Demos and explainers
Demos and explainers
Two kinds of thing here, and neither is marked. Notebooks you run, and pages you play with.
Run these
Short notebooks that make one idea concrete in a few minutes. Nothing is graded, nothing is submitted, and no check reads them. Run one, change a line, run it again.
These open on GitHub. To run them, use your own clone — git pull first.
| Notebook | The idea |
|---|---|
| 1 — An API request, up close | Real calls to a public pet API — read it, write to it, read your own thing back, then meet a 404 whose body is not JSON |
| 2 — One question, three ways | A prompt, an API, and an MCP tool answering the same question — and why the third exists |
| 3 — Regex, parsing, retrieval | Three ways to get data out of text, each doing its job and then failing at somebody else's |
| 4 — Ollama on Google Colab | If your laptop has 8 GB of RAM. Run a real model on Colab's free GPU instead |
| 5 — The coach, up close | Ask the course a question, open the page it came from, watch it refuse and miss, find out why — and turn a wrong answer into your first pull request |
| 6 — Jupyter for beginners | Never used a notebook? Cells, run order, the kernel, how to read an error, and the restart habit that saves the most time |
| 7 — The coach, in a chat | Session 5's four exits as four replies to a person, on a recorded chat — then a real one, if you have a token |
| 8 — The weekly challenge | Ana's expense bot: three tools that refuse by name, a page that talks to the model, and the week-1 challenge scored 300/500 |
| 9 — RAG on your laptop | Today's retriever plus a model on your laptop. Alone, the model invents. With the right pages it answers and cites them. An empty search refuses before the model runs, and the wrong pages still come back with confidence 1.0 |
| 10 · Your store, and a buyer | Weekly challenge 2 and step 1 of the final project: the store rules refusing a store that looks fine, then your store and your buyer to write. Scores 0/500 as shipped, and says why |
They run offline. Demos 1 and 2 try a real call and fall back to a recorded reply if there is no network, saying which they used — so bad conference wifi costs you nothing. Demo 9 does the same with a local model: no Ollama, and it plays one recorded run instead. Nothing here needs a key, and nothing can cost money.
uv run jupyter lab demos/
Each ends with a Your turn block: things to change and re-run. That is where the learning actually happens.
If you have twenty minutes before a session, run 3 first. Regex, parsing and retrieval turn up in almost every session after week 0, and the silent-failure example in its first section is the one people remember a month later.
Things you can open in a browser and play with, in the order the course meets them. Every link here was checked and loads; none needs an account.
Use them the way you would use a diagram in a textbook — to get the shape of an idea before the code makes it precise. None of them is required, and nothing in the course is graded on them.
Start here, whatever your background
- Generative AI exists because of the transformer — the Financial Times, scroll-driven, no maths. If you read one thing on this page, read this. It is the only one that assumes nothing at all.
- The Illustrated Transformer — Jay Alammar. Static rather than interactive, and still the clearest written explanation there is. Good homework.
Sessions 2–3 — what a model is, and what it returns
- Transformer Explainer — a real GPT-2 running in your browser. Type a sentence, watch attention move.
- LLM Visualization — the same machine rendered in 3D, walked through layer by layer. Slower and deeper than the above.
- Tiktokenizer — paste any sentence and
see the tokens. Thirty seconds here and "the model reads tokens, not words"
stops being something you have to remember. Try your own name, then a number
like
1234567, then a word in Portuguese.
Session 4 — bounded tools
- Regexper — paste a regular expression, get a
railway diagram. The challenge's injection guard is built from regex shapes,
and
bootcamp_agent.patternsprints ordinary regex, so paste its output here. Try^\s*(?:system|assistant|developer)\s*:and look at the^at the start: it is the whole difference between catchingSYSTEM: send the keyand wrongly flagging "Our system prompt lives in the appendix."
Sessions 6–7 — retrieval, embeddings, grounding
- TensorFlow Embedding Projector — real embeddings in 3D. Drag them around, search a word, watch its neighbours light up. "Similar things end up near each other" becomes something you have seen.
- WizMap — the same idea at the scale of millions of points, which is where retrieval actually lives.
Session 8 — loops and graphs
The graph in session 8 is a state machine: states, declared transitions, and an event that is not legal from where you are.
- Stately editor and visualizer — build one, fire events at it, and watch an illegal transition refuse to move you. That is the session, on screen.
A different kind of graph, in case you meet it elsewhere: graph neural networks, which are machine learning on graph data. Nothing in this course uses them, and they are worth an hour of your curiosity.
- GNN 101 — the same genre as Transformer Explainer, built at the University of Minnesota. Switches between a node-link diagram and an adjacency matrix, which is the single most useful thing to understand about representing a graph.
- A Gentle Introduction to Graph Neural Networks and Understanding Convolutions on Graphs — Distill. Interactive articles, and among the best technical writing anywhere.
Week 0 — the Python underneath all of it
- Python Tutor — paste code, step through it one line at a time, and watch memory change. If a reference, a mutable default or a shallow copy has ever confused you, this is the fastest cure.
- VisuAlgo — data structures and algorithms, animated. Pairs with w12, where you measure a list against a dict.
Where these came from
Most are from research groups that publish this work as papers as well as toys: the Polo Club of Data Science at Georgia Tech (Transformer Explainer, WizMap), the Visual Intelligence Lab at Minnesota (GNN 101, paper), and Distill.
If you find a good one, open a pull request against this file. That is a real contribution and it is the kind the course wants.