The site brings together a course divided into chapters and its hands-on demonstrations, three paths that lead to the course depending on who you are, a blog that follows the subject, and service pages like this one.

To learn, come in through the course. To be pointed to what fits your situation, teenager, working adult or teacher, come in through a path. To see it all at once, the list that follows covers every page, grouped by theme, with one sentence on each.

The home page

• Home

Where to start

• Understand The course contents: how a model builds its answer, why it makes things up, where its biases come from, and when not to use it. No background needed.
• The demonstrations Hands-on demonstrations, running entirely on your device: no outgoing request, a reproducible seed and a method written out by hand.
• The paths The three paths through the course at a glance: who each one is for and what you get from it, so you can choose without opening all three.
• The blog Texts that extend the course: a technical point explained, a public source taken apart, a reported use. Sources checked, CC BY 4.0 licence.

The course

• Words in pieces A model reads neither letters nor words, but pieces learnt by statistics. The BPE method, live, and what it costs from one language to another.
• The next word How a model works out a probability for every possible word, draws from it, and how temperature, top-k and top-p change that draw.
• Why it makes things up The same toy model answers with the same assurance to a known opening and to one absent from its corpus. Nothing in the answer says which is which.
• Where the biases come from Skew the make-up of a dataset, and the model learnt on it gives the skew back to the exact figure. Bias is a consequence, not an intention.
• What you give it Four things happen to a file handed to an artificial intelligence service: reading, keeping, retraining, giving back.
• A fabricated image Two images fabricated on this page, their spectra side by side, and the trace left by upsampling. Then its limit: a little noise wipes it out.
• A fabricated voice A voice sound fabricated roughly, then carefully: hear the difference, see it on the spectrum, and why it vanishes on the phone.
• Watermarks The declared provenance of a file comes off with no effort. A statistical watermark holds out better, in a text or an image, without being invincible.
• When not to use it Three tests to decide: the cost of an error, whether it can be checked, and legitimacy. And why an AI detector is never grounds for accusing anyone.

The demonstrations

• Words in pieces A model reads neither letters nor words, but pieces learnt by statistics. The BPE method, live, and what it costs from one language to another.
• The next word How a model works out a probability for every possible word, draws from it, and how temperature, top-k and top-p change that draw.
• Why it makes things up The same toy model answers with the same assurance to a known opening and to one absent from its corpus. Nothing in the answer says which is which.
• Where the biases come from Skew the make-up of a dataset, and the model learnt on it gives the skew back to the exact figure. Bias is a consequence, not an intention.
• A fabricated image Two images fabricated on this page, their spectra side by side, and the trace left by upsampling. Then its limit: a little noise wipes it out.
• A fabricated voice A voice sound fabricated roughly, then carefully: hear the difference, see it on the spectrum, and why it vanishes on the phone.
• Watermarks The declared provenance of a file comes off with no effort. A statistical watermark holds out better, in a text or an image, without being invincible.

The three paths

• For teenagers The path through the course for teenagers: the order to follow it at fifteen, from homework handed in to a call that imitates a familiar voice.
• For work How to go through this course on artificial intelligence for professional use: what matters most first, and what you will be able to do at the end.
• For teaching The order for covering the ideas of the course in class, one session ready to use, the traps of the subject and the official frameworks to prepare it.

The blog

• What becomes of an uploaded file What can happen to content handed to a service that uses AI, and the question to ask before uploading a file.
• AI detectors at school What detectors of AI-generated text promise, what they actually do, and who pays for their mistakes at school.
• The checking routine How to check, in three steps, a summary produced by a machine: reopen the source, place the nuance, check the date.
• It does not know that it does not know Why a language model can state an error with the same assurance as an accurate fact, and what its training rewards.
• The price of a language Why the same question, asked in two different languages, does not cost the same to compute: what cutting text into tokens changes.

The service

• The glossary The words this course uses, defined plainly, each tied to its chapter by an anchor: token, hallucination, bias, watermark, black box.
• The official frameworks The OECD and European Commission framework, the two UNESCO frameworks and one non-European framework: what they ask people to know, and where this course gives it.
• Module: the European obligation What Article 4 of the European artificial intelligence regulation requires in literacy, what it does not require, and where this course covers it.
• Following the site Following what is published without giving an email address: what a feed is, how to subscribe, and why this site prefers it to a newsletter.
• About What this site is, what it is not, the method that ties every fact to its source, who stands behind it, and why it is checkable without sending anything.
• Site map Every page on this site, grouped by section and described in one sentence, plus the sitemap, the blog feed and the open data.
• Frequently asked questions Short, checkable answers to the questions that come up most about this site: artificial intelligence, the price, your data, the languages, ODERSA.
• Licence and reuse What the CC BY 4.0 licence of The AI Manual allows: copying, adapting, translating, printing, republishing, including in paid training, with credit.
• Credits The credits of The AI Manual: official frameworks, scientific sources by theme, font licences, demonstration corpora, no third-party dependency.
• Contact Contacting The AI Manual: one address for the service, another for the association. By email, with no form and no account.
• Report an error Reporting an error on The AI Manual: a wrong date, a dead link, a source that has changed, by email, with no form and no account.

Other ways through

This page is made to be read on screen. Other formats carry the same information, for a piece of software or for anyone who would rather follow the site without coming back to it.

• The sitemap, for search engines and the software that crawls the site: /sitemap.xml.
• The blog feed, to follow the articles without coming back to this page: /flux.xml. The page The feed explains how to subscribe to it.
• The open data, for anyone who wants the site’s numbers and body of content outside a web page: /data/impact.json and /data/pages.json.

No page served by this site escapes this map: the map and the sitemap are written by the same build, from the same count.