LLMs: what your AI assistant really has under the hood
tier I · foundations · 9 min · interactive article ✦
You may already use it every day. But if tomorrow, in a meeting, someone asks you “what exactly is an LLM?”, what do you say? This article gives you the answer that holds up in a meeting — with the diagram to redraw on the whiteboard, and a quiz to check it sticks.
A prediction engine, not an encyclopedia
An LLM (Large Language Model) is a program trained on enormous volumes of text to do one thing: predict the most likely continuation of a text. When it “answers” your question, it doesn't look up a knowledge base — it generates, word after word, the most plausible continuation of your request.
what actually happens when you press Enter
The key intuition: an LLM doesn't know, it predicts. That's exactly why it's so fluent — and sometimes so confident when it's wrong.
This mechanic explains its strengths: rephrasing, summarizing, structuring, translating, drafting first versions — anything to do with the form of language, it executes at a formidable level. It also explains its signature weakness: the hallucination, that answer which is false but perfectly phrased. When the model doesn't “know,” it doesn't stay silent: it produces the most plausible continuation. With confidence.
Where does what it “knows” come from?
From its training: months of computation over billions of pages of text, frozen at a given date — that's why a model is unaware of recent events. And from its context window: everything you give it in the conversation (documents, instructions, history). Between two conversations, it remembers nothing. What you don't tell it, it doesn't know — and may well invent.
Remember the hierarchy: what you provide in the context > what it memorized during training. Serious enterprise approaches (like RAG, in tier II) leverage exactly this.
The three takeaways for a manager
Trust it with the form, keep the substance. First draft of an internal memo, summary of a document you provide, rephrasing for another audience: excellent. A precise figure, a legal reference, a recent fact: always to be verified.
Context is your number-one lever. The model knows neither your company nor your project. Everything you put in your request — documents, constraints, examples — directly improves the answer.
The data you type leaves your walls. Depending on the tool and the contract, your inputs may transit through external servers. Before pasting a confidential document, the question isn't technical, it's contractual.
test yourself — just like in the path
Your AI assistant gives you a precise market figure, with a source to back it. Your reflex?
✓ Exactly. An LLM can hallucinate a figure and the reference that goes with it, with the same confidence. Checking against the source is part of the job.
Not quite — reread the section on hallucinations and try again.
Going deeper
Why two different answers to the same question?
Prediction is probabilistic: at each word, the model samples among the most likely continuations, with an adjustable amount of randomness (the “temperature”). Two runs can diverge. For reproducible uses (classification, extraction), you reduce this variability; for creativity, you increase it.
GPT, Claude, Gemini, Mistral… what's the difference?
Same underlying principle, but different training data, sizes and settings — hence different personalities and strengths. For a manager, the choice criterion isn't the technical ranking of the month: it's the contract (data confidentiality), integration with your tools, and quality on YOUR use cases, to be tested.
The question to ask in a meeting
“For this use case, do we expect form from the model (summarize, draft, rephrase) or fact (figures, law, recent data)? If it's fact, the model will need to be connected to reliable sources — that's exactly what RAG is for.”
Sovereign AI & open-source models: why the choice of model is strategic
Not all LLMs are used the same way. With a closed API (ChatGPT, Gemini…), your data leaves for a vendor — often outside your jurisdiction. With an open-weight model (public weights, like Mistral or Llama), you can self-host it: your data never leaves your perimeter. That's the heart of sovereign AI — keeping control of your data and infrastructure, without depending on a single vendor. For a manager, two myths to drop: “open source = less capable” (false, the gap is closing fast) and “confidentiality is a technical detail” (it's a contractual and strategic decision). The real question isn't “which model is best?” but “where does my data live, and who sees it?”.
Frequently asked questions
Can an LLM be wrong?
Yes. An LLM predicts the most likely next word; it doesn't fact-check. It can produce an answer that is false but perfectly phrased (a “hallucination”). For anything that matters, verify the sources.
Does an LLM know my company's documents?
Not by default. Its knowledge stops at its training date and it has no access to your files — unless they are provided to it at question time (that's what RAG is for).
Do you need to be technical to use an LLM?
No. Speaking clearly is enough to start. The skill that makes the difference is framing the request (prompting), not coding.
Is ChatGPT an LLM?
ChatGPT is an application built on top of an LLM (the GPT family). The LLM is the engine; ChatGPT is the interface that lets you talk to it.
This notion is a full level of the IAPLC path.
In the course, you practice it on YOUR context: lessons generated for your job, interactive exercises, gamified progression.