Reading a clear explanation gives you a feeling of understanding that vanishes the moment you have to reproduce anything. That's the flaw in every online course, and it was ours. The Alyriaz Academy now has 109 exercises where you produce something, and where someone responds to you.
The Academy has been around since July. Twenty-three courses, one hundred and fifteen lessons, six levels from complete beginner to expert. Until this summer, you read it, and a four-question quiz validated each lesson.
The quiz has one merit: it's honest about what it measures. It checks that a piece of information has sunk in. It doesn't check that you'd know how to use it on a Monday morning, facing a real problem, with a real budget. Those are two different things, and the second is the one that matters.
One exercise, in full
Here's the one from the course on agents, intermediate level. It's representative of what we're trying to do.
Every Monday, you collect the orders that came in by email over the weekend. For each one, you need to extract the client, the reference and the quantity, fill in a row in a spreadsheet, and alert the workshop if the quantity exceeds stock. The emails are written freely: each client has their own habits, nobody fills in a form.
Three options: an autonomous agent, to whom you describe the objective. A classic pipeline that calls the model at fixed steps. An agent, but constrained by a cap on steps and a validation.
Take thirty seconds before reading on.
Many choose the autonomous agent, and for a reason that seems excellent: the emails are unpredictable. That's true. But it's the content that's unpredictable, not the path. The path, for its part, is always the same: extract, fill in, alert. Always in that order, always the same three steps.
Yet that's exactly the definition of the case where a classic pipeline beats the agent, at a tenth of the cost. You do call a model at the extraction step, where the unpredictability lies, without needing autonomy anywhere else.
The trap in this exercise isn't a wrong answer. It's a line of reasoning that almost holds together, and that conflates two things. That's precisely what a multiple-choice quiz can't catch.
What Mnémosyne does with your answer
You write your decision in your own words. Mnemosyne, the site's AI, reads it and responds to you.
Each exercise carries a marking rubric that you never see. It doesn't stay in the page: it's applied server-side. That stops you from reading the answer by inspecting the code, but above all, it allows precise instructions to be written into it.
The one for this exercise says, in essence: the decisive point is that the path is known in advance. If the person names the confusion between content and path, their justification is excellent, even if they hesitated on the option. And it's asked not to write the automation code on their behalf.
That's the difference from an automated grader: we don't compare your answer to a model answer. We look at whether your reasoning holds, and we tell you where it holds and where it slips. Someone who ticks the right box without knowing why gets a less flattering response than someone who picks the wrong option while articulating the right intuition.
Three forms, for three different things
The exercises aren't all decisions. There are three kinds, and each tests something different.
Putting in order. You receive the steps of a method, shuffled, and you order them. That's what fits when the knowledge is a sequence: splitting a long document before summarising it, setting a baseline score before modifying a prompt, deciding a video's format before generating it. An explanation always follows, and it focuses on the step everyone skips.
The workshop. You write something real: a request, a reusable template, a transparency statement for your clients. You start from your own week, not a textbook case. It's the most demanding form, and by far the most useful.
The decision. A situation, three defensible options, and you rule on it while explaining why. That's the one you just read.
What's around it
The rest of the Academy hasn't changed, and remains free.
460 quiz questions, four per lesson, to validate as you read. A glossary of 133 terms, each explained in two sentences with an everyday analogy. In articles, underlined words show their definition on hover.
And at the end of a level, an exam. The questions come from the whole level, with no correction along the way, and they change with each attempt: retaking it isn't reciting. On success, you download a certificate bearing a verification code that anyone can check.
This certification is issued by Alyriaz. It's neither a diploma, nor a title, nor a recognised professional qualification. It attests to an exam passed here, and that's already more than an attendance certificate.
Where to start
If you've never opened an AI, take the complete beginner level: there you open the page, write your first question, and learn what you should never entrust to a machine. Nothing assumes you know what a model is.
If you already use one and want to understand what you're doing, a thirty-second test recommends your starting course.
And if you only have five minutes, open any lesson and scroll straight down to the exercise. It's the only part you can't skim.