file upload, web search and so on.

Example: ChatGPT (product) uses GPT models. One product can switch between several models, which is why apps let you pick a model from a menu.

Why it matters: two products using similar models can still feel very different, because the features around the model are different.

Token

Meaning: a small chunk of text that the AI reads and writes, often a short word or part of a word. AI apps count usage and limits in tokens.

Example: "unbelievable" might be split into "un", "believ" and "able". In English, 100 tokens is roughly 75 words. Other languages, including Hindi, often use more tokens for the same meaning.

Learn more: What Is LLM Tokenization?

Context window

Meaning: how much text the AI can "see" at once, counted in tokens. It includes your messages, any files you upload and the AI's own replies.

Example: in a very long chat, the AI may forget details from the start because they fell out of its context window.

Learn more: LLM Context Windows Explained

Knowledge cutoff

Meaning: the date after which the model has no information from its training. It does not know about events after that date unless it can search the web.

Example: you ask about a phone launched last month and the AI says it has no information. That is the cutoff at work. Turn on web search, or give it the facts yourself.

Training and inference

• Training: the long, expensive process where a model learns from data. It happens once (plus updates) at the AI company.
• Inference: the model actually answering you. Every time you send a prompt, that is inference.

Simple picture: training is like years of school; inference is answering a question in an exam.

Parameters

Meaning: the millions or billions of numbers inside a model that were adjusted during training. They store what the model learned.

Example: "a 70B model" means a model with about 70 billion parameters. More parameters often means more capable, but also slower and more expensive to run.

Temperature

Meaning: a setting that controls how random the AI's answers are. Low temperature gives safer, more predictable answers. High temperature gives more varied, creative ones.

You'll see it: in AI playgrounds and developer tools, less often in normal chat apps. Temperature and other settings explained goes deeper.

Reasoning model

Meaning: a model that "thinks" through a problem in steps before answering. It is slower but better at maths, logic and planning.

Learn more: Reasoning Models Explained

Benchmark

Meaning: a standard test used to compare AI models, like an exam everyone takes.

Example: news says "Model X scores 90% on a coding benchmark." That means it solved 90% of the problems in that test.

Be careful: a high benchmark score does not always mean the model is best for your task. Try it yourself. What AI Benchmark Scores Mean explains the common ones.

Open model vs closed model

• Closed model: you can only use it through the company's app or service. You cannot download it. Example: most GPT and Gemini models.
• Open model (often called open-weight or "open source"): the company shares the model's parameters so anyone can download and run it on their own computer. Example: Llama, Mistral, DeepSeek and Gemma models.

Why it matters: open models can run privately on your own computer. See Run AI on Your Computer.

Multimodal

Meaning: an AI that works with more than one type of input or output, like text, images, audio and video.

Example: you upload a photo of a maths problem and ask the AI to explain it. That is multimodal.

Key Takeaways

• AI is the big field, machine learning is one way to build it, and an LLM is a language model inside that field.
• ChatGPT is a product; GPT is a model inside it.
• Tokens are small chunks of text, the context window is how much the AI can see at once, and the knowledge cutoff is where its training data ends.
• Training is how a model learns; inference is when it answers you.
• Benchmarks compare models, but your own test on your own task matters more.