# Session 1 - AI Foundations ## Full English script (40 minutes) > **Comment lire ce document** > `▶ SLIDE n` : tu changes de slide. > `[+]` : tu appuies sur flèche droite pour faire apparaître le point suivant, sans changer de slide. > *[Texte en italique entre crochets]* : indication de mise en scène, à ne pas dire. > Le reste est le texte à répéter. **Minutage** | Temps | Bloc | Slides | |---|---|---| | 0-5 | Intro et présentation | 1-5 | | 5-10 | Generative AI and LLMs | 6-8 | | 10-15 | Tokens and context window | 9-14 | | 15-20 | Confidentiality | 15-21 | | 20-26 | Hallucinations | 22-28 | | 26-30 | Harness | 29-30 | | 30-34 | MCP | 31-32 | | 34-38 | Agents | 33-35 | | 38-40 | Wrap-up and homework | 36-39 | --- ## 0-5 min - Opening **▶ SLIDE 1 - AI Foundations** *[Laisser l'animation tourner quelques secondes avant de parler.]* Good morning everyone, and thank you for being here. **▶ SLIDE 2 - Your two speakers** *[Olivier]* My name is Olivier Durand. I am co-founder and Chief Technology Officer at ADALAN. And I am not alone today. *[Etienne]* My name is Etienne Galland. I am co-founder at ADALAN. *[Olivier reprend, les trois premiers `[+]` sont pour lui.]* `[+]` I have been building technology companies for twenty years. `[+]` What I do now is help companies make AI useful. In a pragmatic way, a secure way, and in a way where they keep control of their own data. `[+]` So I am not here to sell you a demo. Most AI demos look brilliant. Most AI projects get stuck. And the reason is almost never the technology. *[Etienne enchaîne sur ses trois `[+]`.]* `[+]` My background is consulting, and financial services is my home ground. `[+]` I know how these businesses actually run. The constraints, the vocabulary, the way a decision gets made in a room like this one. `[+]` So here is how we work together. Olivier knows what the technology can do. I know what your business will accept. *[Se regarder sur cette phrase, pas la salle. C'est le moment du slide.]* `[+]` *[Olivier]* It is this. AI only creates a real advantage when teams change the way they work. That is why you are in this room, and not watching a product video. Our job today is to translate what we know into something you can use, without the jargon. Before we start, one thing about us. We are French, so our English is not perfect. If we say something strange, or if we go too fast, stop us. Really. Raise your hand. It helps us and it helps everyone else. **▶ SLIDE 3 - Five sessions** Let me tell you what these five sessions are about. Today, session one, is the foundation. By the end of these forty minutes, you will understand what these tools actually are, how they work, why they sometimes lie to you, and what is safe to put into them. Everything else we do in the next four sessions builds on today. **▶ SLIDE 4 - This is not a technical training** I want to be clear about one thing. This is not a technical training. `[+]` You will not write code. `[+]` You will not need to remember any formula. `[+]` What you will get is a mental model. A picture in your head of what is happening when you type something into one of these tools. `[+]` Once you have that picture, you stop being a passenger. You start making good decisions on your own. **▶ SLIDE 5 - Who used an AI tool this week?** *[Question à la salle. Attendre les réponses, noter deux ou trois exemples au tableau, ils serviront pendant toute la séance.]* So, a quick question to start. Who used an AI tool this week? For what? *[Écouter. Rebondir avec :]* Good. Keep those examples in mind. I will come back to them. --- ## 5-10 min - Generative AI and LLMs **▶ SLIDE 6 - Two words you hear everywhere** Let us start with the two words you hear everywhere: generative AI, and LLM. The AI we had before sorted things. This email is spam. This transaction is fraud. `[+]` Generative AI creates something new. Text, images, audio, code. That is the difference. `[+]` Today we focus on text, because that is what you will use every day. And the tool that generates text is called an LLM. It stands for Large Language Model. **▶ SLIDE 7 - An LLM predicts the next word** Here is the only technical idea I want you to keep. An LLM predicts the next word. That is it. You give it a beginning, and it computes what is most likely to come next. Then it does it again. And again. Word by word, until the answer is complete. `[+]` It is not looking anything up. There is no database inside it with the right answers. It is not searching. It is predicting. **▶ SLIDE 8 - Imagine a colleague** Think of it like this. Imagine a colleague who has read almost everything published on the internet up to a certain date. Enormous general knowledge. Very confident. Speaks well. Always has an opinion. But this colleague has three limitations. `[+]` He has no access to your CRM. `[+]` He has no access to your files. `[+]` And he stopped reading at a certain date, so he does not know what happened last month. That is your starting point. A brilliant generalist, with no access to your company, and no memory of recent events. Now, if that is true, you might ask: why does it work so well? Why does it feel like it knows things? Because predicting the next word, when you have read that much text, turns out to be an extremely powerful trick. To predict well, the model has to capture how ideas connect. And that gives you something that looks a lot like reasoning. But keep the mechanism in mind, because in ten minutes it will explain everything that goes wrong. --- ## 10-15 min - Tokens and context window **▶ SLIDE 9 - It reads tokens** Two words now that you will see on every invoice and in every product page: token, and context window. The model does not read letters, and it does not read words. It reads tokens. A token is a piece of a word. *[Démo tokenizer. Taper le nom d'un client de la salle, montrer le découpage.]* Look at this. I type the name of a company you all know, and watch how it gets cut into pieces. Three pieces, for one word. **▶ SLIDE 10 - Why this matters to you** Why does this matter to you? Two reasons. `[+]` First, money. These tools are billed by the token. Input and output. So when someone tells you an AI project costs a certain amount per month, the number of tokens is what drives that cost. Long documents cost more. That is the whole story. `[+]` Second, it explains some strange behaviour. The model does not see letters, so it is bad at things like counting letters in a word, or spelling backwards. It is not a sign of low intelligence. It is a sign that it never saw the letters in the first place. **▶ SLIDE 11 - A desk, not a filing cabinet** Now, the context window. It is how much text the model can hold at one time. Here is the image I want you to keep: it is a desk, not a filing cabinet. `[+]` Whatever is on the desk, the model can see. Whatever is not on the desk does not exist for it. And the desk has a fixed size. So when you have a long conversation and the model seems to forget what you said at the beginning, that is the desk being full. The oldest things fall off. **▶ SLIDE 12 - Everything you give it has to fit** Same image, one level sharper. That desk is a box, and the box has a fixed volume. Here is what is inside it. And the first surprise is that it is never empty when you arrive. `[+]` There are the system instructions: the rules the vendor wrote, that you never see. `[+]` There is the list of tools the model is allowed to use. `[+]` There is the way your company configured it: the role you gave it, your house style, your templates. `[+]` And then there is the conversation itself. Every message. Every document you paste. Every result that comes back from a tool. All of that stays in the box. `[+]` What is left is the room the model has to work and to write its answer. `[+]` So the box is never empty when you start, and the answer still has to fit in what is left. When a vendor tells you the context window is very large, ask the honest question: how much of it is still free once everything else is loaded? **▶ SLIDE 13 - And you are the one packing it** Which brings me to the sentence I would like you to leave with. The context window is a box, and you are the one packing it. That is your job, not the model's. `[+]` Everything you add pushes something else out. There is no free space. `[+]` Two pages of the right document beat forty pages of the wrong one. Pasting an entire report because you are not sure which part matters is the most common mistake I see. `[+]` And when the subject changes, open a new conversation. You get a clean box. Keeping a three-hour thread alive to ask an unrelated question is expensive, and it makes the answers worse. Packing that box well is a real skill, and it is what separates people who get good answers from people who do not. We will work on it in session 3. **▶ SLIDE 14 - Put the right things on the desk** And packing is not only about leaving things out. It is also about what you deliberately put in. If you want a good answer, put the right things on the desk. Your context, your constraints, your examples, the tone you want. `[+]` Most bad results come from an empty desk, not from a weak model. That is why the same tool gives one person mediocre answers and another person excellent ones. It is often not the model. It is what was placed on the desk. --- ## 15-20 min - Confidentiality **▶ SLIDE 15 - Where does that desk live?