Skip to content

How AI works

A free tour of the machinery behind AI tools, in five short modules. Each module walks through a few interactive diagrams, or explainers, that you run and change yourself. There is no code and no sign-in, and each module takes about 15 to 20 minutes.

The tour is a companion to the courses, not a step in them. Take the modules in order or dip into one whenever you want a clearer picture of what a tool is doing.

Who it is for

Anyone in a veterinary team who wants to know what goes on inside AI tools. You do not need any maths or a computer science background.

What is real and what is simulated

Some explainers run the real maths. Others are scripted, so an idea is easy to see without a real model behind it. The caption under each explainer says which.

The clinical examples are illustrative. They were written for teaching, not taken from real cases, and a clinician review of them is still to come.

The modules

  1. Neural networks

    Module 1. About 20 minutes.

    Watch a tiny neural network turn two numbers into an answer, see the activation functions that shape each step, and watch backpropagation nudge every weight after a wrong guess. The same parts, scaled up, sit inside every model in this tour.

  2. Language models

    Module 2. About 20 minutes.

    Follow a language model from one next-word prediction to a full reply. You see attention weigh a whole sentence, watch a model learn from free training examples, see one unlucky word steer an answer, and act as the rater who shapes an assistant.

  3. Retrieval

    Module 3. About 15 minutes.

    See a clinical note split into chunks and placed on a map by meaning, a question find the nearest chunks, and an answer built on those chunks cite its sources. One question shows why the closest match is not always the right one.

  4. Agents

    Module 4. About 15 minutes.

    Step through an agent answering a dosing question with tools, and see that the model only ever writes text while the harness does the work. Then change the harness's permissions and watch the same model's behaviour change with them.

  5. Images

    Module 5. About 15 minutes.

    See an image cut into patches that become tokens beside text, and how patch size trades detail for cost. Then watch a diffusion model turn pure noise into a picture, and see that training runs the same process in reverse.