FSH Learn // Free // FR + EN
AI, explained by someone who ships it
Key AI concepts in plain words, each read through a product and tech leader's lens.
- Founders
- Product managers
- Tech leads
- Business leaders
Courses
Short courses that take one AI skill from first principles to a decision you can defend.
Concepts
15 liveModels
- The large language model (LLM)A model that predicts the next token from learned patterns. It has no memory and no judgment of its own, and it produces the most plausible continuation, not the most accurate one.7 min readRead
- HallucinationsThe model makes false but plausible statements because it is trained to answer, not to admit what it does not know. Product design can reduce them; nothing removes them.7 min readRead
- Tokens and costEvery token you send to a model and every token it sends back costs money and time. That makes the budget part of the design from day one.7 min readRead
- Other kinds of modelsAn LLM is only one kind of model among several. Images, speech, search by meaning, numbers, small local models: what each one does, what it changes for you and what it is good for.10 min readRead
- Reasoning effortA setting that trades thinking time and money for quality on hard tasks. On a simple task, more effort costs more and improves nothing.7 min readRead
Context
- The context windowThe fixed amount of text a model can see at once. Anything outside it does not exist for the model, however important it is to you.9 min readRead
- Prompting: briefing a model wellThe prompt is the text a model receives; prompting is the practice of writing it and choosing what the model reads. Clear instructions, each at the right level, beat one long text.11 min readRead
- RAG: look it up before you answerWhen a question comes in, the application fetches the few documents that matter and hands them to the model, instead of sending everything. The answer is only as good as the search.8 min readRead
Building
- The harness: everything around the modelThe same model gives very different results depending on the tool around it: its instructions, what it reads, what it can do and what gets checked. That layer around the model is what you need to choose and tune.8 min readRead
- Skills and plugins: know-how on demandA skill bundles instructions, reference files and sometimes code that an AI assistant loads when the task calls for it. You share it like a document and install it like software.8 min readRead
- AgentsAn agent is a model that, in a loop, picks its tools and acts to reach a goal. A fixed workflow often does the same job for less and is easier to test.10 min readRead
- GuardrailsThe controls placed around a model: validation, limits, human review and defense against prompt injection. You size them to the cost of an error.8 min readRead
- Evaluations (evals)Repeatable tests that show whether an AI feature is good enough, before and after each change. Without them, you are judging from a demo.7 min readRead
- Structured outputsA structured output forces the model's answer to follow a schema fixed in advance. Its shape is constrained; its correctness still has to be measured.7 min readRead
- Tools and MCP: connecting a model to your systemsA model acts on your systems only through tools: it requests the call, your application runs it. MCP standardizes these connections. Each one opens access you should limit and give an owner.9 min readRead