In plain words
An AI assistant such as Claude, ChatGPT, Gemini or Mistral's Le Chat is built on a large language model. Imagine someone who has read an immense library and become very good at predicting how a text continues: from your question, they write the most plausible answer, word after word. That makes them excellent at drafting, summarising, explaining, translating or programming. But "plausible" doesn't mean "true": they can be wrong with complete confidence. So use it like a brilliant colleague whose work you check.
1. The basic vocabulary
| Term | Meaning |
|---|---|
| Generative AI | AI that produces content: text, code, images, sound, video |
| LLM (Large Language Model) | The "engine" of a text assistant: a neural network trained on huge amounts of text |
| Model | A specific version of an LLM (for example a Claude, GPT, Gemini, Mistral, Llama, Qwen… family) |
| Prompt | What you write to the model: question, instructions, documents provided |
| Token | The piece of text the model handles: a short word, part of a word, a punctuation mark |
| Context window | The amount of text (in tokens) the model can take into account at once: the conversation, attached documents, its answer |
| Training | The very costly phase in which the model learns from data; it stops at a cut-off date |
| Inference | Using the model once it's trained: when it answers you |
| Hallucination | An invented answer presented as accurate |
2. How it works, without the maths
- Splitting: your text is split into tokens. In English, a word is on average a little more than one token.
- Prediction: for every possible token, the model calculates the probability that it comes next, taking the whole context into account.
- Choice and repetition: it picks a token, adds it to the text, and starts again, until the end of the answer.
The model doesn't "look up" the answer in a database: it reconstructs it from what it has learnt. That explains both its flexibility (it understands requests worded in a thousand ways) and its mistakes (it can reconstruct a false detail, a reference or a command option that doesn't exist).
After base training, models are fine-tuned to follow instructions, hold a conversation, refuse dangerous requests and admit their uncertainties. Recent models can also "think" before answering: they write out intermediate reasoning, which markedly improves results on complex problems (maths, code, analysis).
3. What it can do, and its limits
| Strengths | Limits |
|---|---|
| Drafting, rewording, summarising, translating | Can invent facts, sources, commands |
| Explaining a concept at different levels | Its knowledge stops at its cut-off date, unless it can search the web |
| Writing, reviewing and fixing code, scripts, configurations | Doesn't know your environment: you have to describe it |
| Analysing a document, a table, an image, an error log | A context window is still limited: a very long document may be partly "forgotten" |
| Proposing plans, templates, procedures | Complex calculations are more reliable when it runs code than "in its head" |
| Talking it through to refine step by step | Its answers vary from one time to the next for the same question |
Always check what matters
A command to run in production, a figure in a report, a legal reference, medical advice: check in the official documentation or with a competent person. Asking the model for its sources, then opening them, is a good habit. On this site, commands are checked and sources are cited at the bottom of each page.
4. Tools, search and agents
A model on its own only produces text. Today's assistants give it tools:
| Capability | What it changes |
|---|---|
| Web search | The model consults recent pages and cites its sources: it goes beyond its cut-off date |
| Code execution | It calculates, analyses a data file, produces a chart, by actually running code instead of "guessing" the result |
| Reading files | PDFs, spreadsheets, images, screenshots, logs |
| Connectors | Access to your applications (email, calendar, cloud storage, ticketing tools), within the permissions granted |
| RAG (Retrieval-Augmented Generation) | The system fetches the relevant passages from your documents and gives them to the model before it answers: the technique behind "chat with your documents" |
| MCP (Model Context Protocol) | An open protocol that standardises the connection between an assistant and tools or data sources |
An agent is a model that chains several steps by itself with its tools to reach a goal: search, read, write a file, run a command, check the result, correct. That's the principle behind programming assistants such as Claude Code, which read a project, modify the code and run the tests.
An agent really acts
An agent with access to a terminal, an email account or a server can delete, send or modify for real. Give it the fewest permissions possible, work on a copy or a test environment, review important actions before approving them, and keep backups. Also beware of prompt injection: a document or a web page can contain hidden instructions intended to hijack the agent.