** *[Enchaîner directement, sans respirer.]* Now. If everything I type goes onto that desk, there is an obvious question. Where does that desk actually live? This is the question I get in every room, so let us deal with it properly. **▶ SLIDE 16 - Three things people mix together** There are three different things people mix together when they worry about confidentiality. I want to separate them. `[+]` One: transit. Is my data encrypted when it travels? Yes. Standard encryption, same as your online banking. This is not the problem. Moving on. `[+]` Two: storage. Your conversation is stored somewhere. There is a history. Some tools now have memory across sessions, which means what you write today can come back next week in a different conversation. `[+]` Three: training. This is the real question. Will what I type be used to improve the model? **▶ SLIDE 17 - It depends on the contract** And here is the key point of this whole section. The answer does not depend on the technology. It depends on the contract. `[+]` The same AI does not follow the same rules depending on how you access it. `[+]` A personal account on your phone and the tool your company has deployed can be the exact same model, with completely different terms. So the sentence I want you to remember is not "AI is dangerous". **▶ SLIDE 18 - Two risks that are yours** `[+]` First, client confidentiality. When you paste a contract, a specification document, or a proposal into a tool your company has not approved, you may be breaking the NDA you signed with that client. That risk is not technical. It is contractual. `[+]` Second, personal data. A contact list, a meeting report with names in it, a CV. That is personal data, that is GDPR, and the responsibility goes up to the company. **▶ SLIDE 19 - The practical rule** One sentence, and it covers ninety percent of situations. If you would not send it by email to an outside supplier, do not paste it into a tool your company has not approved. *[Marquer un silence de trois secondes.]* **▶ SLIDE 20 - So clean it first** *[Démo anonymisation. Prendre un extrait de proposition commerciale, remplacer noms, montants et raisons sociales, relancer, montrer que la réponse est aussi bonne.]* Watch this. I take a real piece of a proposal. I replace the client name, the amounts, and the company names with generic placeholders. Thirty seconds of work. Now I ask the same question. `[+]` Same quality of answer. The model did not need the real names to help me with the structure, the wording, or the argument. `[+]` So the answer is not "you can do nothing". The answer is "anonymise". **▶ SLIDE 21 - A simple traffic light** `[+]` Green: public or generic content. Rewriting, brainstorming, structuring an argument, drafting a generic email. `[+]` Orange: internal content that is not sensitive. Use it, but anonymise first. `[+]` Red: named client data, contracts, negotiated pricing, anything covered by an NDA. Not in an unapproved tool. Full stop. *[Si la charte interne existe, la citer ici. Sinon, dire :]* If you are not sure which tools your company has approved, that is a question worth asking this week. It is a five minute question with a very large impact. *[Si quelqu'un demande "et si on ne veut rien envoyer dehors ?" : "There are models you can run inside your own walls. That is part of what we do at ADALAN, and it is a subject for another session." Ne pas développer.]* --- ## 20-26 min - Hallucinations **▶ SLIDE 22 - Hallucination** Now we get to the most important part of today. You have all heard the word hallucination. It means the model produced something that sounds completely credible and is simply false. I want to change how you think about it. **▶ SLIDE 23 - A hallucination is not a bug** A hallucination is not a bug. `[+]` It is the normal behaviour of the system, applied to a question it does not have the answer to. **▶ SLIDE 24 - It was never trained to say "I do not know"** Remember what we said at the beginning. The model predicts the most likely next word. It was trained to produce a plausible answer. It was never trained to say "I do not know". `[+]` So when you ask something it does not know, it does not stop. It does what it always does. It produces the most plausible continuation. `[+]` And plausible is not the same thing as true. **▶ SLIDE 25 - Let me show you** *[Démo hallucination en direct. Demander une information vérifiable de leur métier que le modèle ne peut pas connaître : un chiffre interne, une référence client, une clause d'un contrat récent.]