What is an LLM?¶

Level: Beginner

The core idea fits in one sentence: an AI assistant does not know, it computes what is most likely. This tip explains what that means in practice and why it changes the way you use the tool.

A model trained to continue a text¶

A large language model, or LLM, is a program trained on huge amounts of text. During this training, it learns the patterns of language: which words follow which, in what context, with what style.

When you ask a question, it does not look up the answer in a database of facts. It produces the most likely text following your message, one fragment at a time. These fragments, words or pieces of words, are called tokens.

The closest analogy is your phone's autocomplete, at a far greater scale and with far more finesse.

Why the answer seems to understand¶

The text it produces is fluent, coherent and often right, because the training texts contain an enormous amount of knowledge. The model reproduces reasoning, explanations, styles. It gives the impression of understanding because it imitates the form of understanding very well.

This works remarkably well for rephrasing, summarizing, structuring and suggesting ideas. It is more fragile as soon as a precise fact, a calculation or recent information is needed.

What this means for you¶

• A fluent answer is not necessarily accurate. A confident tone does not measure reliability.
• Two identical requests can give two different answers. A controlled amount of randomness goes into choosing each fragment.
• The model has a knowledge cut-off date. What happened after its training is unknown to it, unless the tool provides a search or a document.
• It only knows what you show it in the conversation, on top of what it has learned.

The memory of a conversation¶

A model has a limited working area, called the context window. It holds your current conversation. When the conversation gets very long, the first exchanges may be taken into account less well.

This is one of the reasons to start a new conversation for each topic, as in the tip Structuring your first message. Between two conversations, the assistant has no reliable memory, unless your tool offers an explicit memory feature.

What an LLM is not¶

• A search engine: it does not fetch a page, it generates a text.
• A calculator: it can get operations wrong, especially long ones. Some tools pair it with a calculation module.
• A database: it has no exact archive of what it read.
• An accountable expert: it does not answer for what it states.

Your tool may add features (web search, document reading, calculation). They improve results, but the core mechanism stays the same.

The right reflex¶

The most important consequence of this mechanism is the risk of hallucination, covered in the tip Hallucinations: why AI makes things up.

Hands-on exercise (5 minutes)¶

1. Ask a question you already know the answer to.
2. Ask the same question again in a new conversation.
3. Compare the two answers. Are they identical?
4. Check an important point against a reliable source.