5. Closed models, open models, local AI
| Closed (proprietary) models | Open-weights models | |
|---|---|---|
| Examples | Claude (Anthropic), GPT (OpenAI), Gemini (Google) | Llama (Meta), Mistral and Mixtral (Mistral AI), Qwen (Alibaba), Gemma (Google), DeepSeek |
| Access | Through the publisher's online service or its API | Model files can be downloaded: you can run them yourself |
| Performance | Generally the most capable | Very good, a notch behind the best closed models depending on the task |
| Privacy | Data goes through the publisher | Can stay entirely on your machine |
| Licence | The publisher's terms of use | Variable: some licences are open, others impose restrictions |
Running an AI yourself
Free tools let you use an open model without an internet connection:
| Tool | Role |
|---|---|
| Ollama | Downloads and runs models in one command; Windows, macOS, Linux |
| LM Studio | Graphical interface to find, download and chat with models (free of charge, not open source) |
| llama.cpp | The open source engine many of these tools are built on |
| Open WebUI | Web chat interface, to connect to Ollama, self-hostable on a homelab |
Example with Ollama, once installed:
ollama run mistral
The limiting factor is the graphics card memory (VRAM), or unified memory on an Apple silicon Mac. Models are generally quantised (compressed, for example to 4 bits) to fit in memory, with a slight loss of quality:
| Model size | Memory needed, quantised to 4 bits (order of magnitude) | Typical hardware |
|---|---|---|
| 3 to 4 billion parameters | 3 to 4 GB | Recent laptop, even without a dedicated graphics card (slowly) |
| 7 to 8 billion | 5 to 6 GB | 8 GB graphics card, 16 GB Mac |
| 12 to 14 billion | 9 to 10 GB | 12 to 16 GB graphics card |
| 30 to 32 billion | 20 GB and more | 24 GB graphics card, Mac with 32 GB or more |
| 70 billion and more | 40 GB and more | Several graphics cards, Mac with large unified memory |
Allow extra memory for the context window. A small local model is genuinely useful (summarising, rewording, classifying text), but remains well below the large online models for complex reasoning and code.
6. Privacy and security
What you never paste into an online assistant
Passwords, API keys, private keys, bank details, health data, personal information about customers or employees, confidential company documents: not without your organisation's agreement and a suitable plan. Mask or replace sensitive data (spo, 192.0.2.10, example.com) before pasting a configuration file or a log.
- Read your service's privacy settings: depending on the plan and your choices, your conversations may or may not be used to train models, and are kept for longer or shorter periods.
- In a business, use the business plans, which generally include contractual commitments (no training on your data, retention period, location), and define an internal usage policy. Data protection and cybersecurity authorities publish recommendations, such as the French CNIL and ANSSI (links at the bottom of the page).
- The European AI Act, which came into force in 2024, applies in stages: in particular, it requires transparency about generated content and stricter obligations for high-risk uses.
- Transparency: say when content was produced or assisted by an AI, as this site's About page does.
7. Wording your requests well
The quality of the answer depends heavily on the quality of the request. A few simple principles:
| Principle | Less effective | More effective |
|---|---|---|
| Give the context | "My server isn't working." | "Debian 13 server, nginx no longer starts after an update. Here's the output of systemctl status nginx: …" |
| State the goal and the audience | "Explain RAID." | "Explain RAID 1 and RAID 5 to a beginner who wants to buy a 2- or 4-bay NAS." |
| Specify the expected format | "Write me a procedure." | "Write a procedure in numbered steps, with the exact commands and a table of common errors." |
| Give an example | "Write in my style." | "Here are two paragraphs I wrote; keep this tone and length." |
| Ask it to check or to doubt | (nothing) | "Flag anything you're not sure about, and give the official sources." |
| Iterate | Start again from scratch | "Keep the structure, but shorten part 2 and add an example." |
8. A few concrete uses in IT
- Understanding an error message: paste it with the context (system, version, what you were doing).
- Writing a script (PowerShell, Bash, Python): describe the input, the expected output, the system; test on a test machine.
- Reviewing a configuration (nginx, sshd, firewall): ask for security problems and likely errors.
- Documenting: turning notes into a procedure, a runbook or an incident report.
- Learning: asking for a step-by-step explanation, then exercises to check you've understood.
Key points
- An AI assistant predicts how a text continues: brilliant for drafting, explaining and programming, but it can invent.
- Check what matters, ask for sources, test commands before production.
- Tools (search, code, files, connectors) and agents multiply its capabilities: give them the fewest permissions possible.
- Open models + Ollama = local AI, without your data leaving the machine; graphics card memory is the limit.
- Never paste secrets or personal data into an online service without a suitable framework.
- A good request = context + goal + format + example.