* Let me show you. *[Après la démo :]* `[+]` Look at that answer. Confident. Well structured. Correct format. Completely invented. `[+]` And that is what makes it dangerous. A human who does not know usually hesitates. You can hear it in their voice. The model never hesitates. The false answer and the true answer look exactly the same. *[Plan B, si le modèle refuse de répondre :]* Interesting, it refused. That is progress, and modern tools are better at this than they were. But do not rely on it. It refuses sometimes, not always. **▶ SLIDE 26 - Where the risk is highest** `[+]` Precise facts. Numbers, dates, names, references, quotes. Anything specific. `[+]` Anything about your company, because it has no access to your company. `[+]` Anything recent, because it stopped reading at a certain date. `[+]` And anything where you push it. If you insist, if you say "are you sure, find me a source", it will often produce one. Because a source is a plausible continuation too. **▶ SLIDE 27 - Three habits** What do you do about it? Three habits. `[+]` One. Separate the two types of task. Asking it to reformulate, structure, summarise a text you provided, or brainstorm: low risk, because the information comes from you. Asking it for facts it has to retrieve from memory: high risk. Same tool, completely different risk level. `[+]` Two. Verify anything you are going to put in front of a client. Every number, every reference. If you would check it before sending it from a junior colleague, check it here. `[+]` Three. Give it the information instead of asking for it. If the facts are on the desk, it does not have to invent them. This is the single most effective technique, and it costs you nothing. **▶ SLIDE 28 - Not a source of truth** The line to remember. It is not a source of truth. It is a way of working with information you provide. --- ## 26-30 min - Harness **▶ SLIDE 29 - The model is the engine** Now, a word you will probably never say out loud, but a concept you need: the harness. Here is the situation. The model on its own does almost nothing useful. It takes text in, it produces text out. That is all. Everything else, the interface, the buttons, the memory, the ability to search the web, to read a file, to run a calculation: all of that is built around the model. That layer is called the harness. The model is the engine. The harness is the rest of the car. Steering, dashboard, brakes, fuel tank. `[+]` An engine on a table is not transport. You need the car. **▶ SLIDE 30 - Same engine, different car** Why does this matter to you? `[+]` Because two products can use the exact same model and give you very different results. `[+]` So when someone says "we tested AI and it did not work", the right question is often not "which model did you use". It is "what was built around it". And this is where most of the value is created in a company project. Not in choosing the model, models change every few months anyway. In building the right harness around it: the right access to your data, the right instructions, the right guardrails. `[+]` For you, practically, it means one thing. When you evaluate a tool, do not ask which model is inside. Ask what it can actually reach. --- ## 30-34 min - MCP **▶ SLIDE 31 - A USB-C port between AI and your tools** Which brings us to the next question. How do you connect the model to your tools? MCP is a standard that solves this. It stands for Model Context Protocol. Think of it as a USB-C port between AI systems and your tools. `[+]` Until recently, every connection was custom. You want AI in your CRM? Build a specific bridge. In your email? Another bridge. Expensive, slow, and it broke constantly. `[+]` One standard plug. Any tool that supports it can be connected, quickly. I do not need you to remember the acronym. I need you to understand what it changes. **▶ SLIDE 32 - The AI comes to you** Until now, using AI meant going to the AI. You open a separate window, you copy your text into it, you copy the answer back. What is happening now is the opposite. The AI comes to