What these demonstrations show, and what they never settle
Every demonstration is set and rerun on this device, without any artificial intelligence service and without a connection once the page is open. A visible seed makes every result reproducible, and a written method says exactly what the demonstration calculates.
This page gathers 7 demonstrations, one per mechanism in the course. Each one can be set, rerun, and gives a result straight away: none waits for a server, none asks for an account.
The demonstrations
Words in pieces
An algorithm that merges the most frequent pairs of characters until it has a vocabulary, applied live to a sentence and compared across several languages.
OpenThe next word
Watching, setting by setting, a model work out a probability distribution over the next word, then draw from it.
OpenWhy it makes things up
Putting an opening from the corpus to the same model, then one absent from it, and seeing that nothing in its answer says which is which, except the count it never shows.
OpenWhere the biases come from
A demonstration where you skew the make-up of a corpus and watch the model give the skew back to the exact figure, plus what the absence of a piece of data makes impossible.
OpenA fabricated image
A demonstration that fabricates two images, reveals in their spectrum the regular trace of upsampling, then makes it disappear under noise.
OpenA fabricated voice
A speech sound fabricated twice, roughly then carefully, to compare what the ear takes in with what its spectrum reveals, and how far a little ordinary noise wipes it out.
OpenWatermarks
A green-list watermark placed on a text, a message hidden in the pixels of an image and then destroyed in front of you: two marking mechanisms, and how they differ from a declared provenance.
OpenWhy everything runs on this device, and why that is not a limitation
No demonstration on this page calls an artificial intelligence service. No key, no account created behind the scenes, not one byte going out once the page is loaded. It is a teaching choice, not a technical retreat.
The trust banner, at the top of this page, already suggests the move: open your browser’s Network tab, use any demonstration above, and watch. No request goes out.
A toy model, here
The corpus fits on the page. The table of counts can be read in full. The formula that turns a count into a probability is one line long, and can be redone by hand. What a demonstration gives back follows from the setting: nothing about it stays mysterious.
An online service, elsewhere
The corpus is closed, the parameters run into billions, and no formula appears next to the answer. Seeing a whole mechanism once teaches you to be wary in the right places about what a black box never shows.
A visible seed, a result that does not change
Every demonstration draws on chance, but never on the opaque chance of the device that opens it. The draw starts from a seed: a word or a number, shown in plain sight, which you can change. Two people who enter the same seed get exactly the same result, on two different devices.
That is what makes a demonstration citable in class and reproducible in a piece of homework. A result that cannot be replayed exactly cannot be checked, and cannot be corrected either.
What none of these demonstrations claims to do
Every demonstration shows a mechanism and names its limits. None goes any further.
- Settle whether a specific piece of content, brought by you, is fabricated or not.
- Detect anything at all: every demonstration already knows the answer it is staging, it guesses nothing.
- Mark or rank a piece of work, an exam paper or a person.
- Measure a probability that a given piece of content was made by a machine.
What you need to use them
A browser
Nothing to install. Every demonstration runs inside the page, with what the browser already provides.
No account
No sign-up, no password. Nothing identifies whoever is using the demonstration.
No connection, once the page is open
With the page loaded, everything is calculated on the device. A room with no network, or a plane, is enough.
Every demonstration also carries its method, written in plain terms and possible to follow by hand, on paper: the calculation it makes, step by step, up to the result. That is what makes it possible to take it up in class without a screen, or when JavaScript is missing.
Going further
The full course, in its own order, is on the understand page. To prepare a session around these demonstrations, the for teaching path offers an order and a session ready to use as it stands. To copy, translate or print a demonstration, the licence says on what condition.