you. Into your email client. Into your CRM. Into your shared drive. Into your meeting tool. `[+]` That is the shift, and it is happening this year, not in five years. And it has a direct consequence for what we discussed earlier. When the AI is connected to your CRM, the confidentiality question changes shape. It is no longer about what you paste. It is about what the tool is allowed to reach on its own, and who decided that. So the good news and the warning are the same sentence: soon you will not have to copy and paste anything. --- ## 34-38 min - Agents **▶ SLIDE 33 - A model, a harness, tools, and a loop** Last concept. Agents. Take everything we have seen. A model. A harness around it. Tools connected through something like MCP. Now add one more thing: a loop. `[+]` Instead of answering once and stopping, the system does something, looks at the result, decides what to do next, and continues. Until the objective is reached. That is an agent. **▶ SLIDE 34 - From instruction to objective** Here is the practical difference, and it is a big one. Today, mostly, you give an instruction. "Write me a follow-up email for this client." With an agent, you give an objective. "Follow up with the clients who did not answer this week." And it decides the steps. Check who did not answer. Read the history of each conversation. Draft the right message for each one. Prepare them for you. `[+]` You moved from asking for a task to delegating a goal. **▶ SLIDE 35 - The honest version** Now, the honest version. `[+]` Agents are the area with the biggest gap between the demos and reality right now. `[+]` They work well when the task is well defined and the tools are reliable. `[+]` They fail in ways that are hard to predict when the task is vague. `[+]` And the mistakes compound. If the model gets step two wrong, it builds steps three, four and five on top of that mistake. `[+]` That is why almost every serious deployment today keeps a human at the validation point. The agent prepares, the human approves. The reason I am telling you about this in session one, even though you may not use one tomorrow, is that this is the direction everything is moving. If you understand model, harness, tools and loop, you can read any announcement in this space and understand what is actually being claimed. --- ## 38-40 min - Wrap-up and homework **▶ SLIDE 36 - The one sentence** Let me close with one sentence that ties everything together. A model on its own knows nothing about your company and will invent things confidently. What we build around it is what makes it useful and safe. That is the whole session. Everything else was detail. **▶ SLIDE 37 - The words** Quick recap. Generative AI creates content. An LLM predicts the next word. Tokens are the pieces it reads and what you pay for. The context window is its desk, a box you pack yourself. Hallucination is what happens when it does not know and answers anyway. The harness is the car around the engine. MCP is the plug that connects it to your tools. An agent is all of it, in a loop, working toward an objective. *[Distribuer le glossaire FR/EN maintenant, pas avant : sinon ils lisent au lieu d'écouter.]* I am giving you a one page glossary with all of these terms, in French and in English. **▶ SLIDE 38 - Test one real question** Homework, and it is genuinely short. Between now and next session, test one real question from your own work. Something you actually do. Then bring me back one of two things. `[+]` Either a result that was useful. Tell me why. `[+]` Or a hallucination you caught. Tell me how you spotted it. `[+]` We will open the next session with your examples, not mine. That is when this becomes real. **▶ SLIDE 39 - Questions?** Thank you. Questions? *[Laisser ce slide affiché pendant toute la séquence de questions.]* --- ## Vocabulaire à répéter à voix haute Mots où la prononciation peut coincer : - **hallucination** - a-lu-si-NAY-shun - **harness** - HAR-niss (le "h" se prononce) - **plausible** - PLAW-zi-bul - **retrieve** - ri-TREEV - **guardrails** - GARD-rails - **credible** - KRED-i-bul - **compound** (verbe) - kom-POUND - **deployment** - di-PLOY-ment - **acronym** - AK-ro-nim ## Phrases de secours À garder en tête si tu perds le fil ou si une question te bloque : - "Let me rephrase that." - "Good question. Let me come back to it at the end, I want to keep the thread." - "I do not know. I will check and send you the answer." - "Does that make sense so far?" - "Let me give you a concrete example."